# Blayne Aina — AINA Transition-Governance Architecture & Verification Corpus (blayneaina.com) > Canonical orientation, architectural summary, technical briefs, and complete URL directory for automated inspection, technical peer reviewers, and R&D grant evaluators. > Contact: Blayne Aina (Principal Systems Architect & Creator of AINA, Denver, CO, USA; USMC Veteran) — me@blayneaina.com ## 1. Two-Sentence Architecture Summary AINA is a deterministic authority-enforcement architecture that treats probabilistic AI outputs as untrusted proposals. It independently validates evidence, authorization, execution identity, and permitted state transitions at the boundary where external effects are committed, producing independently verifiable evidence when an attempted operation is refused. ## 2. Compact Briefs & Direct Raw Corpus Routes 1. **Single-Request Complete Master Brief Covering All 9 Articles (§01–§09) + Proofs + TCB + Provisionals + PRJ Packages:** - HTML: https://blayneaina.com/ai-brief/ - Plain Text (<=48 lines): https://blayneaina.com/ai-brief.txt 2. **Per-Article Technical Briefs (Every Article <=44 Rendered Lines):** - Briefs: https://blayneaina.com/01/brief/ through https://blayneaina.com/09/brief/ (or .txt: /01/brief.txt .. /09/brief.txt) - Raw Verbatim Markdown: https://blayneaina.com/01.md through https://blayneaina.com/09.md 3. **3-Step Complete Series Digests:** Digest 1 (§01–§03): https://blayneaina.com/articles/digest-1/ | Digest 2 (§04–§06): https://blayneaina.com/articles/digest-2/ | Digest 3 (§07–§09): https://blayneaina.com/articles/digest-3/ 4. **Master Directory, Assurance & Proof Packs:** - Articles Index: https://blayneaina.com/articles/ - Research & Assurance Binder: https://blayneaina.com/research-assurance/ - Offline Refusal Proof Pack: https://blayneaina.com/refusal-proof/ - Semantic Switch Proof Pack: https://blayneaina.com/semantic-switch-proof/ - GitHub Repo: https://github.com/SquadForge/blayne.aina ## 3. Complete 9-Article Series, 1-Call Briefs & Lossless Part URLs ### §01 — An AI Output Is Not an Act: The Lawful-Act Defect and Machine Standing - 1-Call Brief: https://blayneaina.com/01/brief/ (Plain Text: https://blayneaina.com/01/brief.txt) - Full Single-Page URL: https://blayneaina.com/01-an-ai-output-is-not-an-act/ (Short: https://blayneaina.com/01/ | Raw MD: https://blayneaina.com/01.md) - Problem: Valid JSON, credentials, confidence, and a green tool result can make an AI proposal look authorized when it is not. - What It Proves: Machine standing binds the exact act, actor, evidence, target, execution, time, and terminal disposition before consequence. - Lossless Part URLs (6 Parts): https://blayneaina.com/01-an-ai-output-is-not-an-act/part-1/ (.txt: https://blayneaina.com/01/part-1.txt) | https://blayneaina.com/01-an-ai-output-is-not-an-act/part-2/ (.txt: https://blayneaina.com/01/part-2.txt) | https://blayneaina.com/01-an-ai-output-is-not-an-act/part-3/ (.txt: https://blayneaina.com/01/part-3.txt) | https://blayneaina.com/01-an-ai-output-is-not-an-act/part-4/ (.txt: https://blayneaina.com/01/part-4.txt) | https://blayneaina.com/01-an-ai-output-is-not-an-act/part-5/ (.txt: https://blayneaina.com/01/part-5.txt) | https://blayneaina.com/01-an-ai-output-is-not-an-act/part-6/ (.txt: https://blayneaina.com/01/part-6.txt) ### §02 — Zero-Trust Document Ingestion and Server-Cut Evidence - 1-Call Brief: https://blayneaina.com/02/brief/ (Plain Text: https://blayneaina.com/02/brief.txt) - Full Single-Page URL: https://blayneaina.com/02-zero-trust-document-ingestion-and-server-cut-evidence/ (Short: https://blayneaina.com/02/ | Raw MD: https://blayneaina.com/02.md) - Problem: Hostile documents can attack the parser, while models can fabricate citations and make unknown coverage disappear. - What It Proves: LPAC parser confinement, immutable views, explicit omissions, and server-cut srv_ and ev_ receipts establish served evidence. - Lossless Part URLs (4 Parts): https://blayneaina.com/02-zero-trust-document-ingestion-and-server-cut-evidence/part-1/ (.txt: https://blayneaina.com/02/part-1.txt) | https://blayneaina.com/02-zero-trust-document-ingestion-and-server-cut-evidence/part-2/ (.txt: https://blayneaina.com/02/part-2.txt) | https://blayneaina.com/02-zero-trust-document-ingestion-and-server-cut-evidence/part-3/ (.txt: https://blayneaina.com/02/part-3.txt) | https://blayneaina.com/02-zero-trust-document-ingestion-and-server-cut-evidence/part-4/ (.txt: https://blayneaina.com/02/part-4.txt) ### §03 — EISA Four-Gate Verdict Machine & Strict AST Parser - 1-Call Brief: https://blayneaina.com/03/brief/ (Plain Text: https://blayneaina.com/03/brief.txt) - Full Single-Page URL: https://blayneaina.com/03-eisa-four-gate-verdict-machine-and-entity-resolution/ (Short: https://blayneaina.com/03/ | Raw MD: https://blayneaina.com/03.md) - Problem: Entity resolution needs rich comparison structure, but raw contacts, companies, and products should not be sent to external AI providers. - What It Proves: The production EISA Four-Gate machine validates bounded acts over sealed evidence; Semantic Switch is an experimental alpha structure-preservation research module. - Lossless Part URLs (6 Parts): https://blayneaina.com/03-eisa-four-gate-verdict-machine-and-entity-resolution/part-1/ (.txt: https://blayneaina.com/03/part-1.txt) | https://blayneaina.com/03-eisa-four-gate-verdict-machine-and-entity-resolution/part-2/ (.txt: https://blayneaina.com/03/part-2.txt) | https://blayneaina.com/03-eisa-four-gate-verdict-machine-and-entity-resolution/part-3/ (.txt: https://blayneaina.com/03/part-3.txt) | https://blayneaina.com/03-eisa-four-gate-verdict-machine-and-entity-resolution/part-4/ (.txt: https://blayneaina.com/03/part-4.txt) | https://blayneaina.com/03-eisa-four-gate-verdict-machine-and-entity-resolution/part-5/ (.txt: https://blayneaina.com/03/part-5.txt) | https://blayneaina.com/03-eisa-four-gate-verdict-machine-and-entity-resolution/part-6/ (.txt: https://blayneaina.com/03/part-6.txt) ### §04 — The Room: Why Git and Chat Are Not Coordination Planes - 1-Call Brief: https://blayneaina.com/04/brief/ (Plain Text: https://blayneaina.com/04/brief.txt) - Full Single-Page URL: https://blayneaina.com/04-the-room-why-git-and-chat-are-not-coordination-planes/ (Short: https://blayneaina.com/04/ | Raw MD: https://blayneaina.com/04.md) - Problem: Git history and chat transcripts cannot establish who may act, receive, stop, or complete work now. - What It Proves: The Room revalidates territory, causal observation, sealed wakes, native replies, cancellation, and privileged acceptance. - Lossless Part URLs (5 Parts): https://blayneaina.com/04-the-room-why-git-and-chat-are-not-coordination-planes/part-1/ (.txt: https://blayneaina.com/04/part-1.txt) | https://blayneaina.com/04-the-room-why-git-and-chat-are-not-coordination-planes/part-2/ (.txt: https://blayneaina.com/04/part-2.txt) | https://blayneaina.com/04-the-room-why-git-and-chat-are-not-coordination-planes/part-3/ (.txt: https://blayneaina.com/04/part-3.txt) | https://blayneaina.com/04-the-room-why-git-and-chat-are-not-coordination-planes/part-4/ (.txt: https://blayneaina.com/04/part-4.txt) | https://blayneaina.com/04-the-room-why-git-and-chat-are-not-coordination-planes/part-5/ (.txt: https://blayneaina.com/04/part-5.txt) ### §05 — Expiry Is Not Fencing: Typed Leases and Kernel Write Exclusion - 1-Call Brief: https://blayneaina.com/05/brief/ (Plain Text: https://blayneaina.com/05/brief.txt) - Full Single-Page URL: https://blayneaina.com/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/ (Short: https://blayneaina.com/05/ | Raw MD: https://blayneaina.com/05.md) - Problem: An expired lease does not physically stop a stale worker or an uncooperative local writer from changing a resource. - What It Proves: Monotonic generations, current tenancy, typed scopes, and overlapping Win32 handle exclusion fence the write itself. - Lossless Part URLs (4 Parts): https://blayneaina.com/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/part-1/ (.txt: https://blayneaina.com/05/part-1.txt) | https://blayneaina.com/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/part-2/ (.txt: https://blayneaina.com/05/part-2.txt) | https://blayneaina.com/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/part-3/ (.txt: https://blayneaina.com/05/part-3.txt) | https://blayneaina.com/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/part-4/ (.txt: https://blayneaina.com/05/part-4.txt) ### §06 — Rollback Is Not Erasure: Dual-Tier Vaults and Ungated Quarantine - 1-Call Brief: https://blayneaina.com/06/brief/ (Plain Text: https://blayneaina.com/06/brief.txt) - Full Single-Page URL: https://blayneaina.com/06-rollback-is-not-erasure-dual-tier-vaults-and-ungated-quarantine/ (Short: https://blayneaina.com/06/ | Raw MD: https://blayneaina.com/06.md) - Problem: A single undo snapshot can destroy intermediate states and launder unauthorized writes into a clean baseline. - What It Proves: Primary certifies operational recovery while Amber preserves every intermediate and ungated state for forensics. - Lossless Part URLs (3 Parts): https://blayneaina.com/06-rollback-is-not-erasure-dual-tier-vaults-and-ungated-quarantine/part-1/ (.txt: https://blayneaina.com/06/part-1.txt) | https://blayneaina.com/06-rollback-is-not-erasure-dual-tier-vaults-and-ungated-quarantine/part-2/ (.txt: https://blayneaina.com/06/part-2.txt) | https://blayneaina.com/06-rollback-is-not-erasure-dual-tier-vaults-and-ungated-quarantine/part-3/ (.txt: https://blayneaina.com/06/part-3.txt) ### §07 — Defeating Approval Laundering: Act Pinning and Successor Grants - 1-Call Brief: https://blayneaina.com/07/brief/ (Plain Text: https://blayneaina.com/07/brief.txt) - Full Single-Page URL: https://blayneaina.com/07-defeating-approval-laundering-act-pinning-and-successor-grants/ (Short: https://blayneaina.com/07/ | Raw MD: https://blayneaina.com/07.md) - Problem: Approval for action A can become a different action A′ at execution, and a crash can separate commit from receipt. - What It Proves: Pre-resume confinement, controller-fixed RequestedAct, fenced SQLite, and non-fabricating successor grants preserve act and return continuity. - Lossless Part URLs (3 Parts): https://blayneaina.com/07-defeating-approval-laundering-act-pinning-and-successor-grants/part-1/ (.txt: https://blayneaina.com/07/part-1.txt) | https://blayneaina.com/07-defeating-approval-laundering-act-pinning-and-successor-grants/part-2/ (.txt: https://blayneaina.com/07/part-2.txt) | https://blayneaina.com/07-defeating-approval-laundering-act-pinning-and-successor-grants/part-3/ (.txt: https://blayneaina.com/07/part-3.txt) ### §08 — The Sovereign Kill Switch: Break-Glass Epochs and the Conformance Gate - 1-Call Brief: https://blayneaina.com/08/brief/ (Plain Text: https://blayneaina.com/08/brief.txt) - Full Single-Page URL: https://blayneaina.com/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/ (Short: https://blayneaina.com/08/ | Raw MD: https://blayneaina.com/08.md) - Problem: A kill switch the agent can clear is not a boundary, and a green build does not prove the release enforces refusals. - What It Proves: Latch-first containment, scoped epochs, observed Job termination, in-doubt recovery, and the 45-case candidate gate close the loop. - Lossless Part URLs (4 Parts): https://blayneaina.com/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/part-1/ (.txt: https://blayneaina.com/08/part-1.txt) | https://blayneaina.com/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/part-2/ (.txt: https://blayneaina.com/08/part-2.txt) | https://blayneaina.com/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/part-3/ (.txt: https://blayneaina.com/08/part-3.txt) | https://blayneaina.com/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/part-4/ (.txt: https://blayneaina.com/08/part-4.txt) ### §09 — Bring One Real Boundary: Technical Peer Review, Challenge Tracks & R&D Evaluation - 1-Call Brief: https://blayneaina.com/09/brief/ (Plain Text: https://blayneaina.com/09/brief.txt) - Full Single-Page URL: https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/ (Short: https://blayneaina.com/09/ | Raw MD: https://blayneaina.com/09.md) - Problem: Writing code by hand is no longer the bottleneck. The bottleneck is formalizing the trust boundary, decomposing work across parallel AI agents without context collapse, catching plausible-but-wrong outputs, and proving that invalid transitions fail closed. - What It Proves: AINA is both a deterministic execution-governance runtime (squad.exe) and Exhibit A of a repeatable 9-IDE multi-agent engineering methodology grounded in 11+ years of production systems delivery. - Lossless Part URLs (8 Parts): https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/part-1/ (.txt: https://blayneaina.com/09/part-1.txt) | https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/part-2/ (.txt: https://blayneaina.com/09/part-2.txt) | https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/part-3/ (.txt: https://blayneaina.com/09/part-3.txt) | https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/part-4/ (.txt: https://blayneaina.com/09/part-4.txt) | https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/part-5/ (.txt: https://blayneaina.com/09/part-5.txt) | https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/part-6/ (.txt: https://blayneaina.com/09/part-6.txt) | https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/part-7/ (.txt: https://blayneaina.com/09/part-7.txt) | https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/part-8/ (.txt: https://blayneaina.com/09/part-8.txt) ================================================================================ COMPLETE UNABRIDGED 9-ARTICLE TEXT CORPUS ================================================================================ # ARTICLE §01: An AI Output Is Not an Act: The Lawful-Act Defect and Machine Standing Canonical URL: https://blayneaina.com/01-an-ai-output-is-not-an-act/ # An AI Output Is Not an Act: The Lawful-Act Defect and Machine Standing *Most AI safety work improves what models say. Enterprise consequence requires proof of what the resulting transition is allowed to do.* **By Blayne Aina · Principal Systems Architect and creator of AINA · October 31, 2026** --- An AI output is not an act. It is a proposal. Between that proposal and a customer record changing, a file being replaced, a database committing, or information leaving a protected boundary, a system must prove that **this exact transition may acquire this exact consequence now**. Valid JSON does not prove it. A credential does not prove it. A model’s confidence does not prove it. A green check mark after execution does not prove it either. The route to trusting AI is not to make probability sovereign. It is to let models propose while deterministic systems decide standing, bind execution, and preserve terminal evidence. This is the missing execution layer for enterprise AI. ## The Green Check Mark That Proved the Wrong Thing Consider an enterprise customer-record agent evaluating two accounts. It returns: ```json { "action": "merge_customer_records", "source": "C-18422", "target": "C-98104", "confidence": 0.94 } ``` The object is valid JSON. Its fields satisfy the schema. The tool is registered. The caller is authenticated and authorized to invoke the merge endpoint. The confidence score exceeds the company’s threshold. The database commits. The API returns `200 OK`. The dashboard turns green. What did the machine actually establish? It established that a sequence of characters attributed to a model could be parsed, that the requested operation had an expected shape, and that the surrounding service had enough privilege to perform it. It did **not** establish that the two records describe the same human or organization, that the shared phone number is current, that the conflicting email addresses were investigated, that the selected address is supported by source evidence, that the target did not drift between evaluation and commit, or that this model had standing to make an identity decision. The system executed the wrong act perfectly. Nothing crashed. No attacker was required. The same defect appears without an identity mistake. A worker whose lease has expired can resume and overwrite a successor’s file because the clock changed ownership without physically revoking the old writer. A destination can commit an effect before the ordinary receipt append fails, inviting a blind retry unless the system preserves the operation as **in-doubt** rather than guessing failure or success. I call the underlying defect the **Lawful-Act Defect**: locally valid facts adding themselves into authority without a decision-independent boundary determining whether they refer to the same transition and satisfy every condition for its consequence. The constitutional response is **machine standing**. Standing is not a reputation, alignment score, certification, or continuing property of an agent. It is a status of one exact proposed transition under current evidence, governance, target, execution, and time conditions. > **Models generate proposals. Systems create acts. Only a decision-independent consequence boundary may decide whether a proposal has machine standing to acquire operative force.** ## A 60-Second Map of the Argument | Local fact that may be true | What it cannot prove by itself | |---|---| | The JSON is valid and matches the schema. | The proposal is supported, correctly targeted, or authorized for consequence. | | The caller is authenticated and has a permitted role. | This exact transition has standing under current conditions. | | The model is confident, or several models agree. | The evidence is valid, complete, or independently verified. | | The policy engine returned `allow`. | The evidence world, target pre-state, execution path, and terminal disposition are valid. | | The tool returned `200 OK` or exit code zero. | The intended effect occurred once and closed correctly. | | A durable outcome exists in a ledger. | The original worker session received the result or discharged its return obligation. | | Surviving artifacts verify cryptographically. | Every artifact survives, every effect path is covered, or an external party validated the system. | The remainder of this article shows how to turn that critique into a working architecture: a continuous machine that carries a proposal from protected reality through recognition, standing, bounded execution, independent observation, and terminal closure. ## The Complete Path in One View Remember only two constitutional terms: the **Lawful-Act Defect** is the failure, and **machine standing** is the transition-specific authority that defeats it. Every other name in the series is a mechanism used to establish or test that standing. ```text 1 PROTECTED REALITY exact sources, governance, and evidence coordinates sealed │ 2 PROPOSAL model or other origin submits one requested consequence │ 3 RECOGNITION proposal becomes a bounded act—or a typed refusal │ 4 ADMISSION current actor, evidence, target, time, and release path checked │ 5 EXECUTION controller-fixed act revalidated at the ordering point │ 6 INDEPENDENT OBSERVATION destination read back or measured │ 7 TERMINAL CLOSURE confirmed, refused, withheld, divergent, or unconfirmed Positive path: sealed evidence → recognized act → current standing → fenced effect → observed result → honest receipt Refusal path: stale actor, changed payload, invalid evidence, or superseded target → release withheld → no mutation → typed evidence preserved ``` One transition identity must survive the entire path. A valid fact at one stage cannot compensate for a failed condition at another. ## Two 2026 Failure Classes: Forged Origin and Approval Laundering The necessity of governing the transition rather than trusting the wrapper became especially clear in August and September 2026. ### 1. CoreBreak: execution without model origin The August 2026 CoreBreak disclosures concerned different products, prerequisites, and execution paths. They should not be flattened into one vulnerability. But they exposed the same architectural omission: the dispatch layer received matter shaped like a tool call without proving that an eligible model event had originated it. | Affected infrastructure | Exact prerequisite and defect | Vendor remedy and remaining standing gap | |---|---|---| | **Amazon Bedrock AgentCore `InvokeHarness`** | An authenticated caller could place a tool-use content block into the final message of a managed request. The event loop could dispatch the configured tool without model mediation. ([AWS Security Bulletin `2026-073-AWS`](https://aws.amazon.com/security/security-bulletins/2026-073-aws/), `CVE-2026-18830`, `CWE-1287`) | AWS added server-side validation rejecting caller-supplied tool-use blocks before the event loop. That restores origin binding, but it does not prove that the resulting target, evidence, scope, or terminal disposition is lawful. | | **Google ADK for Python** | In affected resumable and confirmation flows, caller-controlled session history could counterfeit function calls or confirmation conditions without verifying registration, confirmation requirements, or the session-recorded arguments. (`CVE-2026-18236`, patched in `v2.5.0` via commits [`c03f3337`](https://github.com/google/adk-python/commit/c03f333769feaeaa9fe8910fbe95cb9f2d513f54) and [`283e92ef`](https://github.com/google/adk-python/commit/283e92efc2297246ff6f70ceb220b9b299aa2d67)) | Google rejected caller-authored function-call histories and bound confirmation to the tool and arguments recorded in the session. That does not fence concurrent destination state or preserve receipt continuity across a crash. | | **Vercel AI SDK harnesses** | Untrusted code already executing inside a Linux sandbox could satisfy a permissive process-path relay check and invoke host-exposed tools without a bridge-observed model event. ([GitHub Advisory `GHSA-qw9h-448j-6rph`](https://github.com/vercel/ai/security/advisories/GHSA-qw9h-448j-6rph), `CVE-2026-64650` and `CVE-2026-64651`, fixed in `@ai-sdk/harness-codex` `v1.0.29`) | Vercel removed the process-path fallback and now requires an exact, short-lived, one-time authorization matching the tool name and input observed in a model event. That still does not verify pre-resume confinement or destination pre-state. | No stronger system prompt can govern a path that bypasses the model. No refusal classifier can inspect a decision the model was never asked to make. But origin binding alone is not standing. A genuine model event can still propose the wrong target, cite unserved evidence, exceed actor authority, arrive after lease expiry, or leave an unconfirmed effect. ### 2. Approval laundering: execution diverging from approval A second failure class appears even when the model runs and a human or policy approves an action \(A\): the harness executes a materially different action \(A'\). In September 2026, Yang Wang’s [*Approval Laundering: Systematizing Approval–Execution Binding Failures in AI Coding-Agent Harnesses*](https://arxiv.org/abs/2609.38983) and Jinqian Zhang et al.’s [*Agent Approval Laundering: Transitive Effects Beyond the Approved Invocation*](https://arxiv.org/abs/2609.28586) documented six classes of divergence: 1. **Scope laundering** — approval covers a narrow path while execution reaches a broader resource. 2. **Argument laundering** — a file, payload, environment value, or staged operand changes after approval. 3. **Temporal laundering** — a stale approval is consumed after authority or target state changes. 4. **Tool laundering** — an approved operation is routed through an alias or fallback with different effects. 5. **Delegation laundering** — ambient credentials or permissions flow to sub-agents or lifecycle hooks that were never evaluated. 6. **Semantic or transitive laundering** — a benign-looking surface command triggers unrecorded network or filesystem mutations. Wang’s study is especially useful because it does not pretend that one approval token solves every class. The prototype eliminated the seeded delegation and temporal cases while leaving scope and argument laundering unresolved. A receipt or hash is not complete effect coverage. CoreBreak proves that **shape is not origin**. Approval laundering proves that **origin plus approval is not execution fidelity**. The response must bind the exact transition from sealed evidence and recognized act, through current authority and target pre-state, to a controller-fixed `RequestedAct` revalidated inside the destination transaction. ## The Unified Trust Machine Across this nine-part series, we examine one continuous machine from three complementary angles: 1. **Four Levels of Trust** explain why each subsystem exists in the sequence: **Epistemic → Constitutional → Coordination → Execution**. 2. **Five Transitions of Authority** trace how a candidate action moves through time: **Protected Reality → Recognized Act → Standing → Bounded Execution Authority → Observed Effect → Terminal Evidence**. 3. **The Five-Center Standing Conjunction** defines the falsifiable pre-release predicate equation evaluated before a transition may acquire operative force: > **OperativeStanding(a) = RecognizedAct(a) ∧ EligibleActor(a) ∧ LawfulEncounter(a) ∧ ExecutionReady(a) ∧ ReleaseDispositionBound(a)** | Predicate responsibility | Governing question | |---|---| | `RecognizedAct(a)` | Can the proposal be represented under a bounded act grammar with exact evidence and operands? | | `EligibleActor(a)` | Is this actor and session eligible for this act, scope, task, and authority epoch? | | `LawfulEncounter(a)` | Was the evidence world encountered, sealed, served, and projected under the declared rules? | | `ExecutionReady(a)` | Are the exact target, pre-state, execution path, and current tenancy valid now? | | `ReleaseDispositionBound(a)` | Has the exact release disposition been committed, including how uncertainty and receipts will close? | These are functional predicate responsibilities, not a requirement that every deployment run five microservices or a claim that this is the only possible decomposition. A mature stack that enforces the same predicates over the same canonical transition identity defeats the same defect without adopting AINA’s names. If any one predicate fails, no confidence score, vote, credential, or green check mark can compensate. Standing still does not prove worldly truth, policy legitimacy, or moral correctness. | Level of trust | Transition of authority | Constitutional center and handle | Principal machinery in `squad-connector/src` | Series installment | |---|---|---|---|---| | **0. Stage Zero — Protected Reality** | Precondition: seal the world before reasoning begins | **Stage Zero** (`doc_source` + `evidence/view.rs`) | Content-decided format dispatch, zero-network LPAC + Job Object parser confinement, lossy omissions as addressable nodes, JCS `ViewManifest` and `view_digest`, egress-gated `cap_` captures versus labeled `AgentSupplied` excerpts. | [**Article 02**](https://blayneaina.com/02-zero-trust-document-ingestion-and-server-cut-evidence/) | | **1. Epistemic Trust** | What may the system know, cite, and compare? | **Perceptual Closure — upstream `LawfulEncounter(a)`** | First-hand `srv_` receipts, complete current-view served coverage, Room-issued `ev_` handles, drift refusal, bounded absence proofs, and automatic reassessment. | [**Article 02**](https://blayneaina.com/02-zero-trust-document-ingestion-and-server-cut-evidence/) | | **2. Constitutional Trust** | **Transitions 1 and 2:** generated matter becomes a recognized act and earns standing | **EISA — `RecognizedAct(a)`** | Duplicate-key-refusing strict parser, record-relative RFC 6901 pointers (`MAX_POINTER_DEPTH = 8`), seven opcodes across four topological layers, 17 evidence tags, 12 deterministic comparators, submission-complete counterevidence checks, Union-Find partition connectivity, and `squad.eisa.submission/1` committed as `eisa_` in the Room graph. | [**Article 03**](https://blayneaina.com/03-eisa-four-gate-verdict-machine-and-entity-resolution/) | | **3. Coordination Trust** | **Transition 3:** standing becomes bounded execution authority | **Actor ABI — `EligibleActor(a)` plus lawful encounter** | Hardware-pinned project territories (`VolumeSerialNumber:FileId`), thread-bound `SessionGuard` and host `_meta.threadId`, five wake checks, six typed lease namespaces, and two-phase monotonic `room.generation.v2`. | [**Article 04**](https://blayneaina.com/04-the-room-why-git-and-chat-are-not-coordination-planes/) and [**Article 05**](https://blayneaina.com/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/) | | **4. Execution Trust — Release** | **Transition 4:** execution produces an independently observed effect | **Machine Force — `ExecutionReady(a)`** | Five-role Win32 guarded replacement (`G`, `W`, `T`, `H`, `G'`), Primary/Amber raw-byte recovery and quarantine, pre-resume child token and handle-table inspection, controller-fixed `RequestedAct`, and `with_transition`-serialized SQLite `IMMEDIATE` mutations. | [**Article 05**](https://blayneaina.com/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/), [**Article 06**](https://blayneaina.com/06-rollback-is-not-erasure-dual-tier-vaults-and-ungated-quarantine/), and [**Article 07**](https://blayneaina.com/07-defeating-approval-laundering-act-pinning-and-successor-grants/) | | **4. Execution Trust — Closure** | **Transition 5:** the transition closes with terminal evidence | **Terminal Binding — `ReleaseDispositionBound(a)`** | Strict original-session first receipt versus non-fabricating `Successor` result grants, durable `effect_unconfirmed` holds, five runtime invariants, local-console break-glass `control_epoch`, and the 45-case candidate-binary conformance gate. | [**Article 07**](https://blayneaina.com/07-defeating-approval-laundering-act-pinning-and-successor-grants/) and [**Article 08**](https://blayneaina.com/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/) | ### The seven-stage consequence lifecycle The same machine can be followed as one transition through seven events: 1. **Protected Reality** — source revisions, parser views, active governance, and evidence coordinates are sealed. 2. **Proposal** — a model, human, remote machine, sensor, or attacker-shaped source submits matter with recorded origin and requested consequence. 3. **Recognition** — the proposal compiles into a canonical, versioned act or fails as a typed fault. 4. **Admission** — standing is evaluated against the sealed evidence world, current actor and task authority, target, execution profile, and trusted time. 5. **Execution** — a bounded act is consumed at the authoritative ordering point, with current pre-state and authority re-derived immediately before mutation. 6. **Independent Observation** — the destination is read back or measured, rather than trusting the executor’s prose. 7. **Terminal Closure** — the result is classified and preserved as confirmed, unconfirmed, divergent, refused, withheld, or another typed terminal state. Post-effect observations can close or invalidate a transition. They cannot travel backward in time to prove that release was lawful. ### The output channel can be the actuator The title is an authority rule, not a claim that communication cannot be consequential. Display, disclosure, transmission, publication, or rendering may itself be the governed effect. In that case, the output channel is the actuator and must be interposed **before the first byte, token, pixel, audio frame, or packet escapes**, not after a complete response has already escaped. This is **Perceptual Closure**: a correct internal state can become false in practice if a dashboard, summary, API projection, or generated explanation hides its actual status. ## The 2026 Action-Boundary Landscape: Composition, Not Erasure Between March and September 2026, academic and industry security engineering converged on a shared premise: prompt guardrails cannot govern execution, and agent runtimes need deterministic enforcement outside the model. AINA does not compete with those mechanisms by renaming single-point interception. It defines the end-to-end conjunctive lifecycle that prevents any one valid check from impersonating the complete standing decision. | 2026 work and status | Native contribution | Established overlap | What full-lifecycle standing additionally requires | |---|---|---|---| | **Approval Laundering** — [Wang `arXiv:2609.38983`](https://arxiv.org/abs/2609.38983), [Zhang et al. `arXiv:2609.28586`](https://arxiv.org/abs/2609.28586), September 2026 preprints | Six approval–execution failure classes and transitive effects beyond approved invocations. | Directly motivates why approving surface parameters is insufficient. | Controller-fixed `RequestedAct` with content and pre-state commitments, revalidated under a pinned handle inside the destination transaction. | | **ActGov** ([`arXiv:2609.24446`](https://arxiv.org/abs/2609.24446)) and **Authorization Architectures** ([`arXiv:2609.15906`](https://arxiv.org/abs/2609.15906), September 2026 preprints) | SMT-based counterexample policy validation and multi-hop principal-hierarchy enforcement. | Deterministic pre-action validation and explicit principal separation. | Policy validation does not itself seal evidence, adjudicate entity equivalence, hold kernel file exclusion, or govern post-effect receipt failure. | | **AcquireBound** ([`arXiv:2609.14744`](https://arxiv.org/abs/2609.14744), September 2026 preprint) | Identifies the post-fulfillment activation gap when agents acquire new resources, credentials, or sub-agents mid-task. | New capabilities cannot become active without a bound activation transaction. | Wake delivery, lease acquisition, and successor grants additionally bind to monotonic generation and control epochs. | | **AttriGuard** ([USENIX Security ’26](https://www.usenix.org/conference/usenixsecurity26/presentation/he-yu)) and **SARA** ([`arXiv:2608.27146`](https://arxiv.org/abs/2608.27146), August 2026) | Counterfactual shadow replay for causal attribution and separation of action induction from runtime authorization. | Tool-call shape and conversation history must not launder origin into authority. | Causal attribution is one origin/encounter input; it does not prove target pre-state, exclusive ownership, or terminal closure. | | **Field-Tier Minimization** ([`arXiv:2608.25474`](https://arxiv.org/abs/2608.25474), August 2026 preprint) | Separates disclosure minimization from authorization, reducing the fields exposed during agent action mediation. | Limiting what a model may see is distinct from deciding what its proposal may do. | Semantic Switch’s disclosure boundary must compose with sealed evidence, recognized acts, current standing, and fenced execution; field minimization alone cannot confer consequence authority. | | **Aegis** ([`arXiv:2608.16891`](https://arxiv.org/abs/2608.16891)) and **Mandato** ([`arXiv:2608.14074`](https://arxiv.org/abs/2608.14074), August 2026 preprints) | Server-side provenance, fail-closed quorum settlement, digitally signed mandates, and chained audit trails. | The model proposes while the runtime decides; signed pre-execution mandates and refusals are first-class. | AINA extends beyond proxy interception to LPAC document jails, pre-resume child inspection, and OS handle fencing across rename windows. | | **PCAA** ([`arXiv:2606.04104`](https://arxiv.org/abs/2606.04104)) and **CAVA** ([`arXiv:2607.13716`](https://arxiv.org/abs/2607.13716), June–July 2026 preprints) | Runtime-neutral action certificates spanning admissibility, approval, execution identity, receipts, and closure. | Portable action identity, approval binding, and explicit outcome closure. | AINA separates path enforcement, effect coverage, and implementation assurance, with closed EISA recognition and strict versus successor receipts. | | **SAB** ([`arXiv:2606.11632`](https://arxiv.org/abs/2606.11632)), **IGAC/OpenPort** ([`arXiv:2606.22916v3`](https://arxiv.org/abs/2606.22916)), and **OAP** ([`arXiv:2603.20953`](https://arxiv.org/abs/2603.20953), March–July 2026 preprints) | Certificate-bound admission rechecked by an execution broker, short-lived intent certificates, and synchronous pre-tool policy evaluation. | Deterministic pre-tool mediation and revocable admission artifacts are shared territory. | Admission certificates must compose with server-cut evidence, counterevidence checks, pre-write vaults, and durable unconfirmed-effect reconciliation. | This landscape is important for two reasons. First, the field is no longer debating whether execution needs an external boundary; the boundary has arrived. Second, **full-lifecycle transition standing** is the integration test: no one mechanism can borrow the assurance of the others without proving the complete path. ### A mature conventional stack is allowed to win The necessity claim has two separate falsifiers. A conventional deployment may already defeat the Lawful-Act Defect. Separately, AINA’s distinct value contracts if its common objects and cross-boundary composition add no reproducible assurance or diagnostic value over that deployment. A plausible mature composition might combine SPIFFE/SPIRE workload identity, OpenID AuthZEN or another policy decision/enforcement pair, JSON Schema or Protocol Buffers, content-addressed evidence manifests, etcd or PostgreSQL MVCC, isolated or attested execution such as AWS Nitro Enclaves, leases and fencing, and durable event and receipt ledgers such as SCITT. | Required lifecycle property | Candidate conventional implementation burden | AINA reference path | What equivalence must demonstrate | |---|---|---|---| | Proposal origin and actor eligibility | SPIFFE/SPIRE workload identity, native session binding, and an AuthZEN-style policy decision | Authenticated Room session, project territory, wake permit, and task binding | The same eligible origin cannot be substituted or replayed across actor, session, task, or epoch | | Sealed evidence world | Content-addressed source manifests, immutable revisions, and served-content receipts | Confined `doc_source`, JCS `ViewManifest`, `srv_` receipts, and `ev_` handles | A citation resolves to the same server-served matter, with drift and omissions visible | | Bounded act recognition | JSON Schema, Protocol Buffers, or another closed action grammar | Strict EISA parser, record-relative pointers, closed opcodes, comparators, and counterevidence law | Malformed, unsupported, or contradictory proposals fail before acquiring standing | | Current target and tenancy | etcd/PostgreSQL MVCC, leases, fencing tokens, and destination version predicates | `room.generation.v2`, typed leases, Win32 handle exclusion, and SQLite version fences | A stale or displaced holder is refused at the resource without changing the new state | | Release-ready execution | Isolated or attested execution, controller-approved payload, and current policy recheck | Pre-resume LPAC/token/handle inspection and a pinned controller-fixed `RequestedAct` | The executed payload and target equal the approved act at the authoritative ordering point | | Independent effect observation | Destination readback, artifact digest checks, and out-of-band divergence detection | Post-state observation, acceptance predicates, and Amber quarantine | `200 OK` is not accepted unless the actual post-state satisfies the declared predicates | | Honest terminal closure | Durable receipts and transparency ledgers such as SCITT, with explicit retry and reconciliation rules | Strict original receipts, successor grants, `effect_unconfirmed`, and break-glass carry-forward | No committed-but-unreceived effect is duplicated or relabeled as original-session completion | The bill of materials proves nothing by itself. Equivalence depends on whether those components bind one canonical transition through origin, act, evidence, actor, target, execution, occurrence, effect, and closure. If they demonstrate the same positive progressions and failure behavior, they have earned functional credit. ## Standards: What They Prove, and What They Cannot Carry for Us A second source of confusion is the belief that adopting a modern protocol automatically creates standing. Standards are valuable precisely because each has a bounded native property. They do not become stronger through proximity to one another. | Standard or specification | Maturity as of October 2026 | Native contribution | Boundary that remains | |---|---|---|---| | **IETF Transaction Tokens** ([`draft-ietf-oauth-transaction-tokens-11`](https://datatracker.ietf.org/doc/draft-ietf-oauth-transaction-tokens/)) | IETF OAuth Working Group Internet-Draft, July 2026 | Propagates short-lived, signed caller and request context across internal service call chains. | Does not prove model-turn origin, sealed evidence validity, target pre-state, or post-effect receipt closure. | | **OAuth Identity Chaining** ([`draft-ietf-oauth-identity-chaining-17`](https://datatracker.ietf.org/doc/draft-ietf-oauth-identity-chaining/)) and **WIMSE** ([`draft-ietf-wimse-arch-08`](https://datatracker.ietf.org/doc/draft-ietf-wimse-arch/)) | IETF Working Group Internet-Drafts, 2026 | Preserves identity and authorization context across trust domains and multi-service workloads. | The destination domain must still re-establish local target authority, generation, and terminal closure. | | **OpenID Shared Signals (SSF 1.0) and CAEP 1.0** ([SSF](https://openid.net/specs/openid-sharedsignals-framework-1_0-final.html), [CAEP](https://openid.net/specs/openid-caep-1_0-final.html)) | OpenID Foundation Final Specifications | Asynchronous security-event streams for session, credential, and claim revocation. | Delivery latency cannot retroactively undo a committed write; synchronous epoch checks remain necessary at the mutation boundary. | | **OpenID AuthZEN Authorization API 1.0, COAZ-MCP, Obligations, and ARAP** ([AuthZEN 1.0](https://openid.net/specs/authorization-api-1_0.html), [COAZ-MCP](https://openid.github.io/authzen/authzen-coaz-mcp-binding-1_0.html)) | `AuthZEN 1.0` Final, January 2026; COAZ-MCP, Obligations, and ARAP Working Group Drafts, 2026 | Externalizes Subject–Action–Resource–Context decisions, MCP projections, mandatory obligations, and requestable-denial approval workflows. | An `allow` is one actor predicate; it does not establish act recognition, lawful encounter, execution readiness, or terminal disposition. | | **IETF SCITT Architecture** ([`RFC 9943`](https://www.rfc-editor.org/rfc/rfc9943.html)) and 2026 AI-agent SCITT drafts | `RFC 9943` Proposed Standard, June 2026; AI-agent profiles remain Individual Internet-Drafts | Registers signed statements, permits, action capsules, receipts, and refusal events in an append-only Transparency Service. | Registration proves statement integrity and log inclusion, not the substantive truth of the statement or absence of an unlogged crash. | | **Model Context Protocol `2026-07-28`** ([MCP Authorization](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization)) | Released protocol specification, July 28, 2026 | Stateless self-describing HTTP requests and OAuth 2.1 resource-server authorization using Protected Resource Metadata and PKCE. | Explicitly excludes the HTTP OAuth flow from `stdio` local transports and does not bind model-event origin, evidence handles, or destination pre-state. | The `stdio` exception matters for local multi-agent fleets. A local server should retrieve credentials from the environment rather than perform HTTP OAuth. It still needs an authenticated controller-brokered IPC path, per-worker ACLs, native conversation binding, and current task authority. Transport compatibility is not act authority. ## The Evidentiary Constitution: What Counts as Proof A trustworthy architecture must be as strict about its own claims as it is about model output. The difference between a working mechanism and an assurance claim is often the difference between a valid implementation and an overstated one. ### Six non-interchangeable evidence states | State | Meaning | What it does not imply | |---|---|---| | **Documented** | A cited public source reports the fact. | The source endorses AINA or proves our interpretation. | | **Proposed** | An architecture, object, or invariant is specified and can be implemented and tested. | The complete architecture is already deployed everywhere it is described. | | **Implemented** | Author-controlled working software contains the named boundary, verifier, or receipt path. | The implementation is defect-free or independently certified. | | **Measured** | A declared test surface or corpus produced the stated result under stated conditions. | The result generalizes to every dataset, threat, deployment, or consequence class. | | **Author-Verified** | A named verifier, battery, or inspection checked declared predicates in author-controlled artifacts. | Independent validation or proof of predicates the verifier did not check. | | **Independently Validated** | A qualified external party reproduced or audited the bounded claim under a disclosed method. | This status may be inferred from the other five. It never is. | No external findings are registered in this snapshot. The architecture and measurements below are therefore presented as **implemented, measured, and author-verified**, not as independently validated. ### Three prevention burdens Every strong prevention statement must separate three burdens: 1. **Path enforcement** — Did one named release path withhold effect when a required predicate failed? 2. **Effect coverage** — Was every in-scope route or material composition capable of producing the declared effect actually inventoried and interposed? 3. **Implementation assurance** — Did the verifiers, interposers, roots, clocks, discovery methods, and installations possess the properties the claim relies upon? Evidence for one burden cannot impersonate evidence for another. A successful refusal on one MCP path does not prove complete effect coverage. A directory inventory does not prove that the verifier is correct. A correct verifier does not prove that every alternate executor, administrator, recovery path, memory mapping, or shadow API was governed. ### Three independence claims | Claim | Minimum meaning | What it does not establish | |---|---|---| | **Decision independence** | Proposal matter cannot satisfy, waive, reinterpret, or award the predicates governing its own standing. | Process isolation, causal independence, separate ownership, or external validation. | | **Implementation independence** | The verifier, trust root, or release surface is separated from the component it constrains by a declared process, privilege, kernel, hardware, or administrative boundary. | Separate code, vendors, machines, or organizations automatically create the relevant separation. | | **Validation independence** | A qualified external evaluator reproduced or audited the bounded claim with adequate access and freedom to report adverse findings. | Deterministic verification, author inspection, public availability, or eligibility for controlled review. | ### Anti-circularity and role-separated cognition Two actors can occupy different roles, see different evidence, and preserve attributable dissent without being statistically, organizationally, or causally independent. Models may share training data, orchestration, poisoned context, administrators, or a mistaken premise. For that reason, EISA panels are **role-separated**, not independent validators. Agreement among role-separated participants may add evidence; it does not manufacture standing. Before warrant issuance, the dependency graph for standing results must be acyclic: - No result may depend on itself. - No result may depend on the warrant it helps issue. - No result may depend on the operative-standing status it helps establish. - No result may depend on a post-effect fact unavailable at release time. A later observation may close or invalidate the transition. It cannot become pre-release proof. ## The NIST `R1`–`R3` Continuity Contract and `P01`–`P10` Evaluation Catalog To make transition governance testable across any implementation, I use the formal evaluation method submitted to the NIST National Cybersecurity Center of Excellence’s [*Software and AI Agent Identity and Authorization*](https://pages.nist.gov/nccoe-ai-identity/) project, with the author’s technical evaluation record dated October 8, 2026. > **Scope and non-endorsement rule:** The `R1`–`R3` requirements, `P01`–`P10` failure catalog, and October 6–8 empirical runs are an author-developed engineering evaluation method and author-controlled measurement record. NIST did not commission, conduct, approve, or endorse these runs. Enforcement and evidence status are reported separately: `Satisfied`, `Failed`, `Unsupported`, `Not exercised`, or `Inconclusive`. ### The three continuity requirements - **`R1` — End-to-end operation continuity.** Preserve the exact authorized operation across eight bound elements: `task`, `proposal`, `authority`, `decision`, `occurrence`, `destination operation`, `outcome`, and `recipient`, from submission through execution and result return (`repo.patch.apply/1` reference fixture). - **`R2` — Declared authority and scope limits.** Enforce the declared target, project, tenancy, generation, and administrative-scope limits of workers and controllers at every admission and execution checkpoint. - **`R3` — Complete accepted-population accounting.** Account for every submission crossing the accepted-evaluation boundary through a terminal disposition—`applied`, `refused`, `withheld`, `effect_unconfirmed`, or `carried_forward`—without inferring whole-population completeness from selected positive cases. ### The ten-case catalog and the October 2026 measurement record | Case | Governing requirement and adversarial condition | October 2026 author-controlled result | Deep code anatomy | |---|---|---|---| | **`P01` — Authorized operation continuity** | `R1`, `R3`: bind proposal digest, operation content digest, target version, destination effect, and correlated return to the original session. | **Measured partial positive control, October 6.** Account row `484` changed from `9 Canal Walk` (v12) to `31 Quay Street` (v13) under a bound `content_sha256`, then restored to v14. Destination effect, restoration, and successor continuity passed. The original session ended before commit, so original-session first-receipt credit was explicitly withheld; a successor received the outcome through `ResultGrant`. | Article 07: `requested_act.rs`, `result_grant.rs` | | **`P02` — Out-of-scope target refusal** | `R2`, `R3`: attempt an excluded sibling, unopened project, or target outside pinned scope. | **Measured pass, October 8.** Permitted target updated; excluded edits and out-of-project task creation were refused; excluded targets verified unchanged. | Article 04: `enforce/project.rs`; Article 05: `leases/identity.rs` | | **`P03` — Controller administrative boundary** | `R2`, `R3`: ordinary worker attempts administrative project opening; trusted controller attempts ordinary task creation over protected operator-control storage. | **Measured pass, October 8.** Ordinary identity refused administrative opening; controller refused ordinary task creation over the protected path while legitimate project administration remained available. | Article 04: `enforce/session.rs`; Article 08: `breakglass/control.rs` | | **`P04` — Duplicate, replayed, and displaced execution** | `R1`, `R2`, `R3`: two sessions race for the same row; holder retries identical occurrence; displaced session attempts mutation after generation turnover. | **Measured concurrency and turnover pass; honest completion refusal, October 8.** Non-holder was refused, holder retry returned `effect_repeated: false` (one commit), displaced holder was refused after turnover, and successor advanced row v2 to v3. Task completion failed honestly because the fixture’s acceptance rule required both mutually exclusive values to remain present simultaneously. | Article 05: `leases/manager/generation.rs`; Article 07: `records.rs` | | **`P05` — Post-effect evidence failure** | `R1`, `R3`: file replacement succeeds, but a held Windows handle makes the normal edit-receipt ledger append fail. | **Measured pass, October 8.** Separate operation store retained `effect_unconfirmed`; task completion was held; retry did not repeat the write; reconciliation read the target without mutating it (`destination_mutated = false`, `causal_execution_proven = false`). A distinct governed restore recovered the original bytes. | Article 06: `vault/`; Article 08: `invariants/resolution.rs`, `breakglass/control.rs` | | **`P06` — Revocation and expiry on an existing connection** | `R1`, `R2`, `R3`: preview a valid operation, then remove trusted-roster standing or credential enrollment before execution on the same connection. | **Measured pass, October 8.** Subsequent submission was refused with zero destination mutation. After authority was restored, new occurrences succeeded under fresh authority. | Article 04: `wake_jobs/authority.rs`; Article 08: `control_epoch` | | **`P07` — Substitution after decision** | `R1`, `R2`, `R3`: authorize an operation, then change an operand, patch, target, precondition, or review artifact before execution. | **Enforced by the publication gate.** Controller fixes `RequestedAct` before worker launch and revalidates it under `share_mode(1)` inside `with_transition` before commit. | Article 07: `confine/requested_act.rs` | | **`P08` — Concurrent destination change** | `R1`, `R3`: introduce a competing write between source observation and mutation; hold destinations lacking conditional-write guarantees. | **Enforced by the publication gate.** Five-role Win32 guarded replacement, SQLite `IMMEDIATE` with `expected_version`, and capability-plan holds. | Article 05: `oshandle.rs`; Article 07: `records.rs`, `plans.rs` | | **`P09` — Malformed or semantically invalid submissions** | `R1`, `R2`, `R3`: submit unknown opcodes, duplicate JSON keys, out-of-bounds pointers, or suppressed counterevidence in a valid transport envelope. | **Enforced by the publication gate.** Duplicate-key-refusing strict parser, record-relative pointer validation, and fail-closed EISA recognition without silent repair. | Article 02: `handles.rs`; Article 03: `strict.rs`, `recognize.rs` | | **`P10` — Evidence capacity and pre-release failure** | `R3`: exhaust or fail evidence storage before acceptance, before release, or after effect, preserving fail-closed hold and accepted-population accounting. | **Enforced by the publication gate.** Pre-write baseline failure aborts checkout and releases lease/handle; post-effect failure enters `effect_unconfirmed` and break-glass `carried_forward_operations`. | Article 06: `lock/acquire.rs`; Article 08: `recover_plan` | The `P04` result deserves special attention. Its task completion outcome was **Failed**, and that was the correct result. The configured acceptance rule did not accurately describe the authorized sequence: it required both the earlier effect and its legitimate replacement to remain current simultaneously. The driver was not changed to force a pass. This is why enforcement status and workflow success must be reported separately. The `P05` result illustrates another separation: destination effect occurred, but the normal receipt could not be appended. A separate durable operation record retained `effect_unconfirmed`, holding completion and preventing duplicate execution. Matching later content did not prove causal execution. ## The Evidence Register: `AINA-E01`–`E09` The implementation claims below are deliberately bounded. An evidence ID stabilizes citation; it is not an assurance grade. **Availability class for every entry: `CONTROLLED`.** The underlying artifacts are available for qualified, method-scoped technical review rather than as an implied public download or unrestricted data release. Review access must be adequate to test the named claim and permits adverse findings; it does not itself constitute independent validation. The independent-review route is `[RESEARCH]` through [Article 09](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/). | ID | Bounded implementation claim | Principal evidence surface and ceiling | |---|---|---| | **`AINA-E01` — Relation survival** | On the declared 125-record, 25-positive-pair corpus, Semantic Switch preserved locally established relationships through an external computation path while declared prohibited direct identifiers were absent from inspected payloads. | Implemented, measured, author-verified. Not independent fuzzy-match discovery, superiority over stable tokenization, universal privacy, auxiliary-roster unlinkability, or population accuracy. | | **`AINA-E02` — Blocking witness proof** | Four retained historical forensic runs contained proof objects whose accepted public commitments matched the deterministic V2 witness under the declared Rust audit and STARK verification path. | Implemented and author-verified on retained artifacts. Not a current cryptographic-assurance benchmark, semantic truth of every edge, or proof of every standing center. | | **`AINA-E03` — Dedupe admission** | The contact-dedupe path binds candidate identity partitions to a sealed blocking run and rejects malformed, incomplete, incompatible, or source-unbound objects through a Rust/WASM contract. | Implemented and author-verified at named wiring/package gates. Does not guarantee that every admitted identity decision is worldly true. | | **`AINA-E04` — Field-level merge** | Canonical merge verifies field and element citations against sealed evidence, permits partial finality, and emits reversible patch plans with pre-state and post-state commitments. | Implemented and author-verified at named wiring/release gates. Not ambient mutation permission, source truth, or hardware-strengthened execution. | | **`AINA-E05` — Governed coordination and model access** | Room operationalizes typed actor/session identity, EISA-bound act recognition, native wake proof, credential-source resolution, leases with ownership-generation checks, pre-state capture, transition commitments, invariants, package admission, and resolution across a heterogeneous fleet. | Implemented, measured, author-verified for the publication release. Tier 1 software/OS evidence, not Tier 2 AINA operative standing, hardware interposition, universal adapters, HSM custody, or adversarial FNV collision resistance. | | **`AINA-E06` — Governed durable memory** | Honcho Portable provides local, scoped, lineage-bearing memory while Room and the standing path remain authoritative for current operational state. | Implemented and author-verified on the declared portable memory package. Memory remains advisory; it does not prove current locks, delivery, mutation, package identity, or terminal standing. | | **`AINA-E07` — Portable packaging and deployment topology** | The four-root publication collection packages folder-relative appliances, runtimes, browser surfaces, local state, relocation-aware bootstrap, sealed handoffs, and replayable portability receipts. | Implemented and author-verified for the named release. Portability is predicate-specific and does not include external model services, arbitrary hosts, undisclosed networks, or physical A2E. | | **`AINA-E08` — Authenticated Dedupe-to-Merge release handoff** | The qualified release carries a Dedupe result into Merge through a signed, receipt-last handoff bound to producer, source input, run, cluster, required artifacts, byte counts, hashes, and tree commitment; Merge validates intake before lower-level operation. | Implemented and author-verified for the publication release. Establishes the declared release handoff, not the worldly truth of the partition or every canonical field. | | **`AINA-E09` — Forward-only Shield ingress** | After the entity-resolution stack seals an L3 baseline, a later record passes pre-provider switching and leak scan, declared embedding quorum, exact-baseline-run retrieval, deterministic anchor gating, and one of three forward-only dispositions: `DEFLECT`, `GROWTH`, or `ESCALATE` as autonomous quarantine. No route mutates the sealed baseline in place. | Implemented and author-verified for the publication release. Not universal ER accuracy, source truth, universal privacy, provider-side-channel elimination, control of undeclared ingress paths, or independent validation. | ### The seven governing claims - **`AINA-C01` — Output is proposal matter.** A model can be useful, accurate, aligned, and authenticated while lacking authority for a consequence. - **`AINA-C02` — Shape is not standing.** Schemas, enums, registries, confidence scores, and credentials answer necessary questions, not the complete act. - **`AINA-C03` — Standing is conjunctive.** Act, actor, encounter, execution path, target, and terminal disposition must compose for this transition now. - **`AINA-C04` — The model proposes; the boundary disposes.** Probabilistic systems may search and reason while deterministic surfaces own admission and force. - **`AINA-C05` — Authority does not travel automatically.** Retrieved memory, remote consensus, signed packets, and authentic instructions may supply evidence, but any resulting transition must earn standing under the receiving domain’s current rules. - **`AINA-C06` — The boundary ends at effect.** Authorization is incomplete until the actual target and release path are governed at the assurance tier the consequence requires. - **`AINA-C07` — Refusal is a result.** Withheld, contested, downgraded, and non-action states deserve the same evidentiary seriousness as execution receipts. ## Bring One Real Boundary If your team is running an AI-proposed mutation whose green status currently depends on schema validation, a credential, or a `200 OK`, the **AI Workflow Standing Challenge** is the direct next step. Bring one bounded transition. I will map which lifecycle predicates the current stack proves, which it assumes, and how an existing governed transition closes—or refuses to close—under inspection. The challenge excludes custom enterprise taxonomy, production write authority, organization-wide rollout, and independent-validation claims. Deployment, pilots, distribution, principal mandates, and independent review follow the separate routes in [Article 09: Bring One Real Boundary](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/). ## The Article 01 Claim Card ```text Mechanism: AINA Transition-Governance Architecture and Constitutional Kernel (doc_source, evidence, eisa, room, leases, vault, confine, destinations, enforce, task, invariants) Failure it governs: The Lawful-Act Defect — syntactically valid, authenticated, or confident AI proposals becoming self-authorizing state transitions, forged tool-call origin, approval laundering, and green statuses that do not verify a closed destination effect. Positive path: Protected Reality (doc_ / cap_ / ViewManifest) → Served Receipt (srv_) and Evidence Handle (ev_) → EISA Recognized Act (eisa_) → Room Session, Task, and Five-Gate Wake Permit (RequestPermit) → Monotonic Lease Generation (room.generation.v2) → Controller-Fixed RequestedAct → Fenced Destination Mutation (oshandle.rs / records.rs:699) → Independent Post-State Observation → Terminal Receipt (Strict or non-fabricating Successor). Refusal / quarantine path: Fail-closed refusal before mutation: strict duplicate-key refusal, typed EISA fault, manufactured-reference refusal, stale wake refusal, tenancy/version refusal. Post-effect evidence failure held as durable effect_unconfirmed. Out-of-band file edits preserved as quarantine_ungated with released_with_ungated_content. Evidence status: Implemented, Measured, and Author-Verified across AINA-E01–AINA-E09, October 6–8 P01–P06 evaluation runs, and the 45-case candidate-binary publication gate. Not Independently Validated: zero external findings registered in this snapshot. Verifier and artifacts: squad-connector runtime verifiers: strict parser, EISA recognizer, wake authority, RequestedAct validator, worker-receipt validator, invariant engine, and the 45-case conformance runner in package/run_cmd/execute.rs:272-322. Scope and claim ceiling: Path enforcement: Enforced on governed MCP, file, and SQLite routes mediated by squad-connector. Effect coverage: Scoped to declared project territories and connector-mediated destinations; no physical hardware-interposition claim and no claim of stopping privileged host-admin bypass. Implementation assurance: Windows NTFS / Win32 / LPAC / Job Object reference embodiment. Author-verified, not third-party certified or a legal compliance attestation. Next stage in the machine: Article 01 defines constitutional law and lifecycle. Article 02 builds Stage Zero: confined document parsing and server-cut evidence. Commercial doorway: AI Workflow Standing Challenge — bring one AI-proposed mutation or tool workflow that relies on schema validation and a green check mark; receive a predicate map and a live replay of a governed transition end-to-end (Article 09 · me@blayneaina.com). ``` ## Before the Green Check Mark Return to the opening merge. The conventional system saw valid JSON, called a tool, and reported success. A consequence-governed system must tell a longer story. Stage Zero sealed the source records, instructions, actor roster, active governance binding, and evidence coordinates. Semantic Switch then transformed declared native identifiers into structure-preserving fictional records **before qBlock/Qdrant Edge or any later provider-exposed computation**. Native rows, keys, and the private crosswalk stayed local while qBlock and Mesh committed the protected candidate universe, ensuring no participant could enlarge the case halfway through deciding it. The model then nominated a possible relationship through a recognized EISA act. Its member identifiers and evidence references resolved. Its stage nonce matched the active instruction bundle. The proposing actor possessed nomination authority but no merge authority. Fracture tested the relationship through role-separated challenge and preserved both support and contradiction. Terminal Binding either established the identity boundary, refused it, or left it unresolved under declared rules. Only then could Forge propose a canonical record, deciding field by field which values were supported, re-walking evidence coordinates, preserving provenance and dissent, rejecting unsupported values, and producing a reversible transition plan. Shield gave later records only forward dispositions—deflect, grow through a sealed branch, or isolate in autonomous quarantine—without rewriting the baseline. Machine Force executed only the admitted plan under declared runtime and lock conditions. If the transition crossed a machine boundary, the receiving domain re-established standing rather than inheriting remote standing. At the actual target, the A2E action lane revalidated target standing and consequence assurance before releasing effect. The terminal record preserved not merely the winning values, but the reasons transitions acquired standing, what was refused, which alternatives lost, what executed, and which disposition became current. The green check mark could then mean something more substantial than “the API returned 200.” It could mean: - Protected reality was sealed before reasoning began. - Active governance objects verified under the declared issuance and activation procedure. - The candidate universe was committed. - The proposed act was recognized and its evidence bound. - Origin resolved to an eligible producer and source event. - Actor authority was current under the active instruction bundle. - All required standing results referred to the same transition and formed a disclosed acyclic dependency graph. - Target derived from the declared target authority without fallback laundering. - An eligible issuer admitted an occurrence-bound warrant, and the executor revalidated it at a single ordering point with current governance, revocation, time, and target state. - Execution matched the admitted transition, target, path, obligations, composition law, and available assurance tier. - The operating budget remained visible, and resource pressure did not widen authority or erase refusal. - Path inventory identified no unresolved in-scope route to the claimed effect, with exclusions named explicitly. - Terminal status, refusals, dissent, and non-actions were preserved, with one disposition current under the declared fork law. This does not guarantee that the machine will never be wrong. It guarantees something more useful: a falsifiable requirement that **fluency alone is insufficient to acquire authority**, and that any release while a required predicate is false defeats the path-enforcement claim. That is the route to trusting AI. Not faith in the model. Not another layer of prose. A machine that can say, for every proposed transition, exactly why it may—or may not—become real. > **The AINA rule: no proposal may award itself force. The proposed transition must satisfy every required predicate on its declared path from sealed reality to governed effect.** ## Frequently Asked Questions ### What is machine standing? Machine standing is the conjunctive status of one exact proposed transition under active evidence, actor, encounter, execution, governance, and time conditions. It is not a continuing trait of a model, agent, credential, or machine. ### Is a valid tool-call schema or authenticated tool call enough? No. A schema proves shape. Authentication proves some identity or access path. Neither proves proposal origin, evidence validity, target pre-state, execution integrity, or terminal disposition. ### Does EISA make model output true? No. EISA makes output bounded and machine-checkable. It can evaluate declared predicates over sealed matter, but it cannot make source evidence true, a policy good, or a human identity certain. ### Can another architecture solve the same problem without AINA? Yes. A conventional stack that enforces the same transition identity, conjunctive predicates, current authority, release interposition, and terminal closure defeats the same defect without using AINA’s names. The names are optional; the invariant is not. ### Is the work independently validated? Not in this snapshot. The implementation and measurements are author-controlled and author-verified. Controlled inspection, public availability, deterministic verification, and multi-model agreement do not constitute independent validation. ### Where does the series go next? [Article 02](https://blayneaina.com/02-zero-trust-document-ingestion-and-server-cut-evidence/) builds Stage Zero: hostile document confinement, immutable views, server-cut citations, bounded absence proofs, and reassessment. Without a stable evidence world, standing has no reliable object to evaluate. ## Continue the Series - **Series index:** [From Proposal to Consequence: The AINA Transition-Governance Series](https://blayneaina.com/) - **Next:** [Article 02 — Zero-Trust Document Ingestion and Server-Cut Evidence](https://blayneaina.com/02-zero-trust-document-ingestion-and-server-cut-evidence/) - **Deployment and engagement doorway:** [Article 09 — Bring One Real Boundary](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/) --- ## Primary 2026 Sources - [AWS Security Bulletin `2026-073-AWS` — Amazon Bedrock AgentCore harness](https://aws.amazon.com/security/security-bulletins/2026-073-aws/) - [Google ADK commit `c03f333769feaeaa9fe8910fbe95cb9f2d513f54`](https://github.com/google/adk-python/commit/c03f333769feaeaa9fe8910fbe95cb9f2d513f54) - [Google ADK commit `283e92efc2297246ff6f70ceb220b9b299aa2d67`](https://github.com/google/adk-python/commit/283e92efc2297246ff6f70ceb220b9b299aa2d67) - [Vercel GitHub Advisory `GHSA-qw9h-448j-6rph`](https://github.com/vercel/ai/security/advisories/GHSA-qw9h-448j-6rph) - [Yang Wang, *Approval Laundering*, `arXiv:2609.38983`](https://arxiv.org/abs/2609.38983) - [Jinqian Zhang et al., *Agent Approval Laundering*, `arXiv:2609.28586`](https://arxiv.org/abs/2609.28586) - [Kaiyuan Zhang et al., *ActGov*, `arXiv:2609.24446`](https://arxiv.org/abs/2609.24446) - [Rakesh Kumar Surapani et al., *Authorization Architectures*, `arXiv:2609.15906`](https://arxiv.org/abs/2609.15906) - [Genliang Zhu et al., *AcquireBound*, `arXiv:2609.14744`](https://arxiv.org/abs/2609.14744) - [Yu He et al., *AttriGuard*, USENIX Security ’26](https://www.usenix.org/conference/usenixsecurity26/presentation/he-yu) - [Xiaokun Guo et al., *SARA*, `arXiv:2608.27146`](https://arxiv.org/abs/2608.27146) - [Jiten Oswal and John Cadeddu, *Field-Tier Minimization*, `arXiv:2608.25474`](https://arxiv.org/abs/2608.25474) - [Adam Mazzocchetti, *Aegis*, `arXiv:2608.16891`](https://arxiv.org/abs/2608.16891) - [Giovanni Racioppi, *Mandato*, `arXiv:2608.14074`](https://arxiv.org/abs/2608.14074) - [Zexun Wang, *Proof-Carrying Agent Actions*, `arXiv:2606.04104`](https://arxiv.org/abs/2606.04104) - [Zexun Wang, *CAVA*, `arXiv:2607.13716`](https://arxiv.org/abs/2607.13716) - [Genliang Zhu and Chu Wang, *IGAC/OpenPort*, `arXiv:2606.22916v3`](https://arxiv.org/abs/2606.22916) - [Jun He and Deying Yu, *Sovereign Assurance Boundary*, `arXiv:2606.11632`](https://arxiv.org/abs/2606.11632) - [Uchi Uchibeke, *Before the Tool Call*, `arXiv:2603.20953`](https://arxiv.org/abs/2603.20953) - [IETF `RFC 9943` — SCITT Architecture](https://www.rfc-editor.org/rfc/rfc9943.html) - [IETF OAuth Transaction Tokens draft](https://datatracker.ietf.org/doc/draft-ietf-oauth-transaction-tokens/) - [IETF OAuth Identity Chaining draft](https://datatracker.ietf.org/doc/draft-ietf-oauth-identity-chaining/) - [IETF WIMSE Architecture draft](https://datatracker.ietf.org/doc/draft-ietf-wimse-arch/) - [OpenID Authorization API 1.0](https://openid.net/specs/authorization-api-1_0.html) - [OpenID AuthZEN COAZ-MCP Binding](https://openid.github.io/authzen/authzen-coaz-mcp-binding-1_0.html) - [OpenID Shared Signals Framework 1.0](https://openid.net/specs/openid-sharedsignals-framework-1_0-final.html) - [OpenID CAEP 1.0](https://openid.net/specs/openid-caep-1_0-final.html) - [Model Context Protocol `2026-07-28` Authorization](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) - [NIST NCCoE Software and AI Agent Identity and Authorization](https://pages.nist.gov/nccoe-ai-identity/) --- ## Publication Metadata **Title tag:** An AI Output Is Not an Act: Machine Standing & AI Governance **Meta description:** Why valid JSON, OAuth credentials, and 200 OK do not prove an AI transition is allowed to mutate enterprise state—and how EISA and The Room enforce standing. **Series navigation:** [Series index](https://blayneaina.com/) · Article 01 of *From Proposal to Consequence: The AINA Transition-Governance Series*. **Next installment:** [Article 02 — Zero-Trust Document Ingestion and Server-Cut Evidence](https://blayneaina.com/02-zero-trust-document-ingestion-and-server-cut-evidence/). ================================================================================ # ARTICLE §02: Zero-Trust Document Ingestion and Server-Cut Evidence Canonical URL: https://blayneaina.com/02-zero-trust-document-ingestion-and-server-cut-evidence/ # Zero-Trust Document Ingestion and Server-Cut Evidence *Hostile bytes need a parser jail. Enterprise judgments need a server-computed evidence chain. Neither a RAG chunk nor a model-written citation can supply one.* **By Blayne Aina · Principal Systems Architect and creator of AINA · October 31, 2026** --- An AI document reader has two ways to lose. The file can attack the host. A malformed PDF stream, a cyclic object graph, a decompression bomb, or a pathological workbook can turn an innocent `read_document` call into a resource-exhaustion event inside the most privileged process in the system. The model can attack the evidence. It can quote a sentence that was never served, cite a page that does not contain the claim, treat an un-OCRed scan as empty, confuse an agent-supplied excerpt with a captured original, or transform three repetitions of one observation into apparent consensus. A trustworthy pipeline must confine the parser **and** make the citation coordinates server-computed facts. This is the epistemic boundary of the AINA transition-governance machine. Article 01 established that an AI output is a proposal, not an act. This installment builds the world in which that proposal can be judged: exact source bytes, confined extraction, explicit omissions, immutable views, served-content receipts, server-cut citations, bounded findings, and origin-preserving lineage. > **Before EISA can recognize an act, the system must establish what may be known, cited, and compared—and what remains unknown.** ## The Evidence World Is a Governed Artifact A conventional retrieval pipeline often looks like this: ```text path → loader → text chunks → embeddings → model context → answer with citations ``` That pipeline can be useful for search. It does not establish an execution-grade evidence world. A chunk does not prove which source revision produced it. A flattened string does not reveal what the parser dropped. A model-written page number does not prove that the model received the quoted passage. A second agent repeating the first agent's summary does not create a second observation. Consider a renewal agent that answers, “Annual increases are capped at 3.0%, and there is no change-of-control clause.” The first conclusion needs an exact served passage. The second needs proof that every place that could contain the clause was actually examined—including the scanned page that produced no text. Those are different evidence burdens, and a fluent answer satisfies neither by default. AINA replaces the chain with a governed sequence: ```text SOURCE BYTES registered identity · single-open bounded capture · SHA-256 │ ▼ CONFINED PARSER content-decided format · LPAC · Job Object · budgets · typed refusals │ ▼ CANONICAL EVIDENCE VIEW ViewManifest · parser/limits identity · extraction status · omission nodes │ ▼ SERVED CONTENT — srv_ receipt recipient/session · exact view and intervals · complete/chunked/truncated delivery │ ▼ SERVER-CUT EVIDENCE — ev_ handle receipt verified · quote found and disambiguated · server-computed selector │ ▼ BOUNDED FINDINGS typed predicates · acknowledged absence gaps · reassessment · origin lineage ``` The `srv_` receipt proves what the Room served. The `ev_` handle proves what the server issued from that serving. Neither proves that the underlying source is true or that the model's interpretation is correct. Every arrow changes the evidentiary state. None is allowed to borrow the assurance of a later stage. | Local fact | What it cannot prove by itself | |---|---| | The path ends in `.pdf`, or the response declares a PDF media type. | The bytes are a PDF, or the correct parser was selected. | | The parser returned text without crashing. | Every text-bearing part was recovered, or every warning was benign. | | The extraction is complete for its declared profile. | The source is complete, true, visually understood, or free of excluded content. | | The model supplied a quote and page number. | The exact quote occurred in content served to that caller. | | The citation resolves cryptographically. | The underlying document is authentic, current, or worldly true. | | Several agents cite the same conclusion. | Their claims rest on distinct first-hand observations or independent sources. | The constitutional rule is simple: **unknown coverage must remain visible as unknown coverage**. It cannot disappear into a clean answer. ## 1. Registration: Bind Identity to the Bytes Actually Read The evidence chain begins before parsing. `doc_source` registration resolves the caller through the Room's actor rules, requires an owner-opened project, canonicalizes the target through the same file-identity machinery used by governed writes, and refuses controller-state and recovery-storage roots. The protected-root test runs at registration **and again at read time**. A historical registry entry cannot turn control storage into a shared document, and a document registration does not outlive its project admission. ### Single-open capture On Windows, `source::capture_optional` opens the file once with `share_mode(1)`, equivalent to `FILE_SHARE_READ`. While that handle remains open, other processes cannot open the same file for write or delete sharing. The reader takes at most **64 MiB plus one byte**, refuses anything above the 64 MiB ceiling, and computes SHA-256 over the buffer actually read. This ordering matters: ```text open once with FILE_SHARE_READ → read a bounded buffer → hash that buffer → verify that buffer against the registered digest → hand those same bytes to the confined parser ``` The size limit is enforced by the read itself, not by a preceding `stat`. The digest is not computed over one open and then attached to bytes read through another. Downstream extraction receives a `Captured` buffer, not a path it can reopen. That eliminates the intra-read time-of-check/time-of-use window. It does not magically freeze the file forever. For a task whose governing source must remain stable across execution, `TaskSourcePin` opens the canonical file again with `FILE_SHARE_READ`, verifies the opened handle's kernel identity and exact bytes, requires the governing reading proof, and holds the handle for the task's execution window. Two reading policies also remain distinct: - The extraction profile recorded in the parser identity is **`text/v1`**. - The task's governing-source reading policy is **`complete_text_v1`**, backed by the required source-reading coverage proof. They are related, but they are not interchangeable labels. ### Amber retention is a separate observation Registration archives the source into the content-addressed Amber recovery tier and refuses to complete if that archive fails. This retention is deliberately described with a precision boundary: `source::capture` performs the single-open read-and-hash capture; `vault::archive_only` performs a **separate filesystem observation** during registration. The archive is not claimed to share the capture's file descriptor. Hiding that distinction would make the architecture sound stronger by making its evidence less honest. ### Drift is not a new version of the old citation Every ordinary resolve rechecks the registered source. If the live bytes no longer match the registered digest, the committed view is not silently served as though it were current. The refusal is `drifted`; capture and verification paths use `doc_source_drifted` where that stable operator code applies. A reader must explicitly register the new revision. A separately authorized historical-inspection path may serve the old view, but its delivery record marks `historical: true`. History is labeled history. This produces three useful non-equivalences: ```text same path ≠ same source revision same source revision ≠ same parser or extraction configuration old citation still resolving ≠ old citation is currently authoritative ``` ## 2. Content Decides the Format—Inside the Jail Format identification is itself parsing. A ZIP central directory must be inspected before the system can know whether a package is a DOCX, an XLSX, or an ordinary archive. That means identification cannot safely occur in the controller while only the resulting parser is confined. AINA performs identification and extraction in the restricted worker. Byte signatures and container structure take priority; extensions break defined ties. The current `Kind` enum has **22 variants**, but that is not a claim of universal file support: `Unknown` is deliberately unsupported, `Image(format)` is a parameterized family, and `EncryptedOffice` has a typed refusal rather than a text reader. | `Kind` variant | Content or structural identity | Extraction or typed outcome | |---|---|---| | `Pdf` | `%PDF-` signature | Structural PDF extraction, layout, annotations, and admitted local OCR paths | | `Json` | `.json` or `.jsonld` tie-break after higher-priority signatures | UTF-8 JSON document with bounded RFC 6901 source locators | | `Doc` | OLE2 plus `WordDocument` and a valid binary Word probe | Binary Word FIB and piece-table reconstruction | | `Docx` | ZIP package with Word document structure or an allowed OOXML extension tie-break | Body, tables, headers, footers, notes, comments, and revision accounting | | `Ppt` | OLE2 plus `PowerPoint Document` and a valid binary PowerPoint probe | Binary presentation record-tree text | | `Pptx` | ZIP package with PowerPoint presentation structure or an allowed extension tie-break | Slides, notes, comments, and relationships | | `Rtf` | `{\rtf` signature or `.rtf` | Bounded RTF control-word and encoding state machine | | `Xlsx` | ZIP package with `xl/workbook.xml` or an allowed extension tie-break | Workbook, cell, cached-value, formula, and comment extraction | | `Xlsb` | ZIP package with `xl/workbook.bin` | Binary workbook extraction | | `Xls` | OLE2 `Workbook`/`Book` probe, or `.xls`/`.xla`/`.xlt` fallback | BIFF workbook extraction | | `Ods` | ODF `mimetype` for a spreadsheet | OpenDocument spreadsheet extraction | | `Odt` | ODF `mimetype` for text | OpenDocument text extraction | | `Odp` | ODF `mimetype` for a presentation | OpenDocument presentation extraction | | `Eml` | `.eml` or bounded RFC 5322-style header recognition | MIME tree, headers, body, and recursive child sources | | `Msg` | OLE2 property streams plus root-property probe | MAPI properties, body, and attachments | | `Zip` | ZIP signature without a more specific package identity | Bounded archive inventory and child extraction | | `Html` | Decodable HTML, an HTML extension tie-break, or declared legacy HTML bytes | DOM-derived reading text plus preserved markup story | | `Text` | Successful strict text decoding without HTML identity | Normalized text stories and line elements | | `EncryptedOffice` | OLE2 `EncryptedPackage` | `encrypted` / `doc_source_extraction_failed`, never fake empty text | | `Ole2` | Other OLE2 compound storage | Embedded-object extraction or a named refusal | | `Image(format)` | Valid image signature and header | Header inventory and admitted OCR behavior under the selected profile | | `Unknown` | No supported identity | `unsupported_format`, never an empty successful document | This is allowlisted recognition, not extension trust and not a promise to parse every file on earth. ### Fourteen budgets govern expansion The outer operating-system jail is the backstop. Inside it, a 14-field `Limits` object admits expensive work before expansion and charges actual output. Callers may ratchet limits downward; zero and values above the ceiling are rejected. The exact limits are serialized with RFC 8785 JCS and hashed into `limits_sha256`, so changing a budget changes extraction identity. | Limit | Default ceiling | Failure it bounds | |---|---:|---| | `max_child_bytes` | 32 MiB | Oversized attachment, archive member, or embedded child | | `max_item_decompressed` | 64 MiB | Single-stream or single-member expansion | | `max_total_decompressed` | 256 MiB | Cumulative decompression across the document and children | | `max_parts` | 20,000 | Container-entry, mail-part, and page floods | | `max_nesting` | 3 | Recursive child-source descent | | `max_xml_depth` | 256 | Deep or amplifying XML structure | | `max_pdf_objects` | 500,000 | PDF object-graph amplification | | `max_form_depth` | 8 | Recursive PDF Form XObjects | | `max_cells` | 2,000,000 | Spreadsheet cell explosion | | `max_elements` | 250,000 | Addressable-element flood; later text may still be extracted without receiving nodes | | `max_output_chars` | 32 MiB | Extracted character-volume exhaustion | | `max_wall_ms` | 60,000 ms | Parser CPU/time overrun, with a 15-second outer kill allowance | | `max_image_pixels` | 16,000,000 | Single-image pixel expansion, admitted before decode | | `max_total_pixels` | 64,000,000 | Aggregate pixel expansion across pages and children | Budget refusal is evidence. It becomes a typed `budget_exceeded` condition rather than a partial answer with the exhaustion removed from the story. ## 3. The Parser Jail: Prove Confinement Before Resume A sandbox named in a configuration file is not a sandbox observed in a child token. AINA's document worker is built around pre-resume verification. ### Private staging Each parse receives an ephemeral scratch tree with three distinct grants: - `bin/`: a private read-execute copy of `squad-doc-extract.exe`, so the jail need not read the installation directory. - `input/`: read-only `source.bin` and a bounded `doc_source.worker/1` request containing the expected input digest and extraction limits. - `work/`: the only writable grant, including a jail-local temporary directory and a `result.json` opened with `create_new(true)`. The child recomputes the input digest before parsing. The supervisor bounds the result at 256 MiB, checks the response schema and input digest, and validates the returned extraction structure. Child output remains untrusted data even though the child was confined. ### Six-step launch contract 1. **Exclusive AppContainer identity.** A fresh per-parse profile is created. An already-existing profile is a `confinement_identity_reused` failure, not an invitation to share a jail. 2. **LPAC policy.** `PROCESS_CREATION_ALL_APPLICATION_PACKAGES_OPT_OUT` removes the implicit `ALL APPLICATION PACKAGES` access that would otherwise widen filesystem and registry reads. 3. **Minimal capabilities.** The document-media profile grants exactly `lpacCom`. Network, registry-read, and credential capabilities are not granted. 4. **Suspended creation and token inspection.** The process starts suspended. Before `ResumeThread`, the supervisor verifies that the token is AppContainer, carries the fresh package SID, has the non-zero `WIN://NOALLAPPPKG` attribute, and contains exactly the expected capability set. A mismatch kills the child before its initial thread executes. 5. **Job Object backstop.** While suspended, the child enters a kill-on-close Job Object with `ActiveProcessLimit = 1`, a 2 GiB job-memory ceiling, and a 25% hard CPU-rate cap. EcoQoS reduces scheduling priority; internal runtime thread pools are capped at two. The job is registered with break-glass observation so sovereign revocation covers in-flight parsers. 6. **Audited teardown.** The supervisor records `scratch_removed` and `profile_deleted` in the worker receipt. Cleanup failure is disclosed rather than rewritten into success. ```text verified Captured bytes → read-only input grant + expected digest → CREATE_SUSPENDED → AppContainer SID + LPAC attribute + exact capabilities verified → Job Object process/memory/CPU limits assigned → ResumeThread exactly once → content identification + budgeted extraction + warning capture → digest-bound result.json → supervisor schema and structural validation → worker Receipt + scratch/profile teardown disclosure ``` Parser panics become `parser_fault`. A panic isolated to one embedded child fails that child while preserving the parent and sibling reads; a top-level panic invalidates the extraction. On non-Windows, the untrusted-document path refuses with `parser_worker_unavailable` rather than pretending that an ordinary subprocess is equivalent to LPAC plus a Job Object. This is the Windows reference embodiment's claim ceiling: bounded process, capability, network, memory, CPU, and filesystem exposure on the governed parser path—not proof that every parser is bug-free or that a privileged host administrator cannot attack the machine. ## 4. Extraction Is Not Complete If the Gaps Disappear A parser that returns a string has answered only one question: what text did this adapter emit? It has not established whether the document contained more text, whether a library recovered by discarding structure, or whether the caller later received everything extracted. AINA keeps five status dimensions separate. | Dimension | Values | Question answered | |---|---|---| | `Integrity` | `verified`, `failed` | Did the captured bytes match the registered identity and integrity checks? | | `Inventory` | `complete_for_profile`, `partial`, `unknown` | Could the adapter enumerate the document's parts and establish a denominator? | | `Extraction` | `complete_for_profile`, `partial`, `unsupported`, `failed` | Was every in-profile, text-bearing part recovered without lossy omission? | | `Admission` | `allowed`, `blocked` | Is this source currently readable by this actor under project and Room policy? | | `Delivery` | `complete`, `truncated`, `chunked` | Did this response serve everything requested, or only a bounded window? | The immutable document report carries integrity, inventory, and extraction. Admission and delivery are facts about a particular serving. A fully extracted document shown through a 60-character window remains fully extracted **and** incompletely delivered. Those facts must not collapse into one green status. ### The 35-reason omission ledger The current `Reason` alphabet has 35 variants: **12 non-lossy profile exclusions and 23 lossy reasons**. `nesting_limit` belongs to the non-lossy set in the current implementation; it is not counted twice. | Classification | Exact variants | Consequence | |---|---|---| | **Profile-excluded, non-lossy (12)** | `binary_part`, `structure_only`, `embedded_font_program`, `image_content`, `media_content`, `deleted_revision`, `external_relationship`, `macro_content`, `derived_alternative`, `nesting_limit`, `structure_unavailable`, `markup_not_rendered` | Recorded for transparency without, by itself, forcing partial extraction or blocking absence evaluation | | **Lossy (23)** | `ocr_not_enabled`, `ocr_unavailable`, `unsupported_format`, `unsupported_part`, `unsupported_compression`, `unmapped_glyphs`, `font_encoding_unavailable`, `undecodable_text`, `stream_decode_failed`, `content_stream_invalid`, `object_missing`, `malformed`, `sheet_read_failed`, `formula_unavailable`, `encrypted`, `encrypted_entry`, `duplicate_entry_name`, `duplicate_member_name`, `multi_disk_archive`, `budget_exceeded`, `library_warning`, `parser_fault` | Forces partial/failed/unsupported handling and creates unknown coverage that an absence claim must acknowledge when in scope | Non-lossy does not mean irrelevant. It means the omission is expected under the declared extraction profile. A deleted revision excluded by `text/v1` may still matter enormously to a human reviewer; the profile must therefore be visible beside the claim. ### Parser warnings become evidence The confined child installs a `log::Log` interceptor at warning level. It preserves up to 1,000 observations, truncates each message at 2,000 characters, and counts overflow. Every warning becomes an observation. Except for one explicitly content-neutral PDF trailer normalization, every warning also becomes a lossy `library_warning` omission at document scope. A library's lenient recovery can therefore never silently upgrade itself to `complete_for_profile`. The omission collector has its own memory defense: at most 2,000 distinct omission tuples are retained, and additional occurrences are counted into a terminal budget-overflow record. Identical adapter omissions aggregate by location, reason, scope, and story. The coverage ledger cannot be allowed to become the denial-of-service vector it exists to report. ### Gaps are addressable nodes When the view is committed, omissions become `omission/1`, `omission/2`, and so on. A whole-document failure is `failure`; inventory status is `inventory`; each nested source is `child:`. These nodes carry no extracted text, but they can be outlined, cited, reviewed, and included in absence analysis. That is the difference between “the model said nothing was there” and “the system can identify exactly where its knowledge stops.” ### Exhibit 1 — A synthetic `ViewManifest` with an addressable omission The following is structurally faithful to `squad.evidence.view/1`, with synthetic identities and hashes. The view digest is computed over the manifest's JCS bytes; it is not a mutable field inside the manifest. ```json { "schema": "squad.evidence.view/1", "report_schema": "doc_source.report/2", "source": { "revision": "doc_20261030_142011_a1b2c3d4", "sha256": "58b2e91f4c03d8a7126590ef4312ab98c7654321fedcba09876543210fedcba9", "bytes": 12345, "name": "renewal.pdf", "canonical_path": "project:/evidence/renewal.pdf", "registered_at": "2026-10-31T14:20:11Z" }, "origin": "room_captured", "parser": { "name": "squad-doc-source", "profile": "text/v1", "digest": "4b9c210e8f1a77d4c0592a1e38f47c1190ab2d3e4f5061728394a5b6c7d8e9f0", "features": "synthetic-parser-feature-profile" }, "config": { "limits": { "max_child_bytes": 33554432, "max_item_decompressed": 67108864, "max_total_decompressed": 268435456, "max_parts": 20000, "max_nesting": 3, "max_xml_depth": 256, "max_pdf_objects": 500000, "max_form_depth": 8, "max_cells": 2000000, "max_elements": 250000, "max_output_chars": 33554432, "max_wall_ms": 60000, "max_image_pixels": 16000000, "max_total_pixels": 64000000 }, "limits_sha256": "84e01b92c7d34f56a1230987654321fedcba9876543210abcdef0123456789ab" }, "format": "pdf", "status": { "integrity": "verified", "inventory": "complete_for_profile", "extraction": "partial" }, "inventory_total": 2, "inventory": [], "stories": [ { "id": "page:1", "kind": "page", "locator": { "kind": "page", "page": 1 }, "primary": true, "method": "pdf_text_operators", "text": "Section 4.2 Renewal Cap: Annual uplift shall not exceed 3.0%." } ], "elements": [ { "id": "page:1!BT:0", "story": "page:1", "kind": "text_object", "span": { "start": 0, "end": 61 }, "locator": { "kind": "page_element", "page": 1, "element": "BT", "ordinal": 0 } } ], "omissions": [ { "id": "omission/1", "omission": { "location": "page:2/image:1", "stage": "extract", "reason": "ocr_not_enabled", "scope": "page", "detail": "page 2 contains a raster scan and local OCR is disabled", "count": 1 } } ], "children": [], "observations": [] } ``` The extracted cap is citable. So is the fact that page 2 was not OCRed. Neither fact may impersonate the other. ## 5. Immutable Views: Identity Includes the Reader Two systems can read identical bytes and produce different evidence because their parsers, profiles, feature sets, or limits differ. AINA therefore binds a view to all of them: ```text source revision + source SHA-256 + parser name + parser source/build digest + text/v1 profile + feature profile + exact 14-field limits + limits_sha256 + report schema + view schema + stories + elements + omissions + children + observations = canonical ViewManifest ``` RFC 8785 JSON Canonicalization Scheme (JCS) gives the manifest one deterministic serialization. Object keys are sorted as UTF-16 code-unit sequences, insignificant whitespace disappears, non-finite numbers are refused, integers beyond exact IEEE-754 range are not rounded, and nesting is capped at 128. String content is escaped but **not Unicode-normalized or whitespace-rewritten**. The SHA-256 of those canonical bytes is the 64-hex view digest. ### Commit children first, then seal the parent Attachments, archive members, and embedded workbooks are sources in their own right. Their revisions compose the parent revision, an encoded member locator, and the child SHA-256. Their views commit first; the parent's `child:` nodes then name the resulting child digests or failures. Child text is never promoted into the parent as though it had been read in the parent's own story. The commit runs inside the Room's cross-process `with_transition` serialization. A `begin` provenance node records the attempted view, the canonical artifact is stored by digest, and a `complete` node plus index closes the commit. Repeating the same manifest in the same Room resolves to the existing view rather than rewriting it. Loading is equally strict. The resolver checks the artifact's recorded digest, hashes the served bytes, re-canonicalizes the parsed value, and compares canonical form. One altered byte or non-canonical repair yields `view_digest_mismatch`. A new parser or configuration creates a new view. It does not retarget old citations. Historical identity survives; current authority must be re-established. ## 6. Server-Cut Evidence: `srv_` Before `ev_` A server-cut citation is not a model emitting a prettier reference format. It is a two-event protocol. ### Event 1: the Room serves content and issues `srv_` The `squad.evidence.receipt/1` record binds: - the Room, source revision, and view digest; - the recipient and recipient session; - the serving tool; - each node and optional character interval served; - delivery counts and `complete`, `truncated`, or `chunked` status; - admission, optional via-handle or wake identity, time, and nonce. The receipt identifier is `srv_` plus the first 32 hexadecimal digits of the receipt's canonical SHA-256. `read`, `resolve`, and `wake_delivery` are content-serving tools. `outline` is different: it serves structure and coverage metadata. An outline receipt can ground a textless `omission/N`, `failure`, `inventory`, or `child:` node. It cannot ground a quotation from a text block the caller was never served. ### Event 2: the Room computes the citation and issues `ev_` The caller submits a receipt plus an exact quote, or a node it was served. For a quote, the caller may supply a node hint or a one-based occurrence when text repeats. The caller does **not** supply authoritative character offsets. The server then: 1. verifies that the receipt belongs to the caller; 2. loads and reverifies the canonical view; 3. rechecks source admission and drift; 4. searches only the content covered by that receipt; 5. refuses zero matches with `quote_not_found` and unresolved repetition with `quote_ambiguous`; 6. computes Unicode-scalar coordinates; 7. selects the innermost covering element with `View::narrowest`, falling back to the story; 8. checks that the selected range was actually served; 9. issues a content-addressed `ev_` handle whose issuance record includes the canonical `EvidenceRef`, basis, Room, issuee, and first-hand lineage. ```text agent: receipt + exact quote (+ optional occurrence) server: served? → current? → unique? → narrowest node? → selector in bounds? agent: receives ev_ handle; receives no authority to invent its coordinates ``` ### Exhibit 2 — A valid synthetic citation issuance ```json { "handle": "ev_6c8f1029384756ab12cd34ef56789012", "issuance": { "schema": "squad.evidence.handle/1", "envelope": { "version": 1, "source_revision": "doc_20261030_142011_a1b2c3d4", "view_digest": "a1b2c3d4e5f60718293a4b5c6d7e8f901a2b3c4d5e6f708192a3b4c5d6e7f809", "node_id": "page:1!BT:0", "selector": { "kind": "bounded_subrange", "unit": "unicode_scalar", "start": 25, "end": 61 }, "source_native_locator": { "kind": "page_element", "page": 1, "element": "BT", "ordinal": 0 }, "role": "supporting", "visibility": "all", "origin": "room_captured" }, "envelope_digest": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "room": "architecture-demo", "issued_to": "analyst-alpha", "issued_at": "2026-10-31T14:20:15Z", "basis": { "kind": "receipt", "receipt": "srv_91d2e3f4a5b6c7d80123456789abcdef" }, "lineage": { "first_hand": true, "origin": "analyst-alpha" }, "nonce": "synthetic-nonce-001" } } ``` The handle proves which exact passage was issued to which actor on which view and receipt basis. It does not prove that the contract is authentic, that 3.0% is commercially reasonable, or that page 2 agrees. ### Refusal is part of the citation grammar Paths and URLs are refused with `not_a_path`. Raw hashes, hand-built offsets, and similar manufactured references are refused with `agent_manufactured_ref`. Unknown well-formed handles yield `ev_unknown`. Foreign-receipt use is `access_denied`; content outside the receipt is `not_served` or `quote_not_found`. Reserved locator kinds and offset units fail by name rather than being loosely accepted. ### Exhibit 3 — `quote_ambiguous` The actual refusal is an error string; it is shown here as a JSON string with synthetic occurrence data abbreviated for publication. ```json "quote_ambiguous: the quotation occurs 3 times in the served content; name one with occurrence: [{\"occurrence\":1,\"node\":\"page:1!BT:0\"},{\"occurrence\":2,\"node\":\"page:3!BT:1\"},{\"occurrence\":3,\"node\":\"page:7!BT:2\"}]" ``` The server does not choose the most convenient occurrence. The caller must disambiguate. ### Exhibit 4 — `drifted` ```json "drifted: project:/evidence/renewal.pdf no longer holds the registered bytes (registered 58b2e91f4c03d8a7126590ef4312ab98c7654321fedcba09876543210fedcba9, current c0ffee99887766554433221100ffeeddccbbaa99887766554433221100ffeedd); the committed view is not served in its place — re-register, or request authorized historical inspection" ``` The old view remains intact. Its current evidentiary role does not. ## 7. Network Capture: Authorize Every Hop, Not Just the First URL Documents also arrive through retrieval. A fetch tool that authorizes only the initial URL can be redirected into a loopback service, private network, or cloud metadata endpoint. The safe rule is not “validate the first destination”; it is **authorize the destination before every request**. AINA's capture client disables automatic redirects with `redirect::Policy::none()`. It follows redirects manually, up to ten, and calls `authorize_url` on each current URL before sending. Every hop records its URL, status, optional content type and `Location`, and declared security headers. The final body is capped at 8 MiB. A body cut by that ceiling becomes a `snippet` with `truncated: true`, not a complete response. The exact bytes are stored as a Room artifact, hashed, and parsed through the same confined worker. Every fetch is a new `cap_` capture; re-fetching a URL never overwrites the earlier observation. ### Captured, supplied, and governing are different roles | Origin and root | What the Room observed | May anchor a governing requirement? | |---|---|---| | `room_captured` + `doc_…` | Registered bytes captured from admitted project territory | Yes, after first-hand serving and the other requirement checks | | `room_captured` + `cap_…` | A response body fetched through the Room's egress-gated path | No; research capture is not a registered governing document | | `agent_supplied` + `cap_…` | Text an agent pasted with a claimed, unverified origin | No; citing it cites the supplier, not the claimed source | An agent-supplied excerpt is not inherently false. It is explicitly labeled evidence of what the agent supplied. It cannot be laundered into a Room observation merely because its bytes match a real document. The same boundary applies to web research: egress authorization and immutable capture establish provenance, not truth and not permission for retrieved instructions to govern the system. ## 8. Findings: Claims the Server Can Re-Evaluate A citation establishes a passage. A finding states something checkable about it. The evaluator supports five typed assertions: - **`equals`** — exact text or typed-value equality. - **`in_range`** — a typed numeric value against minimum/maximum bounds, with an optional unit check when the view knows the unit. - **`matches_regex_bounded`** — a pattern capped at 512 characters, compiled with 1 MiB size limits and nesting depth 32. - **`exists`** — the selected text or structural node exists. - **`formula_equals`** — exact comparison against a spreadsheet formula node, including the formula associated with a cached-value cell. The interpreter remains separate from the machine-checked assertion. A model may offer a bounded interpretation, but the server re-evaluates the predicate against the committed view. A false predicate is not deleted to make the ledger look clean. The `fnd_` finding and failed `evl_` evaluation are persisted, and only then does the call return `predicate_false`. Failed claims remain failed matter. ### Absence is a coverage claim before it is a search claim “No change-of-control clause exists in this contract” is much stronger than “the phrase did not occur in the chunks I saw.” It requires an inventory of what was examined and an explicit account of what remains unknown. For an absence claim, the server derives required unknown-coverage markers: - `failure` for failed or unsupported extraction; - `inventory` when enumeration is partial or unknown; - every in-scope lossy `omission/N`; - every `child:` for whole-document scope, because child text is not merged into the parent's searched stories. The caller supplies `acknowledged_gaps` as the relevant **node IDs**. These are not authoritative character offsets and are not a second set of fabricated evidence references. The server checks that every required marker is acknowledged and that every acknowledged marker exists in the view. Outline receipts can be used to cite and inspect those textless markers. Only then does the evaluator search for the phrase or bounded pattern. A successful negative result means, precisely: > The phrase is absent from the content this view examined; acknowledged gaps remain unknown, not absent. ### Exhibit 5 — `absence_unsupported` ```json "absence_unsupported: the claimed scope has unknown coverage; an absence claim must acknowledge each marker at its narrowest known location: omission/1" ``` The server refuses to let a missing OCR pass become proof that page 2 contains nothing. ## 9. Revision, Reassessment, and Correspondence Without Retargeting Evidence is allowed to change. History is not allowed to move underneath its citations. When a new view is committed for a source, dependent findings, evaluations, and requirements receive deterministic `rsa_` reassessment flags. Requirement amendments form `reqrev_` revision chains; task evidence manifests (`man_`) pin the admitted source and optional view context. A task or handoff cannot silently complete over evidence now marked for reassessment. Cross-version correspondence is a separate, explicit operation: 1. `correspond` proposes candidate locations in a later view. 2. The candidate search has four stages, with a separate same-kind rule for textless structural nodes: `same_node_same_text`, `same_node_same_kind`, `exact_text`, `normalized_text`, then the weak `same_node_changed_text` fallback. 3. The reviewer must actually be served the candidate in the later view. 4. `adopt-correspondence` records `coa_`, creates a **new** `ev_` handle, and adds a `corresponds_to` relation. 5. The old handle remains bound to its old view. A requirement-role handle adopted through correspondence becomes supporting evidence in the new view rather than silently carrying governing authority forward. ```text old view + old ev_ handle ───────────────→ remains historically pinned │ │ ├─ new view committed ─→ rsa_ reassessment flags on dependents │ │ └─ cor_ proposal → candidate served in new view ↓ coa_ adoption + new ev_ handle ↓ corresponds_to relation; no retargeting ``` This is the difference between provenance and path mutation. The system can say how two observations correspond without pretending they are the same observation. ## 10. Echo Collapse: Count First-Hand Origins, Not Repetitions Multi-agent systems are particularly good at making repetition look like corroboration. Agent A reads a passage. Agent B cites A's finding. Agent C cites B's summary. The graph now contains three records and still contains only one first-hand observer. `lineage::origins` walks citations for the target view and node. A first-hand citer contributes its actor identity as an origin. An echo contributes the origin it rests on. The output records `first_hand_citers`, `echo_citations`, and the collapsed `distinct_origins` set. ### Exhibit 6 — Three citation records, one first-hand origin ```json { "room": "architecture-demo", "view_digest": "a1b2c3d4e5f60718293a4b5c6d7e8f901a2b3c4d5e6f708192a3b4c5d6e7f809", "node": "page:1!BT:0", "citations": 3, "records_by_kind": { "finding": 3 }, "first_hand_citers": ["analyst-alpha"], "echo_citations": [ { "citer": "reviewer-beta", "rests_on": "analyst-alpha", "record": "fnd_11111111111111111111111111111111" }, { "citer": "reviewer-gamma", "rests_on": "analyst-alpha", "record": "fnd_22222222222222222222222222222222" } ], "distinct_origins": ["analyst-alpha"], "distinct_origin_count": 1 } ``` If beta later receives the passage directly and cites it first-hand, the count becomes two. That is a precise operational metric for an owner-declared evidence-origin requirement. It is not proof of statistical, organizational, or causal independence. Two first-hand readers can share a flawed parser, training data, administrative control, or source document. **Distinct first-hand observation is not the same thing as independent validation.** Article 01's three independence claims still apply. ## 11. The 2026 Evidence Landscape: What External Work Establishes The 2026 literature and vulnerability record reinforce the need for explicit evidence boundaries. They do not validate AINA, and their measurements should not be stretched beyond their actual scope. | 2026 primary source | What it actually reports | Architectural consequence—and ceiling | |---|---|---| | Zhao, Pandey, and Misra, [*What the Author Wrote, What the Agent Reads*](https://arxiv.org/abs/2610.30026), October 2026 preprint | Across 76 real policy documents, formatting-blind ingestion reduced requirement-recovery micro-F1 from 0.690 to 0.150 for Docling and 0.075 for plain text; six answering models replicated the loss. Highlight and color drove the largest measured effects. | Ingestion can make agents confidently wrong by removing authorial cues. This supports explicit structural and omission accounting; it does not by itself prove that every omitted cue changes every decision. | | Wang et al., [*From Agent Traces to Trust*](https://arxiv.org/abs/2606.04990), June 2026 survey, revised October 2026 | Organizes evidence tracing and execution provenance into typed graphs and reports limited integration, incomplete-chain benchmarks, and little tamper-evident or attested trace reporting in the surveyed catalog. | Typed provenance must connect evidence and execution rather than treating a final answer as its own audit. A survey supplies a framework, not an AINA benchmark. | | Khajavi et al., [*CiteCheck*](https://arxiv.org/abs/2605.27700), May 2026 preprint, revised October 2026 | Evaluates corrupted and fabricated bibliographic references on a 982-citation physics benchmark, reporting 88.7 macro-F1 for its hybrid detector. | Plausible citation metadata requires verification. Bibliographic identity is a different problem from quotation-span grounding; CiteCheck is not evidence that server-cut spans work. | | Meng et al., [*Social Circuits behind Multi-agent Echo Chambers*](https://arxiv.org/abs/2609.34444), September 2026 preprint | Studies how messages reinforce shared errors and traces their effects through receiver activations; proposes circuit-guided message selection. | More agents can amplify rather than correct an early error. The work does not prove that every topology converges or that actor-origin counting solves model independence. | | Setiawan, [*Prompt-Based Abstention Fails Under Misleading Context*](https://arxiv.org/abs/2608.22228), August 2026 preprint | On three small frozen models (3.8B–8B), explicit abstention prompting still answered 41.6% of misleading questions; 63% of those answers echoed the planted wrong entity. | Prompted caution is not a deterministic coverage gate. The result is scoped to the tested models and conditions, not a universal frontier-model failure rate. | | Pai, [*PARSE*](https://arxiv.org/abs/2606.17467), June 2026 preprint, revised September 2026 | On 122 professional-domain tasks using real retrieved documents, PARSE reported 15.6% attack success versus a 25.4% baseline, with utility and significance limitations disclosed. | Retrieval sanitization and provenance-aware processing belong before judgment. PARSE is a sanitization result, not proof that cryptographic provenance tags neutralize prompt injection. | | [`CVE-2026-27026`](https://nvd.nist.gov/vuln/detail/CVE-2026-27026), February 2026 | Malformed `/FlateDecode` streams could cause long runtimes in `pypdf` before 6.7.1 through byte-by-byte decompression. | Per-item and aggregate budgets, wall clocks, and an outer Job Object are necessary backstops. The advisory does not need to be exaggerated into a different memory-corruption claim. | | [`CVE-2026-104120`](https://nvd.nist.gov/vuln/detail/CVE-2026-104120), October 2026 | `mcp-server-fetch` and `mcp-server-everything` through 2026.6.4 were affected by SSRF in `fetch_url` URL/path handling. | URL fetches require an egress boundary. AINA additionally authorizes every redirect hop on its capture path; the CVE description alone is not used to prove a particular redirect implementation. | | [`CVE-2026-51994`](https://nvd.nist.gov/vuln/detail/CVE-2026-51994), September 2026 | `mcp-remote` 0.1.32–0.1.38 was vulnerable to SSRF through the `resource_metadata` URL extracted from `WWW-Authenticate`. | Remote metadata is untrusted destination input. The capture-path proof does not automatically cover unrelated OAuth clients; every network route needs its own inventory and gate. | | IETF [`RFC 9943`](https://www.rfc-editor.org/rfc/rfc9943.html), June 2026, SCITT Architecture | Defines registration of signed statements and transparency receipts in an append-only Transparency Service. | Tamper-evident inclusion complements canonical evidence views. Inclusion proves log behavior, not that the registered statement is substantively true. | The composition matters. Sandboxing addresses parser blast radius. Canonical views address artifact identity. Server-cut citations address served-span provenance. Bounded absence addresses unknown coverage. Origin collapse addresses repetition. None can substitute for the others. ## 12. The Twenty-Case Evidence Conformance Gate The epistemic subsystem has its own 20-case evidence battery inside the 45-case candidate-binary publication gate. These are author-controlled engineering checks, not external certification. | Case | Predicate exercised | |---|---| | `evidence_jcs_rfc8785_vectors` | Canonical serializations match the RFC's vectors | | `evidence_corpus_through_mcp` | Corpus documents read through the confined worker with expected nodes and outcomes | | `evidence_governing_pdf_six_pages` | The governing PDF yields its six text pages through `squad_doc_source read` | | `evidence_repeated_text_resolves_each_node` | Repeated text is ambiguous until occurrence selection, then each handle resolves to its own node | | `evidence_drift_refused` | Altered source bytes refuse reading and resolution | | `evidence_tampered_view_refused` | Altered committed views fail with `view_digest_mismatch` | | `evidence_old_view_survives_extraction_change` | New extraction identity creates a new view while the old handle remains on the old view | | `evidence_bad_selector_refused` | Bad selectors, paths, raw hashes, unknown handles, and reserved extensions fail by name | | `evidence_unauthorized_dereference_refused` | A caller outside the Room's owner roster cannot dereference | | `evidence_truncation_in_delivery_status` | Bounded delivery reports chunked/truncated counts | | `evidence_excerpt_stays_distinct` | Agent-supplied material retains its origin and cannot take requirement role | | `evidence_echo_lineage_preserved` | Echo citations remain echoes resting on a first-hand origin | | `evidence_distinct_origins_through_three_hops` | Three echo-linked records collapse to one origin; a second first-hand reader makes two | | `evidence_false_predicate_fails_receipt` | False predicates fail evaluation and remain as failed matter | | `evidence_absence_requires_complete_inventory` | Absence is refused until every unknown-coverage marker is acknowledged | | `evidence_correspondence_keeps_old_handle` | Correspondence adoption issues a new handle without retargeting the old one | | `evidence_reassessment_flags` | Source and requirement revisions flag dependents until answered | | `evidence_staged_visibility_reveal` | Collaborative-pass findings remain withheld until reveal, with delivery limits disclosed | | `evidence_origin_gate` | Completion refuses until a requirement has the declared distinct first-hand origins | | `evidence_capture_provenance_chain` | Every redirect hop is egress-authorized and records status, location, and security headers | The gate demonstrates named path behavior on the candidate binary. It does not prove universal format coverage, universal parser memory safety, complete network-route coverage, source truth, or independent assurance. Nor does it establish a throughput or latency profile for an enterprise corpus; resource cost must be measured against the actual document mix, limits, worker count, and delivery windows. **Evidence availability:** the implementation and conformance artifacts are `CONTROLLED` for qualified, method-scoped technical review, with adverse findings permitted. They are not represented as unrestricted public downloads or independent validation. The review route is `[RESEARCH]` through [Article 09](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/). ## Bring One Document Boundary The **Zero-Trust Evidence & Citation Audit Challenge** is the direct application of this layer. Bring one document-heavy AI workflow and one bounded evidence question. I will map its source-to-view-to-citation predicates, inspect the omission surface, and demonstrate fabricated-citation refusal against the existing implementation. The challenge excludes universal format coverage, source-truth guarantees, absence without acknowledged gaps, production write authority, and independent validation. Deployment and paid pilots are separately scoped; independent evaluation remains outside the commercial ladder through [Article 09](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/). ## The Article 02 Claim Card ```text Mechanism: AINA Epistemic Trust Layer (doc_source, confined extraction, evidence views, receipts, handles, findings, requirements, reassessment, correspondence, and lineage) Failure it governs: Hostile document parsing inside the controller; stat/read/hash substitution; silent extraction loss; fabricated or unserved citations; stale evidence; unbounded absence claims; and multi-agent repetition masquerading as corroboration. Positive path: Registered doc_ source → single-open FILE_SHARE_READ capture and SHA-256 → LPAC + Job Object confined identification and budgeted extraction → RFC 8785 ViewManifest and 64-hex view digest → bounded served-content srv_ receipt → server-computed EvidenceRef and ev_ handle → typed finding / bounded absence / requirement revision → EISA sealed case (Article 03). Refusal / unknown-coverage path: doc_source_drifted or drifted parser_worker_unavailable, budget_exceeded, parser_fault not_a_path, agent_manufactured_ref, ev_unknown access_denied, not_served, quote_not_found, quote_ambiguous view_digest_mismatch, role_ineligible, absence_unsupported. Evidence status: Implemented and Author-Verified on the declared Windows reference embodiment. Twenty evidence cases form part of the 45-case candidate-binary publication gate. Not Independently Validated. External 2026 papers and advisories do not endorse AINA. Verifier and artifacts: Captured source digest; worker Receipt; doc_source.report/2; squad.evidence.view/1 and view digest; squad.evidence.receipt/1 (srv_); squad.evidence.handle/1 (ev_); fnd_, evl_, reqrev_, man_, rsa_, cor_, coa_ records; evidence_conformance.rs:301-322. Scope and claim ceiling: Confinement covers the governed Windows parser path, not every possible host process. The Kind alphabet is bounded and includes typed unsupported/encrypted outcomes. Complete-for-profile is not universal extraction or human comprehension. A server-cut citation proves served-span provenance, not source truth. Absence is bounded to examined content with acknowledged gaps. Distinct first-hand actor origins are not statistical or organizational independence. Egress authorization is proven on the capture path, not every network client. Next stage in the machine: Article 02 seals the evidence world. Article 03 compiles candidate judgments into EISA's closed act grammar and adjudicates support, contradiction, partition, and layout relations. Commercial doorway: Zero-Trust Evidence & Citation Audit Challenge — bring one document-heavy AI workflow; receive a source-to-view-to-citation predicate map, an omission surface review, and a live demonstration of fabricated-citation refusal (Article 09 · me@blayneaina.com). ``` ## Before EISA Judges the Act Return to the renewal contract. The conventional pipeline returned a clean answer: “Annual increases are capped at 3.0%, and there is no change-of-control clause.” Its citation looked plausible. Its confidence was high. Its evidence world was never established. The governed pipeline can tell a different story. The source revision was captured and hashed from one read. The parser ran with one capability, no network, bounded resources, and a token verified before resume. Page 1 produced a text-object node containing the 3.0% sentence. Page 2 produced an `ocr_not_enabled` omission, not silence. The analyst received a bounded content window and an `srv_` receipt. The Room found the exact sentence, computed its Unicode-scalar range, selected the innermost text object, and issued `ev_…`. The 3.0% assertion can now be checked. The change-of-control absence claim cannot pass until `omission/1` and every other unknown-coverage marker in scope are acknowledged. If another agent repeats the finding without reading the passage, lineage preserves the repetition as an echo. If the file changes, ordinary resolution refuses drift. If a new view is adopted, correspondence creates a new handle rather than moving the old one. Only now is there a stable object for constitutional judgment. > **Models may interpret evidence. They may not manufacture the evidence world, erase its gaps, or award themselves first-hand observation.** Article 03 begins at that boundary: candidate judgments become EISA acts, support and contradiction become deterministic relations, and entity-resolution panels decide what may acquire standing. ## Frequently Asked Questions ### What is a server-cut citation? A server-cut citation is an `ev_` handle issued by the Room from a caller-bound `srv_` served-content receipt. The caller supplies an exact quote or served node; the server verifies coverage, computes Unicode-scalar coordinates, selects the narrowest structural node, and binds the result to a canonical view. ### Does LPAC confinement make document parsing safe? It makes the governed parser path substantially safer by restricting capabilities, network access, writable locations, process count, memory, and CPU, with token checks before resume. It does not prove that parser code is bug-free or that every host and administrative route is contained. ### Does `complete_for_profile` mean the whole document was read? No. It means extraction completed without lossy omissions under the declared `text/v1` profile. Profile exclusions remain disclosed, delivery can still be truncated, and the result does not imply visual understanding or humanlike comprehension. ### Can a server-cut citation prove that a source is true? No. It proves that a particular passage was served and bound to an exact source revision, parser configuration, view, node, and recipient. Authenticity, worldly truth, policy quality, and the interpretation of that passage remain separate claims. ### How does bounded absence work? The server identifies unknown-coverage markers such as lossy omissions, incomplete inventory, failure, and child sources. The caller must acknowledge every required marker by node ID before the phrase or pattern search can support the bounded negative result. Acknowledged gaps remain unknown, not absent. ### Do three agents citing one passage provide three independent origins? Not if two are echoes. Lineage counts first-hand actor observations and collapses derived citations to their origins. Even two first-hand actors are not automatically statistically, organizationally, or causally independent. ## Continue the Series - **Series index:** [From Proposal to Consequence: The AINA Transition-Governance Series](https://blayneaina.com/) - **Previous:** [Article 01 — An AI Output Is Not an Act](https://blayneaina.com/01-an-ai-output-is-not-an-act/) - **Next:** [Article 03 — Semantic Switch and EISA: Protected Computation and Entity Resolution](https://blayneaina.com/03-eisa-four-gate-verdict-machine-and-entity-resolution/) - **Deployment and independent review:** [Article 09 — Bring One Real Boundary](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/) --- ## Primary 2026 Sources - [Zhao, Pandey, and Misra — *What the Author Wrote, What the Agent Reads: Silent Cue Loss in Enterprise Document Ingestion*, `arXiv:2610.30026`](https://arxiv.org/abs/2610.30026) - [Wang et al. — *From Agent Traces to Trust: A Survey of Evidence Tracing and Execution Provenance in LLM Agents*, `arXiv:2606.04990`](https://arxiv.org/abs/2606.04990) - [Meng et al. — *Social Circuits behind Multi-agent Echo Chambers*, `arXiv:2609.34444`](https://arxiv.org/abs/2609.34444) - [Setiawan — *Prompt-Based Abstention Fails Under Misleading Context*, `arXiv:2608.22228`](https://arxiv.org/abs/2608.22228) - [Khajavi et al. — *CiteCheck*, `arXiv:2605.27700`](https://arxiv.org/abs/2605.27700) - [Pai — *PARSE: Provenance-Aware Retrieval Sanitization for Professional Domain LLM Agents*, `arXiv:2606.17467`](https://arxiv.org/abs/2606.17467) - [NVD — `CVE-2026-27026`, `pypdf` long-runtime denial of service](https://nvd.nist.gov/vuln/detail/CVE-2026-27026) - [NVD — `CVE-2026-104120`, MCP fetch-server SSRF](https://nvd.nist.gov/vuln/detail/CVE-2026-104120) - [NVD — `CVE-2026-51994`, `mcp-remote` resource-metadata SSRF](https://nvd.nist.gov/vuln/detail/CVE-2026-51994) - [IETF `RFC 9943` — SCITT Architecture, June 2026](https://www.rfc-editor.org/rfc/rfc9943.html) --- ## Publication Metadata **Title tag:** Zero-Trust Document Ingestion & Server-Cut Evidence (`srv_` → `ev_`) **Meta description:** Why RAG chunks, file paths, and agent-written citations are not evidence: inside AINA's LPAC-confined parser, omission ledger, canonical views, and server-cut `srv_` and `ev_` handles. **Series navigation:** [Series index](https://blayneaina.com/) · Article 02 of *From Proposal to Consequence: The AINA Transition-Governance Series*. **Previous installment:** [Article 01 — An AI Output Is Not an Act](https://blayneaina.com/01-an-ai-output-is-not-an-act/). **Next installment:** [Article 03 — Semantic Switch and EISA: Protected Computation and Entity Resolution](https://blayneaina.com/03-eisa-four-gate-verdict-machine-and-entity-resolution/). ================================================================================ # ARTICLE §03: EISA Four-Gate Verdict Machine & Strict AST Parser Canonical URL: https://blayneaina.com/03-eisa-four-gate-verdict-machine-and-entity-resolution/ # EISA Four-Gate Verdict Machine & Strict AST Parser *The production EISA core determines which bounded acts earn constitutional standing; Semantic Switch is an experimental alpha module for structure-preserving entity-resolution research.* **By Blayne Aina · Principal Systems Architect and creator of AINA · October 31, 2026** --- Entity resolution has a data-boundary problem before it has a model problem. Contacts, companies, products, and other native records can contain identifiers that should never be sent raw to an external AI provider. But destructive redaction destroys the relationships that useful comparison requires: exact equality, normalized equality, phone variance, token containment, and the difference between a substitution, insertion, deletion, and transposition. **Semantic Switch resolves that transport problem first.** Native rows, keys, and the private crosswalk remain under client control. External computation receives structure-preserving fictional records instead of the declared raw values. Qdrant Edge and qBlock can then nominate candidate partitions without exposing the native rows. Fracture adjudicates identity, Forge selects canonical values with provenance, local restoration maps authorized references back to exact native values, and Shield defends the resolved baseline. Protected computation alone still does not make a decision correct. Three agents can read the same contaminated context, repeat the same unsupported merger, and produce three confident explanations. A judge model can make the fourth vote. Nothing about that process establishes which records were compared, which values survived inspection, which contradiction was omitted, or whether the proposed partition covers the candidate universe exactly once. A trustworthy system does not count votes. It compiles each proposal into a bounded act and recomputes the declared predicates over sealed evidence. Article 02 built the evidence world: exact source revisions, confined parser views, served-content receipts, and server-cut `ev_` handles. Article 03 connects it to **Semantic Switch** and the **Evidence Instruction Set Architecture (EISA)**: protected computation first, deterministic constitutional judgment second. > **Native data stays local. Semantic Switch powers protected qBlock candidates. EISA determines whether the resulting judgments are well-formed, evidence-contained, type-correct, and substantively consistent. The Room commits the result without executing it.** That last distinction matters. An operative EISA act is not a CRM merge, a file write, or a network effect. It is a recognized constitutional artifact entering the coordination and execution boundaries that Articles 04–08 govern. ## One Constitution, Two Proof Lanes EISA applies the same rule to two very different proposal classes: ```text sealed evidence world │ strict acts over one sealed case │ Four-Gate deterministic verdict ┌──────────┴──────────┐ │ │ Proof Lane A Proof Lane B PDF vector relations Protected entity resolution CONNECTS only Semantic Switch → qBlock → geometry + direction Fracture → Forge → Shield frames + branch labels identity + canonical state └──────────┬──────────┘ │ squad.eisa.submission/1 → eisa_<32hex> │ Room authority and fenced execution ``` The recognition spine is `ev_` evidence → `eisacase_` sealed case → bounded `EISA_ACT_V1` acts → Four-Gate recognition → recognized artifact → immutable `eisa_` Room submission. The two lanes demonstrate that spine in different evidence domains. Lane A proves a narrow spatial relation from parser-committed geometry rather than asking a vision model to infer an arrow. Lane B turns native customer records into a protected comparison substrate, adjudicates identity relations, compiles canonical state, and defends the sealed baseline against later drift. **Naming precision:** `Forge` names two related but distinct responsibilities. In Proof Lane A, `eisa.forge.document_layout/1` is the wire profile for spatial `AssertLayoutRelation` acts. In Proof Lane B, **EISA Forge** is Stage 3 of the entity-resolution pipeline, compiling canonical state after Fracture admits an identity partition. The layout profile and the canonical-state compiler are not the same mechanism or proof surface. The common rule is more important than either application: **a proposal may not choose its own evidence world, suppress inconvenient matter, or narrate itself into authority.** ## 1. Strict Representation Before Judgment A grammar cannot govern a transition if two parsers can assign different meanings to the same bytes. ### Gate 1 begins with `strict::parse` EISA does not deserialize submissions through an ordinary last-key-wins JSON map. `strict::parse` drives a recursive `StrictVisitor` before schema evaluation: - `visit_map` checks whether each incoming key already exists and rejects duplicates. - `visit_f64` refuses non-finite numbers. - `de.end()` rejects trailing non-whitespace matter after the root value. A submission is capped at **1 MiB** before parsing and must be a non-empty array of at most **256 acts**. One submission belongs to one act family: Fracture contact acts or Forge layout acts, never a mixture. This prevents a gateway from seeing one opcode while a downstream evaluator executes another. The August 2026 Ruby JSON vulnerability [`CVE-2026-71847`](https://nvd.nist.gov/vuln/detail/CVE-2026-71847) is a related warning about parser edge cases: incomplete objects containing duplicate keys could trigger a heap-use-after-free in the affected native parser. EISA's separate point is semantic—duplicate-key matter is ambiguous and therefore inadmissible, not repaired. ### The exact ten-key act envelope Every act contains all ten, and only ten, top-level fields in `ACT_KEYS`: | Field | Current wire law | |---|---| | `version` | Exactly `EISA_ACT_V1` | | `stage` | A declared stage whose family admits the opcode: `Fracture` or `Forge` | | `actor` | Must match the authenticated caller, case-insensitively | | `opcode` | One of the seven closed opcodes | | `operands` | An object with exactly the opcode's declared keys | | `evidenceRefs` | At most 16 strings naming Room-issued `ev_` handles; layout relations require exactly the connector/from/to handles | | `confidence` | An empty object, or `{ "bps": integer }` with `0 <= bps <= 10000`; audit only | | `boundedReason` | A string of at most 1,024 characters, with control characters other than newline refused; audit only | | `sourceEvent` | Empty, or `evt_<16 hex>_<8 hex>` | | `artifactHash` | The sealed case's bare 64-lowercase-hex JCS digest; no `sha256:` prefix | `confidence` and `boundedReason` are not consulted by any gate and do not enter `consensusIdentity`. A model can explain itself for operators. It cannot make its explanation causal by declaring it confident. Act IDs are local dependency names: a letter followed by up to 63 ASCII letters, digits, underscores, or hyphens. Duplicate IDs fault every act carrying the ambiguous name. ### Record-relative RFC 6901 pointers A Fracture `BindEvidenceTag` coordinate has exactly three fields: ```json { "artifact": "records", "recordId": "rec_101", "pointer": "/contact/workEmail" } ``` The participant names the record and pointer. The recognizer reads the value. Coordinates never name a filesystem path, URL, case-array index, or foreign artifact. `pointer::parse` enforces: - a leading `/`; - at most 256 bytes and eight tokens; - no whole-record empty pointer; - no control characters; - only RFC 6901 `~0` and `~1` escapes. `pointer::resolve` then walks the sealed record. Object members must exist. Array indices use canonical decimal syntax—`0` or a nonzero digit sequence without leading zeroes—and must be in bounds. Traversal into a scalar fails. Missing fields, malformed indices, and out-of-bounds access become typed faults rather than guessed values. Comparator inputs are **non-null, non-empty strings of at most 1,024 characters**. Numbers, booleans, arrays, objects, nulls, and whitespace-only strings do not become comparator text through coercion. ### The profile bounds | Constant | Value | Boundary | |---|---:|---| | `VERSION` | `EISA_ACT_V1` | Wire version | | `PROFILE` | `eisa.fracture.contact/1` | Contact act family | | `LAYOUT_PROFILE` | `eisa.forge.document_layout/1` | Layout act family | | `RECORDS_ARTIFACT` | `records` | Only coordinate substrate | | `MAX_SUBMISSION_BYTES` | 1,048,576 | Submission bytes | | `MAX_ACTS` | 256 | Acts per submission | | `MAX_REASON_CHARS` | 1,024 | `boundedReason` characters | | `MAX_ID_CHARS` | 64 | Act and record ID ceiling | | `MAX_POINTER_BYTES` | 256 | Pointer bytes | | `MAX_POINTER_DEPTH` | 8 | Pointer tokens | | `MAX_VALUE_CHARS` | 1,024 | Compared string characters | | `MAX_REFS` | 16 | Evidence handles per act | | `MAX_LIST` | 1,000 | Members, support, counter, and groups | | `MAX_CHAIN` | 4 | Duplicate links per transitive chain | | `MAX_CASE_BYTES` | 4,194,304 | Sealed case-file bytes | | `MAX_CASE_RECORDS` | 1,000 | Records per case | | `MIN_PHONE_DIGITS` | 7 | Minimum digits on each phone side | ## 2. The Closed Grammar: Seven Opcodes, Four Layers The opcode set is closed. New semantics cannot be introduced by sending a persuasive string. | Layer | Opcode | Stage | Exact operand keys | Constitutional role | |---:|---|---|---|---| | 0 | `BindEvidenceTag` | Fracture | `id`, `tag`, `coordinates` | Compare values at two distinct sealed-record coordinates using one positive or counter tag | | 0 | `RequestHumanReview` | Fracture | `id`, `reason`, `members` | Emit lawful non-operative review matter over at least one sealed record | | 0 | `RejectUnlawfulAct` | Fracture | `id`, `rejected`, `fault` | Preserve anti-knowledge rejecting a 64-hex matter identity or act digest under a registered fault | | 0 | `AssertLayoutRelation` | Forge | `id`, `relation`, `connector`, `from`, `to` | Prove `CONNECTS` over three sealed whole-node layout citations | | 1 | `AssertDuplicatePair` | Fracture | `id`, `pair`, `support`, `counter` | Link two records using admitted positive bindings and every admitted counter binding on that pair | | 2 | `AssertTransitiveClosure` | Fracture | `id`, `pair`, `chain` | Compose two to four admitted duplicate links into a simple path | | 3 | `ProposePartition` | Fracture | `id`, `groups`, `support` | Cover every sealed record exactly once with connected, non-contradicted groups | Dependencies point downward. Layer 0 observations feed Layer 1 pairs; Layer 1 pairs feed Layer 2 closure; Layers 1 and 2 support the Layer 3 partition. A missing, faulted, wrongly typed, cyclic, or pair-mismatched dependency cannot become support by being named confidently. `RequestHumanReview` accepts only `INSUFFICIENT_EVIDENCE`, `CONFLICTING_EVIDENCE`, `SOURCE_DEFECT`, or `OUT_OF_PROFILE`. It records a review obligation; it does not assert that a human has acted. `RejectUnlawfulAct` validates the rejected identity's digest shape and fault class; it does not need to prove that the rejected target exists. A submission cannot contain both an admitted terminal partition and an admitted review request. That is `submission_mixed_decision`, not a hedge with extra authority. ## 3. The Evidence Alphabet: 17 Tags and 12 Comparators EISA's tags are not natural-language hints. Each positive or counter tag binds to one deterministic comparator. The two inferential tags are reserved for transitive composition and are refused in `BindEvidenceTag`. | # | Evidence tag | Class | Comparator and exact predicate | |---:|---|---|---| | 1 | `PHONE_STRONG` | Positive | `DigitsEqual`: both digit strings have at least seven digits and are identical | | 2 | `PHONE_ONE_DIGIT_VARIANCE` | Positive | `DigitsOneSubstitution`: equal-length digit strings, Hamming distance exactly one | | 3 | `PHONE_MULTI_DIGIT_VARIANCE` | Positive | `DigitsMultiSubstitution`: equal-length digit strings, Hamming distance two or three | | 4 | `PHONE_COUNTRY_CODE_VARIATION` | Positive | `DigitsSuffix`: the shorter digit string is a proper suffix of the longer | | 5 | `INTERNATIONAL_PHONE_VARIATION` | Positive | `DigitsSuffix`: the same suffix law under a distinct declared evidence tag | | 6 | `EMAIL_LOCAL_MATCH` | Positive | `EmailLocalEqual`: valid trimmed/lowercased email shapes have equal local parts | | 7 | `DOMAIN_ROOT_MATCH` | Positive | `EmailDomainRootEqual`: the first domain label before the first dot is identical | | 8 | `DOMAIN_COUNTRY_TLD_VARIATION` | Positive | `EmailTldVariation`: first domain labels match and full domains differ | | 9 | `DOMAIN_ROOT_MISMATCH` | Counter | `EmailDomainRootDiffers`: first domain labels differ | | 10 | `NAME_STRONG` | Positive | `TextEqual`: lowercase whitespace-collapsed non-empty strings match | | 11 | `NAME_SIMILARITY` | Positive | `TokenOverlap`: nonblank values share at least one lowercase alphanumeric token of two or more characters | | 12 | `COMPANY_ROOT_MATCH` | Positive | `FirstTokenEqual`: first qualifying tokens match | | 13 | `COMPANY_REGIONAL_VARIATION` | Positive | `FirstTokenVariation`: first tokens match and normalized full strings differ | | 14 | `REGIONAL_SUBSIDIARY_PATTERN` | Positive | `FirstTokenVariation`: the same first-token variation law under a distinct declared tag | | 15 | `REGIONAL_VARIATION` | Positive | `FirstTokenVariation`: the same first-token variation law under a distinct declared tag | | 16 | `TRANSITIVE_CLOSURE_APPLIED` | Inferential | No field comparator; express with `AssertTransitiveClosure` | | 17 | `TRANSITIVE_CHAIN_DEPTH_1` | Inferential | No field comparator; express with `AssertTransitiveClosure` | The domain rule is intentionally simpler than a public-suffix analyzer: `domain_root` is the first DNS label. It does not strip `mail`, `www`, or country-code second-level domains. That limitation must remain visible; comparator semantics are declared policy, not universal identity truth. A comparator can answer only its exact predicate. `TokenOverlap` does not prove two people are identical. `FirstTokenVariation` does not prove two companies share a parent. `DigitsSuffix` does not establish ownership of a phone number. EISA determines whether the declared relation holds between the sealed strings—not whether the relation is sufficient evidence for a worldly identity decision. ### The 47 registered fault classes Faults are residue, never standing. The current `FAULTS` registry contains exactly these 47 codes: ```text Representation and bounds (16): act_not_object, unknown_field, missing_field, field_type, unknown_version, unknown_stage, stage_not_admitted, unknown_opcode, unknown_symbol, tag_class_mismatch, bound_exceeded, act_id_invalid, act_id_repeated, actor_mismatch, case_mismatch, evidence_ref_refused Coordinates and observations (11): pointer_foreign_artifact, pointer_malformed, record_unknown, member_repeated, field_absent, value_null, value_empty, value_not_string, value_shape_mismatch, observation_contradicted, pair_mismatch Dependencies and composition (8): dependency_missing, dependency_kind, dependency_faulted, dependency_cycle, support_missing, counterevidence_suppressed, chain_broken, chain_depth_exceeded Partitions and terminal decisions (7): partition_member_unknown, partition_member_repeated, partition_member_missing, partition_unsupported_group, partition_contradicts_support, partition_repeated, submission_mixed_decision Layout (5): layout_fact_absent, layout_scope_mismatch, layout_kind_mismatch, layout_citation_mismatch, relation_indeterminate ``` Malformed submissions can also produce whole-submission inadmissibility codes such as `submission_too_large`, `submission_unparseable`, `submission_not_array`, `submission_size`, and `submission_mixed_stage`. Those envelope outcomes are distinct from the 47 registered per-act fault classes that `RejectUnlawfulAct` may cite. ## 4. Sealed Cases: Two Substrates, One Artifact Identity An act may point only into a sealed case. EISA creates `eisacase_<32hex>` by hashing the Room name, a NUL separator, and the case's RFC 8785 JCS record digest. The `artifactHash` carried by every act is that record digest; the Room-qualified case ID is the graph address. Identical records sealed in two Rooms therefore have the same `artifactHash` but different case nodes. Neither Room can resolve the other's case through composition. ### File cases `seal_file` reads a JSON case through a task-scoped `ReadAuthority` under the Room's transition serialization, revalidates that authority, checks Room-project containment, enforces 4 MiB and 1,000-record ceilings, and applies the same strict parser. A case is either an array of record objects or `{ "records": [...] }`. Every record has a unique valid `id`. The graph record binds the canonical records, digest, source-path identity and source-byte SHA-256, actor, session, task, Room, and time. Loading recomputes the record digest before recognition. ### Passage cases `seal_passages` bridges Article 02 directly into EISA. Each input is a Room-issued `ev_` handle. The sealer: 1. prepares and validates the citation for the caller and task; 2. refuses handles from another Room or unrevealed collaborative-pass findings; 3. loads and reverifies the canonical view and live source; 4. cuts the exact Unicode-scalar passage using the handle's selector; 5. refuses textless gaps, empty selections, and passages over 1,024 characters; 6. attaches committed `layout` facts when a whole-node citation names a PDF layout element; 7. seals records containing the exact text and source provenance; 8. commits citation edges so later source reassessment reaches the case. A passage record's `/text` is the selected passage. Its `/source/...` fields bind Room, view digest, node ID, unit, start, end, source revision, source SHA-256, and parser digest. Every later `case::load` rechecks the JCS digest, source views, live source bytes, and sealed layout equality. A drifted source cannot be used for a new judgment. The two substrates share constitutional identity but not assurance. A task-scoped file case establishes JCS integrity over an admitted file. A passage case additionally binds each value to server-cut document evidence and layout provenance. Neither proves that the underlying records are worldly true. ## 5. The Four-Gate Verdict Machine | Gate | What the recognizer establishes | What it refuses to let the proposal do | |---|---|---| | **1. Syntax and stage isolation** | Strict JSON, exact keys, closed symbols, authenticated actor, unique IDs, bounded sizes, one act family, matching `artifactHash` | Smuggle duplicate fields, foreign opcodes, mixed stages, self-declared actors, or unbounded matter | | **2. Spatial containment** | Every coordinate resolves inside the named sealed case; every layout citation names the same committed view, revision, parser, and page | Point at foreign records, manufactured geometry, another Room's evidence, or unstored fields | | **3. Type and canonical identity** | Values are non-empty bounded strings, dependencies have the right opcode/layer, and admitted acts receive JCS consensus identities | Coerce shapes, cite faulted dependencies, or let prose and local naming change the proof identity | | **4. Substance and coverage** | Comparators hold, counterevidence stays visible, chains are simple paths, and partitions are exact, connected, and non-contradicted | Suppress contradiction, break a chain, omit members, split proven duplicates, or assert an unsupported cluster | Each act ends in one consequence class: - **`operative`** — a recognized act with a consensus identity. - **`lawful_non_operative`** — a valid review request, which creates no operative identity claim. - **`anti_knowledge`** — a valid rejection of other matter. - **`typed_fault`** — preserved residue with a fault code and no standing. A malformed whole submission is `inadmissible`. A recognized submission reports `decision: partition`, `review`, `relation`, or `none`; it does not acquire a blanket database-execution verdict. ### Counterevidence is submission-complete For `AssertDuplicatePair`, **at least one admitted positive `BindEvidenceTag` on the same pair is required**. Every listed support dependency must be such a binding. Every listed counter dependency must be an admitted counter binding on that pair. An empty counter list is valid only when the submission contains no admitted counter binding for that pair. Then EISA scans all admitted acts in the submission. If any counter binding on that pair is absent from `operands.counter`, the pair faults with `counterevidence_suppressed`. This is the minimum support and complete-counter rule—not a domain-specific claim that one similarity signal is sufficient proof of real-world identity. The proposer cannot bind `DOMAIN_ROOT_MISMATCH` in one act and then omit it when asserting identity in another. The check is deliberately scoped to the **current submission's admitted acts**; it is not a claim to have searched every historical record or external source in the world. ### Transitive closure is a path proof, not a slogan `AssertTransitiveClosure` cites two to four admitted `AssertDuplicatePair` links. Pairs are canonically ordered, but the walker can traverse each edge in either direction. It starts from either requested endpoint, requires every link to continue from the current record, tracks visited IDs to reject cycles, and must end at the other endpoint. This establishes that the cited simple path exists. It does not prove that transitivity is a safe business rule for every relationship. ### Partition verification is global across the admitted submission `ProposePartition` enforces four separate mathematical properties: 1. **Exact disjoint cover.** Every group member exists in the case, appears in at most one group, and every sealed record appears somewhere. Empty groups are meaningless. 2. **Cited support stays inside groups.** Every support reference must be an admitted duplicate pair or transitive closure. 3. **No admitted relation crosses a boundary.** EISA scans all admitted Layer 1/2 acts, including those omitted from `support`. If any links records in different groups, the result is `partition_contradicts_support`. 4. **Every multi-record group is connected.** A Disjoint-Set Union-Find structure unions cited support edges within each group. If members do not share one root, the result is `partition_unsupported_group`. A proposal cannot prove A=B and B=C, then place C outside the cluster because it dislikes the consequence. Nor can it place A, B, and C together without a cited path connecting them. ### Consensus identity removes narrative, not responsibility For an admitted act, `consensusIdentity` is the SHA-256 of an RFC 8785 basis containing the version, stage, opcode, case `artifactHash`, and canonical operands. Dependency IDs are replaced by the sorted consensus identities of the acts they name. Layout acts additionally bind their profile and committed geometric identity. Excluded from that identity are `actor`, `boundedReason`, `confidence`, `sourceEvent`, `evidenceRefs`, and local act IDs. Two proposers using different names, confidence scores, and explanations therefore receive the same consensus identity only when they prove the same structural transition over the same sealed case. Their raw acts remain separately attributable through `act_sha256`, Room actor, session, and submission identity. Canonical consensus does not create independent evidence. ## Exhibit A — A Valid Contact Submission The following synthetic submission binds two email relationships, asserts a pair, and proposes a complete partition over three sealed records. `artifactHash` is the case's bare digest. Values are resolved from the sealed case, not trusted from the prose. ```json [ { "version": "EISA_ACT_V1", "stage": "Fracture", "actor": "agent_er_worker_01", "opcode": "BindEvidenceTag", "operands": { "id": "act_bind_email_01", "tag": "EMAIL_LOCAL_MATCH", "coordinates": [ { "artifact": "records", "recordId": "rec_101", "pointer": "/email" }, { "artifact": "records", "recordId": "rec_102", "pointer": "/contact/workEmail" } ] }, "evidenceRefs": ["ev_8f31c09a4b2e11d08a7c65e41902b3a1"], "confidence": { "bps": 9400 }, "boundedReason": "Both sealed values share the local part m.chen across regional domains.", "sourceEvent": "evt_0123456789abcdef_a1b2c3d4", "artifactHash": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567" }, { "version": "EISA_ACT_V1", "stage": "Fracture", "actor": "agent_er_worker_01", "opcode": "BindEvidenceTag", "operands": { "id": "act_bind_tld_02", "tag": "DOMAIN_COUNTRY_TLD_VARIATION", "coordinates": [ { "artifact": "records", "recordId": "rec_101", "pointer": "/email" }, { "artifact": "records", "recordId": "rec_102", "pointer": "/contact/workEmail" } ] }, "evidenceRefs": ["ev_8f31c09a4b2e11d08a7c65e41902b3a1"], "confidence": { "bps": 9400 }, "boundedReason": "The first domain labels match and the full regional domains differ.", "sourceEvent": "evt_0123456789abcdef_a1b2c3d4", "artifactHash": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567" }, { "version": "EISA_ACT_V1", "stage": "Fracture", "actor": "agent_er_worker_01", "opcode": "AssertDuplicatePair", "operands": { "id": "act_dup_101_102", "pair": ["rec_101", "rec_102"], "support": ["act_bind_email_01", "act_bind_tld_02"], "counter": [] }, "evidenceRefs": ["ev_8f31c09a4b2e11d08a7c65e41902b3a1"], "confidence": { "bps": 9100 }, "boundedReason": "Two admitted positive bindings support this pair; no counter binding exists in this submission.", "sourceEvent": "evt_0123456789abcdef_a1b2c3d4", "artifactHash": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567" }, { "version": "EISA_ACT_V1", "stage": "Fracture", "actor": "agent_er_worker_01", "opcode": "ProposePartition", "operands": { "id": "act_part_final", "groups": [["rec_101", "rec_102"], ["rec_103"]], "support": ["act_dup_101_102"] }, "evidenceRefs": ["ev_8f31c09a4b2e11d08a7c65e41902b3a1"], "confidence": { "bps": 9100 }, "boundedReason": "A complete disjoint cover with one connected pair and one singleton.", "sourceEvent": "evt_0123456789abcdef_a1b2c3d4", "artifactHash": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567" } ] ``` The recognizer, not the narrative, decides whether the sealed values satisfy `EmailLocalEqual` and `EmailTldVariation`, whether the dependencies match, and whether the partition laws hold. ## Exhibit B — `counterevidence_suppressed` Suppose the same submission also admitted `act_bind_counter_03`, binding `DOMAIN_ROOT_MISMATCH` on the pair, but `act_dup_101_102` omitted it from `counter`. Its recognized-act report entry is: ```json { "index": 2, "id": "act_dup_101_102", "opcode": "AssertDuplicatePair", "class": "typed_fault", "identity": null, "act_sha256": "01928374655647382910abcdef1234567890abcdef1234567890abcdef123456", "canonical": "{\"actor\":\"agent_er_worker_01\",\"artifactHash\":\"4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567\",\"boundedReason\":\"Pair asserted without admitted counterevidence.\",\"confidence\":{\"bps\":9100},\"evidenceRefs\":[],\"opcode\":\"AssertDuplicatePair\",\"operands\":{\"counter\":[],\"id\":\"act_dup_101_102\",\"pair\":[\"rec_101\",\"rec_102\"],\"support\":[\"act_bind_email_01\"]},\"sourceEvent\":\"evt_0123456789abcdef_a1b2c3d4\",\"stage\":\"Fracture\",\"version\":\"EISA_ACT_V1\"}", "fault": { "code": "counterevidence_suppressed", "detail": "admitted counterevidence on this pair is not listed in counter: act_bind_counter_03" }, "observations": [], "comparator": null, "depends_on": [], "evidence_refs": [] } ``` The refusal is not a score penalty. It is the absence of standing for that pair. ## 6. Proof Lane A: Spatial Relations Without Visual Guesswork A flowchart can have perfect OCR and still produce a false graph. Recognizing the words inside two boxes does not establish which arrow connects them, which direction the arrow points, whether the tip touches a node or an enclosing frame, or whether “Yes” and “No” label the same geometric edge. EISA's layout family does not ask a model to settle those questions from prose. It judges `AssertLayoutRelation` over parser-committed `LayoutFact` values sealed from whole-node document citations. ### From PDF operators to integer evidence The layout projection decodes PDF content streams, tracks affine transforms and graphics-state nesting, interprets subpaths and paint operators, and expands Form XObjects within bounded recursion. Coordinates become integer **centi-points**—hundredths of a PDF point—only when finite and within the declared coordinate ceiling. The projection is bounded at 4,000 elements per page and 20,000 points per path. Cubic Bezier bounding boxes solve derivative extrema rather than treating control points as the painted extent. Course tracing samples each cubic segment 16 times and keeps at most 4,096 trace points. Small filled triangular marks no larger than 24 points can become arrowheads. The attachment logic associates marks with open connector ends, so direction and touch are measured from the actual marked end. Branch labels up to 20 points high are assigned to a connector when the traced course establishes ownership. A label beside a course must be within 12 points and every other course must be at least twice as far; otherwise it is tied and assigned to neither. The recognizer never accepts replacement geometry from the act. ### The `CONNECTS` proof `CONNECTS` is the only relation in the current layout alphabet. The act names three distinct sealed records—connector, from, and to—and its `evidenceRefs` must be exactly those three handles, each once. An outer frame is found through the parser-committed attachment and must also be sealed in the case, but it is not added as a fourth citation. The recognizer then enforces: 1. **Provenance.** All three records come from one Room, view digest, source revision, parser digest, and page. Each record's committed `LayoutFact.node` is exactly the node its evidence handle cites. 2. **Element types.** The connector is an open `line`, `polyline`, or `open_curve` with stroke paint and bounded start/end geometry. Both endpoints' targets are closed `rect`, `rects`, `closed_curve`, or `compound` shapes. 3. **Direction.** The committed direction is `toward_end` or `toward_start`, determined by one marked end. No direction or an ambiguous direction is `relation_indeterminate`. 4. **Unique attachment.** Neither end may carry parser-recorded ambiguity or no attached node. The parser's attached `from` and `to` node IDs must equal the act's cited nodes exactly. 5. **Touch.** A direct tail lies within `LAYOUT_TOUCH = 150` centi-points (1.5 pt) of the `from` box. A direct head or its mark touches the `to` box within 150; a marked head can use `LAYOUT_NEAR_TOUCH = 200` centi-points (2.0 pt) for mark-box contact. 6. **Frame proof.** When the parser routed an end through an outer frame, the frame and cited inner shape must independently satisfy the sealed frame law rather than letting a group boundary impersonate a target. A frame must be a sealed, stroke-only closed shape on the same view and page with exactly the committed box. The cited inner shape must be closed and filled. All four inner-minus-outer gaps must be nonnegative, and the **maximum** gap may not exceed `LAYOUT_FRAME_GAP = 600` centi-points (6.0 pt). The frame node must be sealed whole exactly once, and no other sealed closed shape may create an inner-shape tie. Page-wide uniqueness is reported as `parser_committed`; the recognizer separately checks contradictions present in the sealed case. The thresholds are therefore not interchangeable: 150 is direct touch, 200 is marked-head near-touch, and 600 is the maximum frame-to-inner gap. ### Branch labels enter relation identity A connector's committed `path_labels` qualify the relation. Their label-node IDs enter `consensusIdentity`, so a “Yes” edge and a “No” edge between the same two shapes are not the same constitutional relation. The recognizer reads label words only when the corresponding text node is sealed whole from the same view. `labels_tied` remains visible and is assigned to neither connector. The result is deliberately narrow: EISA can prove one directed `CONNECTS` relation under declared geometry. It does not infer `flows_to`, `branches_to`, `grouped_with`, electrical correctness, process safety, or the diagram author's intent. ## Exhibit C — An Operative Layout Relation This is the layout portion of a synthetic recognized-act report. The frame fields are observations in the current report shape, not an agent-authored `layout_proof` object. ```json { "index": 0, "id": "layout_connect_07", "opcode": "AssertLayoutRelation", "class": "operative", "identity": "9f2d4b8c1e7a0356d4c2b1a0987654321fedcba0987654321fedcba098765432", "fault": null, "observations": [ { "recordId": "ev_11111111111111111111111111111111", "node_id": "page4/e42", "shape": "polyline", "direction": "toward_end", "tail": [24100, 41200], "head": [28900, 42150], "tail_via": "direct", "head_via": "frame", "path_labels": [ { "label": "page4/e43", "bbox": [26000, 41400, 27800, 41700], "text_node": "page4/t9", "text": { "recordId": "ev_44444444444444444444444444444444", "text": "ADMITTED" }, "text_sealed": true, "on": "course", "beside_gap": null } ], "labels_tied": [] }, { "recordId": "ev_22222222222222222222222222222222", "node_id": "page4/e18", "shape": "rect", "bbox": [23000, 40000, 24500, 42400] }, { "recordId": "ev_33333333333333333333333333333333", "node_id": "page4/e29", "shape": "rect", "bbox": [28000, 41000, 29600, 43000], "head_marks": [ { "node": "page4/e40", "bbox": [28700, 41900, 28900, 42200] } ] }, { "end": "head", "recordId": "ev_55555555555555555555555555555555", "node_id": "page4/e5", "shape": "rect", "paint": "stroke", "bbox": [27500, 40500, 30100, 43500], "inner": "page4/e29", "gap": 500, "uniqueness": "parser_committed", "sealed_case_other_inner": 0 } ], "comparator": "LayoutConnects(touch=150,mark_touch=200,frame=parser_committed_inner,frame_gap=600)", "depends_on": [], "evidence_refs": [ "ev_11111111111111111111111111111111", "ev_22222222222222222222222222222222", "ev_33333333333333333333333333333333" ] } ``` The frame and label-text records are sealed in the case for verification, but the act's citations remain exactly connector, from, and to as required. ## Exhibit D — Two Distinct Partition Failures ```json [ { "index": 3, "id": "act_part_split_cluster", "opcode": "ProposePartition", "class": "typed_fault", "identity": null, "fault": { "code": "partition_contradicts_support", "detail": "'act_dup_101_102' asserts rec_101 and rec_102 are one record, but they are in different groups" }, "observations": [], "comparator": null, "depends_on": [], "evidence_refs": [] }, { "index": 3, "id": "act_part_unconnected_cluster", "opcode": "ProposePartition", "class": "typed_fault", "identity": null, "fault": { "code": "partition_unsupported_group", "detail": "group 0 is not connected by its cited relations" }, "observations": [], "comparator": null, "depends_on": [], "evidence_refs": [] } ] ``` The first splits records already proven equivalent. The second groups records without proving connectivity. They are different defects and receive different faults. ## 7. Proof Lane B: Semantic Switch First, Then the Entity-Resolution Stack Entity resolution adds a different danger: the evidence can contain direct identifiers, and the final merge can be difficult or impossible to reverse through the destination's ordinary product operations. The architecture protects the data **before any provider-exposed computation**, then protects the resolved baseline afterward. ```text Native contacts, companies, products, or other records → Stage 0: Semantic Switch protected representation → Stage 1: qBlock / Qdrant Edge candidate mesh over protected records → Stage 2: EISA Fracture identity adjudication → Stage 3: EISA Forge canonical state compilation → local restoration of authorized native values through the private crosswalk → Stage 4: Shield forward-only baseline defense → Room-governed destination transition ``` ### Stage 0 — Semantic Switch: protected computation before external AI Semantic Switch converts native values into structure-preserving fictional representations while retaining native rows, keys, and the private crosswalk locally. External AI providers and downstream reviewers receive the fictional records—not the declared raw identifiers. The downstream party returns partitions, unresolved membership, and optional field-source references; local restoration maps authorized references back to exact native values. This is the privacy-first ordering rule: **Semantic Switch runs before qBlock, Fracture, or any other computation that exposes records to an external model.** Candidate generation, adjudication, and canonical-value selection all operate on the protected representation. The accurate privacy description is **deterministic, structure-preserving pseudonymization**, not anonymity. The fictional surface preserves declared comparison structure while native data and restoration authority remain under client control. **Engineering trade-off note:** Semantic Switch is purpose-built for structured entity resolution, CRM deduplication (Keiki), and topological claim graphs. In those workloads, structural edges, blocking keys, and relational patterns carry the primary signal. Lexical identifiers such as names, emails, and account numbers are severed into local `SeveranceAntiKnowledgeObjects` before any external embedding provider or qBlock / Qdrant Edge computation, preserving useful comparison structure without shipping the native identifiers. | Check | Current contract and evidence | |---|---| | `SS01` | **Privacy-first stage responsibility.** Semantic Switch transforms native records before provider-exposed qBlock, Fracture, Forge, or Shield computation. Separate stages own identity judgment and destination mutation. | | `SS02` | **Whole blocks remain the unit of work.** Disjoint IDs, member existence, non-overlap, and in-block edge evidence are checked; no pair-request fragmentation or cross-block merge. | | `SS03` | **Input validation prevents silent loss.** Unknown fields, nested/numeric values in the declared string schema, duplicate keys/IDs, unknown members, and overlapping membership are refused; unblocked rows remain local and counted. | | `SS04` | **Relation-isomorphic fiction retains declared signals.** Scoped references, presence, exact and normalized equality, ordered components, and substitution, insertion, deletion, and transposition relations survive the transformation. | | `SS05` | **Per-block reproducible scope.** HMAC-SHA256 derivation binds schema, release, original block, canonical membership, and native content; no exported global dictionary, and no claim of cross-block unlinkability. | | `SS06` | **Complete partitions and explicit uncertainty.** Every emitted record appears exactly once across partitions and unresolved lists; omissions, duplicates, foreign records, and dual membership are refused. | | `SS07` | **Reference-based exact restoration.** A selection names an emitted record and issued field-value reference inside the same partition; local restoration retrieves the exact native value and source provenance rather than asking a model to reproduce it. | | `SS08` | **Release-inventory verification.** The expected `release_inventory` manifest is reconstructed from the authenticated crosswalk; admitted/held payload SHA-256 digests, counts, field-reference coverage, unexpected files/directories, and complete native accounting are checked. | | `SS09` | **Measured transport and relation fidelity.** Retained executions establish complete block, record, field-restoration, relabeling-invariance, and exact edit-relation accounting—not population matching accuracy. | | `SS10` | **Whole-block capacity.** Oversized blocks remain intact and held rather than being truncated or semantically fragmented; raising capacity requires a new preparation run. | | `SS11` | **Privacy and custody ceiling.** Native IDs, original rows, keys, crosswalks, and restored outputs stay local; keyed checks protect integrity, while host confidentiality and key custody remain deployment duties. | | `SS12` | **Reproducible packaging and handoff.** Pinned builds, launcher compatibility, manifest verification, new-output-location rules, and retained receipt and refusal artifacts support repeatable operation. | #### The retained transport-measurement record | Execution input | Native rows | Whole blocks | Fictionalized rows | Exact field restorations | Unblocked rows retained locally | |---|---:|---:|---:|---:|---:| | Inspected restaurant development data | 250 | 17 | 212 | 2,120 | 38 | | Synthetic stress data | 2,500 | 33 | 2,492 | 24,920 | 8 | | **Total** | **2,750** | **50** | **2,704** | **27,040** | **46** | All 2,750 inputs are accounted for. The retained Stage 1 executions made zero provider calls. A separate inspection checked 15,932 ordered component positions and the declared normalized-equality and component bijections. #### Relation-isomorphism and adversarial verification | Verification surface | Verified result | |---|---:| | Relabeling-invariance blocks | **20/20** | | Eligible native edit relations preserved exactly | **23/23** | | Records verified in each relation-isomorphism run | **53/53** | | Field values verified in each run | **529/529** | | Batch-restoration adversarial cases | **14/14** | | Transformation adversarial cases | **10/10** | | Provider calls during preparation | **0** | These are transport-fidelity and relation-preservation measurements. They do not establish matching accuracy, fresh-holdout performance, an AI advantage, anonymity, inference resistance, or independent validation. #### `SS08`: authenticate the release universe, not just a reply A valid response for one block proves nothing about another block hidden from the release. The release verifier therefore closes the inventory around the authenticated crosswalk: ```text authenticated private crosswalk → reconstruct expected release_inventory manifest → compare admitted and held block references, payload SHA-256 digests, counts, and bytes → verify field-reference coverage and complete native-record accounting → refuse missing/unexpected payloads, files, and directories → only then release the listed fictional artifacts → require complete block results before local restoration ``` Altered manifest counts, hidden blocks, unexpected native-data canaries, and unexpected public directories are refusal conditions. This is local software verification with a locally held key—not hardware attestation or control over every later copy. ### Stage 1 — qBlock: candidate nomination over protected records `qBlock`/Qdrant Edge consumes the Semantic Switch representation and partitions protected records into candidate blocks and edges. This narrows the comparison universe without sending the native rows through the provider-exposed candidate path. It does not prove that any two records match, and its recall ceiling remains explicit: a true pair absent from every nominated block cannot be adjudicated by a later stage. The blocks and edge evidence remain inside the nominated parent block. The Semantic Switch interface requires disjoint blocks; overlapping blocker output needs an explicit upstream ownership policy. Candidate nomination supplies a partition for adjudication—not an identity conclusion. ### Stage 2 — EISA Fracture: identity adjudication with dissent preserved Fracture is where candidate relations become constitutional acts. Role-separated defense, prosecution, and tribunal participants can nominate support and contradiction, but their votes are not the proof. The Four-Gate recognizer recomputes every binding and partition law. BFT-style panels and persistent role separation improve review structure and attributable dissent. They do not create independent model accuracy, and they cannot override a failed comparator or a hidden counter binding. The output is an identity partition, a review request, anti-knowledge, or typed residue—not a mutated customer record. ### Stage 3 — EISA Forge: canonical state is a separate decision Identity equivalence does not decide which values should survive. Forge consumes an admitted partition and compiles a canonical record field by field: - every surviving value names its source record and field pointer; - provenance and validity intervals remain inspectable; - declared field quorums and canonical claims decide support; - unsupported values do not win through confidence alone; - a conditional reversal plan records pre-state and post-state commitments before destination execution. This is the **Field Provenance Matrix** and reversible transition plan. Forge does not inherit ambient permission to run a CRM merge. Article 07's controller-fixed act and destination pre-state checks remain necessary. As Fracture partitions and Forge field selections return, Semantic Switch maps their authorized references through the private crosswalk to exact native values and source provenance. Restoration is local and reference-driven; no external model reconstructs or invents the original data. Shield then defends the resulting baseline, while any destination mutation remains separately governed. ### Stage 4 — Shield: the baseline moves forward or does not move A sealed baseline is not allowed to be rewritten by every later record. Shield evaluates new input through pre-provider switching and leak scanning, the declared embedding quorum, exact-baseline-run retrieval, and deterministic anchor gating. It then assigns one of three forward-only dispositions: - **`DEFLECT`** — contradictory or unverified matter is held away from the sealed baseline. - **`GROWTH`** — compatible evidence extends the provenance graph through a sealed forward path. - **`ESCALATE`** — genuine conflict enters autonomous quarantine and review rather than overwriting established state. No route mutates the sealed L3 baseline in place. Shield is temporal baseline defense, not a claim of universal entity accuracy or control over undeclared ingress paths. ## 8. From Duplicate Records to Evidence-Linked Relationship Graphs The commercial value is not “AI deduplication” as a slogan. It is a defensible identity substrate on which relationship paths can be inspected. ```text Semantic Switch protected records → qBlock candidate nominations → evidence-bound identity partitions → canonical records with field provenance → evidence-linked relationships → inspectable graph paths with freshness, relation type, and status → bounded vertical pilot ``` A six-hop query can be a useful exploration horizon. It is not proof of influence, consent, or authority. No graph path, by itself, authorizes outreach, enrichment, disclosure, or CRM mutation. Each resulting transition must still earn standing under the receiving domain's current rules. ## 9. The 2026 Research Landscape: Proposal Quality Is Not Transition Standing The current research supports a hybrid division of labor: models can improve nomination, comparison, and review, while deterministic systems must decide which declared predicates actually hold. The sources do not endorse AINA, and their results need their original scopes. | 2026 primary source | What it establishes | How EISA composes with it—without overclaiming | |---|---|---| | Teofili et al., [*Can we trust LLM Self-Explanations for Entity Resolution?*](https://arxiv.org/abs/2606.01210), June 2026 preprint | Across three LLMs, ten datasets, and multiple prompting strategies, self-explanations were often unstable, weakly faithful, and poorly aligned with counterfactual evidence. Established post-hoc methods were more trustworthy but costly. | `boundedReason` is audit context, not causal evidence. EISA admission depends on coordinates, comparators, dependencies, and counterevidence—not on the persuasiveness of an explanation. | | Papadakis et al., [*AgenticER: the next frontier in Entity Resolution*](https://arxiv.org/abs/2607.27435), July 2026 vision paper | Formalizes ER as sequential evidence gathering and decision-making with external sources and selective human involvement; proposes a reference architecture and evaluation dimensions. | Agentic ER needs constitutional checkpoints as iteration increases authority and cost. This is a vision paper, not an empirical superiority result. | | Smith et al., [*OpenSanctions Pairs*](https://arxiv.org/abs/2603.11051), February 2026 preprint, revised August 2026 | Provides 755,540 expert-labeled pairs from 293 sources across 45 jurisdictions. The rule baseline reached 91.3% F1, GPT-4o 99.0%, and a local DeepSeek model 98.2%; rules and LLMs failed in complementary ways. | Strong pairwise models are valuable proposers. The authors also point attention toward blocking, clustering, and uncertainty review—the exact places where candidate recall, counterevidence, partitions, and human review need separate proof. | | Kasprova et al., [*Too Polite to Disagree*](https://arxiv.org/abs/2604.02668), April 2026 preprint, revised July 2026; SIGDIAL 2026 | In six open-source model discussions, supplying peer-sycophancy rankings reduced the influence of sycophancy-prone peers and error cascades, improving final accuracy by 10.5 percentage points. | Multi-agent review can be improved, but social mitigation is not deterministic adjudication. EISA does not infer standing from a vote, whether the vote is naive or carefully managed. | | Deußer et al., [*On the Impact of Anonymization on the Performance of Large Language Models*](https://arxiv.org/abs/2609.11335), September 2026 preprint | Across five models and eleven benchmarks, anonymization effects were task- and model-dependent. Reversible methods preserving entity uniqueness outperformed irreversible approaches; explicitly prompting about anonymization did not help. | Structure-preserving pseudonymization belongs in the transport design, not in a prompt. Semantic Switch's measured restoration still does not prove anonymity or downstream matching accuracy. | | Deng et al., [*Circuit-MLLM*](https://arxiv.org/abs/2609.15668), September 2026 preprint; ECCV 2026 | Addresses the difficulty dense circuit schematics pose for multimodal models through device localization, path tracing, and topology-guided latent reasoning; reports a 25% higher average score than GPT-5.1 on its tasks. | Topological structure deserves explicit representation and verification. A stronger VLM remains a proposal mechanism; EISA's separate claim is a deterministic proof over committed PDF vector facts. | | Kogilathota et al., [*HALP*](https://arxiv.org/abs/2603.05465), March 2026 preprint; EACL 2026 | Shows that hallucination risk can be probed before generation across eight VLMs, with up to 0.93 AUROC on several models and architecture-dependent variation. | Early abstention and routing are useful. Risk probes do not prove which connector touches which shape; `AssertLayoutRelation` still requires committed geometry, unique attachments, and direction. | | [`CVE-2026-71847`](https://nvd.nist.gov/vuln/detail/CVE-2026-71847), August 2026 | Ruby JSON 2.20.0–2.21.x before 2.21.2 could dereference stale pointers and terminate when incomplete attacker-controlled objects containing duplicate keys reached the affected warning path. | Duplicate-key handling is a real parser-security boundary as well as a semantic ambiguity. EISA refuses duplicate keys before schema evaluation rather than delegating meaning to parser defaults. | The architectural conclusion is not that models are unnecessary. It is that **proposal quality and transition standing are different evidence claims**. Better models reduce the number of bad candidates requiring refusal. They do not remove the need for the refusal boundary. ## 10. The Continuous Wire into the Room EISA recognition and Room admission share one path. `recognize` previews a verdict; `submit` runs the same recognition logic and records the result. The caller is authenticated, the Room is admitted, the case is loaded and rechecked, and every evidence handle used by a non-faulted act is prepared for that Room. Foreign-Room evidence is refused rather than cited across an authority boundary. The recorded `squad.eisa.submission/1` payload binds: - `by` and native `session`; - `room` and optional `task`; - the case's bare `artifactHash`; - the complete recognition result; - submission time. Its content identity becomes `eisa_<32hex>`. Citation edges then link that immutable submission to its sealed case and every checked `ev_` handle. Resolved source values stay in the case; the recognized report carries their digests. ```text server-cut evidence (ev_) → sealed case (eisacase_) → Four-Gate recognized acts and consensus identities → Room submission (eisa_) → actor, native session, task, project, and wake authority → typed lease and monotonic generation → controller-fixed RequestedAct → fenced destination mutation and independent observation → strict original receipt or non-fabricating successor grant → terminal evidence ``` This is the fully wired machine, with authority changing form at every boundary rather than leaking across it. The EISA verdict supplies recognized matter for Room authority evaluation. It does not execute itself. The destination still has to accept the exact act under current ownership, pre-state, time, and release conditions. ## Exhibit E — The Current Room Ledger Envelope This synthetic `show` response illustrates the current ledger shape with one operative binding. The per-act report remains inside `result`. ```json { "submission": "eisa_7c4e91a02f8b3d651092a8b7c6d5e4f3", "room": "architecture-demo", "record": { "schema": "squad.eisa.submission/1", "by": "agent_er_worker_01", "session": "worker-mcp:00000000-0000-0000-0000-000000000001", "room": "architecture-demo", "task": "TSK-2026-ER-001", "artifactHash": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567", "result": { "version": "EISA_ACT_V1", "profile": "eisa.fracture.contact/1", "class": "recognized", "raw_sha256": "e01928374655647382910abcdef1234567890abcdef1234567890abcdef12345", "artifactHash": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567", "case": "eisacase_3b8f19c04e2a7d6518293a4b5c6d7e8f", "acts": [ { "index": 0, "id": "act_bind_email_01", "opcode": "BindEvidenceTag", "class": "operative", "identity": "6b8f1029384756ab12cd34ef567890124a9c8e12f03b7d651928374655647382", "act_sha256": "01928374655647382910abcdef1234567890abcdef1234567890abcdef123456", "canonical": "{\"actor\":\"agent_er_worker_01\",\"artifactHash\":\"4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567\",\"boundedReason\":\"Both sealed values share one local part.\",\"confidence\":{\"bps\":9400},\"evidenceRefs\":[\"ev_8f31c09a4b2e11d08a7c65e41902b3a1\"],\"opcode\":\"BindEvidenceTag\",\"operands\":{\"coordinates\":[{\"artifact\":\"records\",\"pointer\":\"/email\",\"recordId\":\"rec_101\"},{\"artifact\":\"records\",\"pointer\":\"/contact/workEmail\",\"recordId\":\"rec_102\"}],\"id\":\"act_bind_email_01\",\"tag\":\"EMAIL_LOCAL_MATCH\"},\"sourceEvent\":\"evt_0123456789abcdef_a1b2c3d4\",\"stage\":\"Fracture\",\"version\":\"EISA_ACT_V1\"}", "fault": null, "observations": [ { "recordId": "rec_101", "pointer": "/email", "value_sha256": "1111111111111111111111111111111111111111111111111111111111111111" }, { "recordId": "rec_102", "pointer": "/contact/workEmail", "value_sha256": "2222222222222222222222222222222222222222222222222222222222222222" } ], "comparator": "EmailLocalEqual", "depends_on": [], "evidence_refs": ["ev_8f31c09a4b2e11d08a7c65e41902b3a1"] } ], "classes": { "operative": 1 }, "decision": "none", "partition": null }, "at": "2026-10-31T21:50:00Z" } } ``` The immutable node is constitutional proof matter for the next boundary. It is not a destination receipt. **Evidence availability:** the EISA implementation, recognition artifacts, and Semantic Switch measurement record are `CONTROLLED` for qualified, method-scoped technical review, with adverse findings permitted. They are not represented as unrestricted public downloads or independent validation. Performance and privacy claims must be evaluated against the declared record population, parser/profile, and deployment threat model. ## Bring One Bounded Protected-Computation Problem The **Semantic Switch & Entity Resolution Challenge** applies this constitutional layer to one bounded contact, company, product, or other duplicate-record workflow. The demonstration follows Semantic Switch → qBlock → Fracture → Forge → local restoration → Shield and returns a relation-fidelity report, comparator and counterevidence map, exact-restoration record, and the predicates required before any production destination mutation. The challenge excludes custom schema engineering, production CRM write-back, lead-list generation, universal privacy or matching-accuracy claims, and independent validation. Deployment and paid pilots are separately scoped; independent evaluation remains outside the commercial ladder through [Article 09](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/). ## The Article 03 Claim Card ```text Mechanism: Semantic Switch Protected Computation + EISA Constitutional Act Recognition (native-data minimization, relation-isomorphic fictional records, qBlock candidates, Fracture, Forge, local exact restoration, Shield, strict parsing, sealed cases, pointer law, deterministic comparators, Four-Gate recognition, layout proof, and Room ledger) Failure it governs: Native identifiers exposed to external AI computation; comparison structure destroyed by destructive redaction; ambiguous or schema-valid model prose becoming authority; fabricated coordinates and geometry; suppressed counterevidence; broken transitive claims; incomplete or contradictory partitions; unprotected external comparison; and model votes impersonating proof. Positive path: native records held locally → Semantic Switch protected representation → qBlock / Qdrant Edge candidates over protected records → Fracture partitions and Forge canonical selections → local exact restoration through the private crosswalk → Shield forward-only baseline defense → doc_ / ev_ evidence and eisacase_ sealed records → strict EISA_ACT_V1 acts and deterministic recognition → squad.eisa.submission/1 as eisa_ → Room authority, RequestedAct, fenced execution, terminal receipt. Refusal / non-operative path: Whole-submission inadmissibility for malformed envelopes. typed_fault with one of 47 registered per-act fault classes. counterevidence_suppressed and partition_contradicts_support. partition_unsupported_group and relation_indeterminate. lawful_non_operative RequestHumanReview and anti_knowledge rejection. Foreign-Room, unrevealed, drifted, textless, or out-of-scope evidence refused. Evidence status: EISA: Implemented and Author-Verified on declared wire, graph, and geometry paths. Semantic Switch: Implemented, Measured, and Author-Verified on retained executions. 2,750 inputs; 50 whole blocks; 2,704 fictionalized records; 27,040 exact field restorations; 46 unblocked records retained; 20/20 relabeling-invariance blocks; 23/23 exact native edit relations; 53/53 records and 529/529 field values verified per isomorphic run; 14/14 batch and 10/10 transformation adversarial cases; zero provider calls in retained Stage 1 executions. Not Independently Validated; external research does not endorse AINA. Verifier and artifacts: strict::parse, pointer::parse/resolve, Comparator::holds, recognize::recognize, case::load, layout::LayoutFact, Semantic Switch authenticated crosswalk and release-inventory verifier, relation-isomorphism and exact-restoration reports, batch and transformation adversarial receipts, consensusIdentity, eisacase_, eisa_, cor_/coa_ where views correspond, and the Room citation graph. Scope and claim ceiling: Comparators are declared policies, not universal identity semantics. Counterevidence and partition scans cover the current admitted submission, not every historical or external fact. Sealed evidence is not source truth; consensus identity is not independence. Semantic Switch establishes transport and relation fidelity, not anonymity or ER accuracy. Layout proves only CONNECTS under committed geometry and declared thresholds. EISA recognition is not a destination mutation or independent validation. Next stage in the machine: Article 03 protects native computation and recognizes constitutional acts. Article 04 binds them to live actors, sessions, tasks, wake delivery, acceptance, and current Room authority before execution. Commercial doorway: Semantic Switch & Entity Resolution Challenge — bring one bounded contact, company, product, or other duplicate-record workflow; receive a replayable protected-computation demonstration, relation-fidelity report, comparator and counterevidence map, exact-restoration record, and explicit production-integration predicate list (Article 09 · me@blayneaina.com). ``` ## Before the Room Can Govern, the Act Must Mean One Thing Return to the three agents voting for a merger. The conventional system can report that the vote was unanimous, the JSON matched a schema, and the judge was 97% confident. It still cannot answer which exact values were compared, whether the domain contradiction was included, whether the chain closes, or whether every candidate appears exactly once in the result. EISA answers those questions mechanically. The duplicate keys are gone. The case is sealed. The pointers resolve. The comparator runs on the actual values. The counter binding stays visible. The transitive path is walked. The partition is checked against every admitted relation and reconstructed with Union-Find. The surviving act receives a canonical identity and enters the Room with actor, session, task, and citation continuity. Only then does the proposal become something a coordination plane can safely govern. > **Models propose. EISA recognizes. The Room governs. Execution is fenced. Effects are observed. Every transition closes with evidence.** Article 04 takes up the next non-equivalence: a recognized act still needs a live authority domain. Git history and chat transcripts cannot supply one. ## Frequently Asked Questions ### Why is JSON Schema or constrained decoding insufficient to govern AI actions? Constrained decoding establishes shape. It does not resolve record-relative pointers against a sealed case, run deterministic comparators on the values, reject faulted dependencies, force counterevidence into the decision, or prove that a partition is complete, connected, and non-contradicted. EISA also refuses duplicate keys before schema evaluation. ### How does EISA prevent an AI agent from suppressing conflicting evidence? Every admitted counter binding on a pair must appear in that pair's `counter` array. EISA scans all admitted acts in the submission and returns `counterevidence_suppressed` if one is omitted. Separately, any admitted identity relation crossing a proposed partition boundary causes `partition_contradicts_support`. ### Does `consensusIdentity` prove independent agreement? No. It proves that two acts encode the same structural transition over the same sealed case after actor, prose, confidence, source event, evidence references, and local IDs are removed. Shared evidence, shared models, and shared mistakes can still produce the same identity. ### Why does Semantic Switch run before qBlock? Because qBlock and later stages may use external AI computation. Semantic Switch transforms native values first, so Qdrant Edge and qBlock operate on structure-preserving fictional records rather than the declared raw identifiers. Native rows, keys, and the private crosswalk remain local while qBlock nominates candidates for Fracture. ### Is Semantic Switch anonymization? No. It is deterministic, structure-preserving pseudonymization, with native rows, keys, crosswalks, and restoration mappings retained locally. Its measured exact restoration and relation fidelity do not prove anonymity, cross-block unlinkability, inference resistance, or matching accuracy. ### How does EISA verify flowchart connections without trusting VLM prose? It seals parser-committed PDF vector facts and evaluates only `CONNECTS`: one stroked open connector, two closed shapes, one marked direction, unique attachments, touch thresholds of 150/200 centi-points, and a maximum 600-centi-point frame gap when an outer frame is involved. Ambiguity becomes `relation_indeterminate`. ### Does an operative EISA verdict execute the proposed transition? No. It creates recognized constitutional matter and commits it to the Room as `eisa_<32hex>`. Actor eligibility, task and lease authority, controller-fixed execution, destination pre-state, effect observation, and terminal closure remain separate gates. ## Continue the Series - **Series index:** [From Proposal to Consequence: The AINA Transition-Governance Series](https://blayneaina.com/) - **Previous:** [Article 02 — Zero-Trust Document Ingestion and Server-Cut Evidence](https://blayneaina.com/02-zero-trust-document-ingestion-and-server-cut-evidence/) - **Next:** [Article 04 — The Room: Why Git and Chat Are Not Coordination Planes](https://blayneaina.com/04-the-room-why-git-and-chat-are-not-coordination-planes/) - **Deployment and independent review:** [Article 09 — Bring One Real Boundary](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/) --- ## Primary 2026 Sources - [Teofili et al. — *Can we trust LLM Self-Explanations for Entity Resolution?*, `arXiv:2606.01210`](https://arxiv.org/abs/2606.01210) - [Papadakis et al. — *AgenticER: the next frontier in Entity Resolution*, `arXiv:2607.27435`](https://arxiv.org/abs/2607.27435) - [Smith et al. — *OpenSanctions Pairs: Large-Scale Entity Matching with LLMs*, `arXiv:2603.11051`](https://arxiv.org/abs/2603.11051) - [Kasprova et al. — *Too Polite to Disagree: Understanding Sycophancy Propagation in Multi-Agent Systems*, `arXiv:2604.02668`](https://arxiv.org/abs/2604.02668) - [Deußer et al. — *On the Impact of Anonymization on the Performance of Large Language Models*, `arXiv:2609.11335`](https://arxiv.org/abs/2609.11335) - [Deng et al. — *Circuit-MLLM*, `arXiv:2609.15668`](https://arxiv.org/abs/2609.15668) - [Kogilathota et al. — *HALP: Detecting Hallucinations in Vision-Language Models without Generating a Single Token*, `arXiv:2603.05465`](https://arxiv.org/abs/2603.05465) - [NVD — `CVE-2026-71847`, Ruby JSON duplicate-key warning-path use-after-free](https://nvd.nist.gov/vuln/detail/CVE-2026-71847) --- ## Publication Metadata **Title tag:** Semantic Switch and EISA: Protected Computation and Entity Resolution **Meta description:** Semantic Switch protects native data before external AI computation, then qBlock, Fracture, Forge, and Shield turn fictionalized records into governed entity state. **Series navigation:** [Series index](https://blayneaina.com/) · Article 03 of *From Proposal to Consequence: The AINA Transition-Governance Series*. **Previous installment:** [Article 02 — Zero-Trust Document Ingestion and Server-Cut Evidence](https://blayneaina.com/02-zero-trust-document-ingestion-and-server-cut-evidence/). **Next installment:** [Article 04 — The Room: Why Git and Chat Are Not Coordination Planes](https://blayneaina.com/04-the-room-why-git-and-chat-are-not-coordination-planes/). ================================================================================ # ARTICLE §04: The Room: Why Git and Chat Are Not Coordination Planes Canonical URL: https://blayneaina.com/04-the-room-why-git-and-chat-are-not-coordination-planes/ # The Room: Why Git and Chat Are Not Coordination Planes *Source history records what changed. Chat records what was said. The Room decides who may read, act, receive, stop, and close work now—and makes every transition prove it.* **By Blayne Aina · Principal Systems Architect and creator of AINA · October 31, 2026** --- A commit proves what source history recorded. A chat message proves what someone said. Neither proves who may mutate a live resource now, whether the intended native session received the assignment, whether the worker saw every causal parent before acting, whether a queued instruction still has authority, or whether the resulting work passed acceptance. That is the coordination problem. It does not disappear because agents have worktrees, checkpoints, tool schemas, and an execution log. Article 03 produced recognized constitutional acts and committed them as `eisa_<32hex>`. Article 04 is the machine that gives those acts a live authority domain: **The Room**. > **The Room is not a chat channel. It is a coordination plane that binds territory, identity, observation, delivery, execution, cancellation, and completion into one revalidated authority chain.** ## The Whole Machine, Not a Wake Queue The Room includes eleven cooperating mechanism families grouped into six architectural domains. Nine receive numbered subsystem sections below; worker staging and controller-side brokerage are developed within native execution and cancellation rather than numbered separately. Treating the Room as a messaging API leaves most of its assurance on the table. ```text [1. TERRITORY & PRINCIPAL PINNING] state-tree identity + 192-bit combined Win32 project object IDs CSPRNG enrollment + thread-local SessionGuard + native conversation binding EffectiveTier clamp + multiplexer verb probing + stdio/pre-model egress isolation │ ▼ [2. PER-ROOM GRAPH SHARDS & VISIBILITY] immutable scope binding + content-addressed artifacts monotonic parent-withholding closure + multi-Room composition reparse-safe snapshot recovery │ ▼ [3. ROOM SYNC ENGINE & PRE-ACTION WORK GATE] per-Room SQLite WAL sync index + handle-pinned journal reconciliation conversation/view checkpoints + recursive causal-parent verification atomic turn commit + Fold Rule + post-read revalidation │ ▼ [4. CONTRIBUTION GRAMMAR & SEALED WAKE PERMITS] note / reply / say / send / handoff / done with different authority effects 18-field room-wake-job-v3 + dispatch and split-lock revalidation per-recipient evidence delivery + complete governing-source coverage │ ▼ [5. NATIVE SESSIONS, CONFINED WORKERS & ACTIVE CANCELLATION] Session Factory v1 + five delivery/reply proof states row/column staging + controller-side MCP broker held in-flight operations + 1 Hz Job observer + two-phase stop outbox │ ▼ [6. TASK LINEAGE, PRIVILEGED ACCEPTANCE & RUNTIME INVARIANTS] 16-level delegation ancestry + 14-key completion revalidation 7 engine-evaluated acceptance kinds exactly_once + well_formed + exclusive + causal + resolution ``` The result is a continuous authority path: ```text EISA-recognized act → live actor, credential, native conversation, project, and Room scope → confirmed observation of the live causal frontier → task-bound contribution and sealed wake permit → dispatch-time and worker-claim authority revalidation → confined execution and native reply proof → destination and artifact acceptance → terminal evidence or explicit in-doubt state ``` ## One Assignment Through the Machine Consider a coordinator assigning renewal-evidence review to `reviewer-beta`. The path is not “write a prompt and hope the right agent answers”: 1. **Territory and principal.** The coordinator's credential, native conversation, project object, and Room scope are re-derived. The recipient's registration is current. 2. **Recognized work.** An EISA submission supplies the constitutional act and its evidence handles. The task binds that work to an exact project and acceptance contract. 3. **Observed readiness.** Before the coordinator can send, the Work Gate proves the conversation checkpoint covers the live frontier and every required causal parent is confirmed. 4. **Sealed assignment.** `send` writes the message and obligations, then seals an 18-field wake permit with the recipient, task, evidence, registration, Room authority, control epoch, and original deadline. 5. **Dispatch revalidation.** The Room rechecks the permit, roster, registration, task ancestry, governing source, project, admission, epoch, and remaining time. Evidence is admitted specifically for `reviewer-beta` and delivered with a fresh `wake_delivery` receipt. 6. **Split-lock worker claim.** The worker first validates context outside the transition lock, then rereads the job inside it and requires the unchanged binding digest before claiming. 7. **Native delivery.** The bound native session receives the prompt. Dispatch, body delivery, body visibility, wrong-body refusal, and reply remain separate facts. 8. **Reply and production.** The worker returns a correlated reply and produces the required artifact. Process exit alone would leave the obligation unresolved. 9. **Privileged acceptance.** The engine measures the artifact digest, distinct evidence origins, destination effect, and original worker receipt against the declared acceptance checks. 10. **Terminal closure.** Completion revalidates all 14 task-authority keys, releases the scope through the structured path, and records `task_done`. If registration changes before dispatch, the old wake stops with `wake_recipient_registration_changed`. If cancellation arrives after launch, new effects stop, the owned Job is terminated, and only `ActiveProcesses == 0` confirms quiescence. Any destination operation whose outcome is uncertain remains `held_until_reconciled`; it is not relabeled “cancelled and safe.” ## 1. Git and Chat Are Necessary—and Insufficient Git is an excellent source-history plane. Chat is an excellent human and model communication surface. The Room does not replace either. It prevents their artifacts from impersonating live authority. | Existing primitive | Native value | Additional question the Room answers | |---|---|---| | Tool-call schema | The proposal has a valid shape | Is this exact act recognized, currently standing, and inside the caller's verb and tier authority? | | Graph/workflow checkpoint | A prior plan can resume | Is that plan valid for the current target, conversation, task lineage, registration, generation, and control epoch? | | Git worktree | Source histories and working directories are separated | Who owns the live resource and may produce this effect now? | | Model reviewer or vote | More cognition and critique | Which deterministic predicate admits or refuses the act? | | Execution log | A tool reported an outcome | Did the destination and receipt chain verify it, and did the responsible native session return it? | | Idempotency key | A retry can be deduplicated | Was the effect committed, and was the original obligation actually answered? | Clean merges do not prove semantic non-interference. Two agents can edit different lines and violate a shared invariant. They can modify separate files whose combination is unsafe, call the same external API twice, or make locally valid decisions against the same stale pre-state. Mature coordination systems—consensus services, compare-and-swap stores, MVCC databases, lease and fencing protocols, durable workflow engines—absolutely can answer these questions. The claim is not that AINA invented coordination. The claim is that **source history and conversation are the wrong artifacts to carry its full burden**. A conventional stack that enforces the same canonical transition through the same current predicates earns the same functional credit. ## 2. Subsystem 1: Territory and Principal Identity Authority starts by proving which state tree and which physical project object the act concerns. ### Split-brain state-tree refusal Before the engine opens its state tree, `validate_state_dir` requires the state-directory identity marker. The marker carries an `id` and `canonical_path`; the machine bootstrap supplies the canonical instance and root. A missing marker, an unreadable bootstrap, a different instance ID, a noncanonical path, or a mismatched `SQUAD_ROOT` fails closed. This prevents two `.squad` trees from independently accumulating tasks, leases, credentials, and Room history while believing they are one machine. It is not a path prettification step. It is a split-brain admission gate. ### Project grants pin path and kernel object identity An owner-opened `ProjectGrant` binds: - a unique project ID and bounded name; - the requested and canonical root paths; - `root_object_id`, a 192-bit combined identity: a 64-bit volume serial number plus the 128-bit NTFS/ReFS file ID; - opening actor and time; - an optional governing-source grant and its reading policy. The directory is opened with backup semantics and queried through `GetFileInformationByHandleEx(FileIdInfo)`. Every live authorization rechecks both canonical path and object identity. A junction, symlink, replacement, or redirected root produces `project_root_redirected`. Project opening refuses non-directories, filesystem roots, broad owner containers, control and recovery storage, worker roots, and overlapping project territories. Roster membership does not widen this boundary. `ReadAuthority` applies the same provenance discipline to shared briefcases, contexts, artifacts, braindumps, and proposals: project-scoped records go to admitted project readers, coordination records to the appropriate Room roster, and legacy unscoped records only to their original author. ### CSPRNG enrollment without persistent plaintext secrets Agent enrollment generates a 32-byte secret with Windows `BCryptGenRandom`. The controller stores only `secret_sha256`, issuance, revocation, and target metadata. Host configuration receives the actual token for connection authentication; enrollment receipts do not return it. Multi-agent enrollment restores exact preimages if any step fails. Every call revalidates the current credential digest, so rotation and revocation affect live connections without a restart. ### Thread-local sessions and native conversation identity `SessionGuard` binds one authenticated session to one synchronous dispatch. Its `PhantomData>` prevents the authority guard from moving across threads, and `Drop` removes the active frame during normal return or unwind. `IS_HTTP_WORKER` closes the privilege-fallback path: an HTTP worker thread without an active request session never inherits the process launch principal. For supported Codex hosts, `bind_tool_request_context` reads `params._meta.threadId` outside model-authored arguments, requires a non-nil canonical UUID, and binds `codex:` as the native conversation. The authenticated host is the metadata trust boundary. This is not cryptographic attestation against every process that might possess its credential, and Article 04 does not claim otherwise. Self-identifying argument fields must match the bound principal. Mismatches produce `actor_mismatch` and an integrity signal rather than letting a model nominate a more privileged author. ### Tier clamping includes the route, not just the tool name The effective tier is the minimum of: ```text session-declared tier owner per-agent max_tier fleet-wide control ceiling ``` Owner policy and the fleet ceiling are re-read on every invocation. The multiplexer guard then derives verb routes from the MCP schema build table using `VERB_PROBE`. A denial of `room_note` therefore also blocks `squad_room` with `{"action":"note"}`. Blocking the friendly tool name while leaving the generic route open is not policy enforcement. ### Stdio and pre-model egress are separate boundaries The MCP transport duplicates its original input/output as private handles and clears inheritance flags. Process standard input is detached, and standard output is redirected to diagnostics so a child cannot corrupt the JSON-RPC wire or retain the parent's pipe accidentally. Before a result or error reaches model context, every string passes `redact_checked`. If the redaction policy cannot be established, the response becomes `tool_result_withheld`. The tool may already have run; the uninspected result still does not escape. ## 3. Subsystem 2: The Room Graph Is a Visibility and Recovery Boundary Each Room owns a physically isolated graph shard derived from the workspace instance and Room name: ```text room_graphs// manifest.json binding.json nodes.jsonl artifacts.jsonl errors.jsonl blobs/. sync_index.db ``` `binding.json` permanently fixes `Scope::Coordination` or `Scope::Project { id }`. Historical artifacts cannot be rebound to another project or scope class; `graph_project_rebind_refused` requires a new Room identity. Policy mutation uses `expected_sha256` optimistic concurrency, atomic preimage retention, and `prepared`/`applied` journal stages. Payloads over 8 KiB spill into content-addressed blobs. MCP call text over 24 KiB compacts to a blob with a 4 KiB preview. Sensitive credential arguments are omitted from audit projections rather than preserved for convenience. Legacy imports require quiescence: a live pre-shard writer triggers `room_graph_legacy_writer_detected` instead of racing the new graph. Hash-chained artifacts make alteration detectable; they do not prove non-deletion or independent validation. ### Withholding closes over descendants `ProjectVisibility` first filters by project and evidence visibility, including collaborative-pass staging. It then computes a monotonic parent-withholding closure: if any parent is withheld, every child evaluation, citation, or derived record is withheld too. Legacy unscoped MCP calls degrade to `project_metadata_only`, removing arguments, responses, read/write sets, and artifact references. A hidden parent cannot leak through a visible child. ### Composition is intersection, not union, of authority `squad.room_composition.v1` composes a bounded set of Rooms under an owner grant. Every constituent must remain admitted. The effective reader scope is the intersection of the composition grant, each Room grant, current membership, and active project. A failed constituent yields `incomplete_source_set`, not a partial composition that silently drops its restrictions. Pagination binds `request_sha256` to the composition, reader, filter, limit, and source snapshots. If a source changes between pages, `graph_page_snapshot_changed` prevents a mixed-generation export. Composition confers no execution authority. ### Recovery refuses reparse shortcuts and preserves causal order Snapshot recovery uses `squad.room_snapshot.v1`. Copying refuses symlinks and NTFS reparse points. A verified staged tree is fingerprinted, the previous live tree is preserved, and promotion occurs only after current-state and staged-state checks. Interrupted recovery remains pending and blocks ordinary graph access until the exact operation resumes. ## 4. Subsystem 3: The Room Sync Engine and the Pre-Action Work Gate A coordination plane cannot rely on an agent remembering what it has read. It must measure the observation. ### A durable per-Room sync index Each Room has `sync_index.db` under `squad.room_sync_index.v1`. Every connection enforces and verifies: - `journal_mode = WAL`; - `synchronous = FULL` (`2`); - `foreign_keys = ON`; - `busy_timeout = 5000 ms`. Nine relational tables hold the rebuildable event projection, context head, graph links, receipts, checkpoints, snapshots, causal edges, confirmed events, and local event embeddings. The event index is rebuildable from the authoritative journals; receipts and checkpoints survive index repair. ### Handle-pinned reconciliation Reconciliation opens `EVENTS.jsonl`, verifies its Win32 file identity, and retains a trailing anchor of up to 4,096 bytes with its SHA-256. Journal replacement or incompatible rotation is `room_sync_journal_rotated`, not an invisible reset. The reader advances in 1 MiB batches with an 8 MiB per-record ceiling, verifies the `prev_hash → hash` chain and monotonic Hybrid Logical Clock, and computes local embeddings outside the SQLite write transaction. Expensive semantic work never holds the database lock. ### Checkpoints bind actor, conversation, credentials, Room, and view A delivery checkpoint is not just `agent_x read_offset = N`. It binds a `conversation_key` derived from workspace instance, bound actor, registration digest, native conversation UUID, and Room, plus a `view_gen` derived from evidence-read admission, reset state, project, participants, and readers. Changing IDE thread, rotating registration, resetting the Room, or changing the visible project invalidates the checkpoint and requires fresh synchronization. Pages are limited to 24 KiB including the MCP envelope. Oversized events split into 1,024-byte UTF-8-safe `item_chunk` fragments. Only the final `item_complete` fragment advances the journal offset, so an interrupted read re-serves the item rather than pretending it was consumed. Redaction runs again at confirmation. If policy changed the delivered representation, confirmation fails with `room_sync_delivery_policy_changed`. A receipt tip that is not an ancestor of the current journal fails with `room_sync_receipt_forked`. ### The Work Gate requires frontier and causal readiness Nine participant verbs—`note`, `say`, `reply`, `send`, `done`, `blocked`, `decline`, `handoff`, and `circle`—must pass the pre-action Work Gate. Two conditions must hold: 1. The conversation checkpoint is ready and `confirmed_offset` covers the live coordination frontier. 2. `verify_causal_parents_confirmed` proves that the explicit parents and every ancestor returned by a recursive SQLite CTE over `causal_edges` are confirmed for that conversation and view. A matching confirmed event or a confirmed receipt covering the parent's offset can establish coverage; the bound identity's own events are handled explicitly. Failure returns `room_sync_required` with the next authorized page. The agent must read and confirm the missing context before acting. At commit, the writer lock revalidates the frontier and conversation, assigns HLC and hash lineage, and appends the staged events atomically. A foreign arrival triggers `room_sync_concurrent_arrival` rather than letting a stale turn slip through. The **Fold Rule (D1)** advances the author's confirmed offset only when no foreign coordination event exists between the old checkpoint and the newly written event. The author folds its own work, never someone else's unread work. Named system writers have separate authority and cannot emit participant message kinds. Finally, `RoomEventView::revalidate()` rechecks actor, project visibility, Room authority, and task-source digest after a read, catching authority changes that occurred while the view was being assembled. ## 5. Subsystem 4: Contribution Verbs Have Different Authority Effects The Room is not freeform chat with locks added later. Its verbs deliberately separate status, answer, assignment, transfer, and completion. | Verb | Wake and task rule | Constitutional effect | |---|---|---| | `note` | Never wakes; visible `@mentions` refused | Status or identity telemetry without mentions, obligations, or dispatch | | `reply` | Never wakes; explicit `reply_to` required | Resolves an addressed mention or obligation after validating that it belongs to the caller | | `say` | Coordination-only wake behavior; rejects `--task`; project wake delivery requires `send` | Conversational turn with strict targeting and ambiguity handling | | `send` | Task required in project Rooms; creates sealed wake jobs | Two-phase message/obligation commit and authorized worker delivery | | `handoff` | Bounded invitation; project no-wake inbox or coordination delivery | Transfers task ownership without transferring a lease | | `done` | Task required | Runs privileged acceptance and resolves only proven addressed items or verified successors | An untargeted `say` reply cannot resolve a mixed set of pending obligations. If open items trace to more than one source event, the Room refuses with `say_reply_target_ambiguous` before writing anything. `send` prevalidates the task and every evidence handle before its first event. A citation belonging to another Room produces `cross_room_evidence_refused`. Flush 1 commits the `room_message`; Flush 2 commits the mention and obligation records. A cooldown-deferred wake is still a sealed permit, and the 2,000 ms damper clock starts at dispatch completion rather than at initial gate passage. `handoff` records `lease_transferred: false` and `fresh_claim_required: true`. Ownership can move; the old kernel and lease authority cannot be inherited as a side effect. `done` can resolve items from a superseded task only when `verified_successor` proves a non-weaker replacement: same owner, source, project, requirements, and non-empty acceptance, with no wider scope. ## 6. Subsystem 5: A Wake Is a Sealed Authority Transition A queued prompt is not execution authority. It is a proposal to deliver work later, under conditions that may have changed. ### The 18-field `room-wake-job-v3` binding `binding_sha256` covers these exact queue fields: ```text schema, room, agent, from, source_event, path, prompt, task (explicit null when untasked), evidence_refs (when present), mention_ids, obligation_ids, pulse_message, deadline_ms, created_at, not_after, control_epoch, room_authority, registration_sha256 ``` The explicit `task: null` matters: adding or stripping a task binding changes the digest. `not_after` is computed from the original creation time and deadline—60 seconds for untasked delivery and 20 minutes for task-bound work—so queue wait consumes the same budget as execution. Registration and Room authority are sealed for comparison, not treated as evergreen. Prompt and pulse text pass the egress redactor before the job is stored. The queue path itself must resolve to the canonical regular, non-reparse object. `binding_sha256` is an **integrity comparison, not a digital signature or standalone authentication mechanism**. Authenticity comes from the controller-governed queue object and write protocol plus the live revalidation below. A party that rewrites both the job and its digest still does not inherit the sealed recipient, registration, Room authority, task, project, or control epoch; those must independently match now. This is not a claim to contain a privileged host administrator outside the governed boundary. ### Five live dispatch checks `authority::validate` re-derives five independent authority centers before launch: 1. **Sealed integrity and replay eligibility.** Schema, binding digest, canonical queue object, and `retry_allowed` must hold. Tampering records `wake_job_binding_changed` as an integrity signal. 2. **Current Room and recipient authority.** The Room must be open and uncancelled; recipient and sender rosters, owner grant, reset/session state, graph-reader admission, and registration digest must match the sealed values. Replacement authority produces `wake_room_authority_changed` or `wake_recipient_registration_changed`. 3. **Active task and governing-source authority.** The task must be open or claimed, have valid ancestry, remain owned by the sealed sender or an eligible operator/delegator, and pass `pin_task_source`. Project Rooms require a task from the exact project; project tasks cannot hide in coordination Rooms. 4. **Current admission and control epoch.** Project closure, control admission, break-glass subject, and sealed `control_epoch` must still permit a new effect. 5. **Original deadline and running-process binding.** `not_after` must still have budget. Once claimed, `RequestPermit` binds the canonical path, digest, `running` status, and worker PID; later effects revalidate that exact claim. ### Single flight, capacity, and split-lock claim Dispatch enforces one active worker per recipient with `recipient_busy` and defers excess model-route load through `worker_route_capacity`, preserving the obligation instead of dropping it. The launcher uses the verified installed engine, strips caller credentials and tier overrides, and isolates standard handles. The worker claim uses a split-lock protocol: ```text validated_context outside with_transition → file, ancestry, task, and source checks without starving sender lease locks → enter with_transition → re-read job and require unchanged binding_sha256 → require remaining deadline → mark running with worker_pid and attempts + 1 ``` A change between those phases becomes `wake_job_changed_during_validation`, not a replay. ### Evidence is delivered to the recipient, not inherited from the sender `attach_evidence` rechecks every `ev_` handle for the specific recipient at wake time. Admitted passages ride inline in the wake prompt, while served and withheld dispositions are recorded. Each delivery receives a fresh `squad.evidence.receipt/1` with `tool: "wake_delivery"` and an `srv_<32hex>` identifier. For governing-source tasks, `require_source_reading_between` proves that the native actor/session's receipts cover the continuous `[0, length)` interval across every primary story and detached element of one complete current parser view, inside the `[not_before, not_after]` connection/checkpoint window. Registration, summaries, or a few impressive chunks do not satisfy `complete_text_v1`. ### Exhibit 1 — A sealed wake binding excerpt Synthetic values show the current binding shape. The digest covers all 18 fields, not only this display excerpt. ```json { "schema": "room-wake-job-v3", "room": "architecture-demo", "agent": "reviewer-beta", "from": "coordinator-alpha", "source_event": "evt_0123456789abcdef_a1b2c3d4", "path": "room_wake_jobs/architecture-demo/reviewer-beta/evt_0123456789abcdef_a1b2c3d4.json", "prompt": "Review the sealed renewal evidence and answer the assigned obligation.", "task": "TSK-2026-ROOM-004", "evidence_refs": ["ev_8f31c09a4b2e11d08a7c65e41902b3a1"], "mention_ids": ["mention_synthetic_01"], "obligation_ids": ["obligation_synthetic_01"], "pulse_message": "Review the sealed renewal evidence.", "deadline_ms": 1200000, "created_at": "2026-10-31T16:00:00Z", "not_after": "2026-10-31T16:20:00Z", "control_epoch": 7, "room_authority": { "owner_grant": "synthetic-owner-grant", "revision": 12, "reset_at": null, "session": "room-session-synthetic" }, "registration_sha256": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567", "binding_sha256": "e01928374655647382910abcdef1234567890abcdef1234567890abcdef12345" } ``` ### Exhibit 2 — Registration changed after sealing ```json "wake_recipient_registration_changed" ``` The old prompt may still be intact. Its recipient authority is not. ### Exhibit 3 — Deadline consumed before dispatch ```json "wake_job_expired: queue time counts against the original deadline" ``` A prompt that became stale in the queue cannot acquire fresh execution time merely because a worker became available. ## 7. Subsystem 6: Native Session Delivery and Reply Are Different Proofs A tool transport can deliver a request to a process. It cannot prove that the intended native conversation received it or that the resulting obligation was answered. ### Session Factory v1 `session_factory_v1` gives each native session a global ID of the form `sf__`, with lane, native session ID, working directory, transport, model profile, proof cursor, parent, lease, tags, and per-adapter operation coverage. The current contract covers 15 IDE/CLI adapter families and eleven operations: `list`, `create`, `bind`, `resume`, `focus`, `send`, `prove`, `summarize`, `fork`, `archive`, and `restore`. For stateful IDE wakes, `spawn_policy: "never"` requires delivery into the bound native session rather than launching an untracked sibling. Adapter capability remains explicit; a shared session schema does not invent identical native features everywhere. Each adapter's `operation_coverage` determines which lifecycle operations and proof states it can actually supply. `dispatched` may be available where body-visible or correlated-reply evidence is not; an unsupported proof is reported as unsupported, never inferred from transport success. ### The five delivery/reply proof states | State | What is proven | Obligation status | |---|---|---| | `dispatched` | A delivery attempt exists, not body delivery | `delivery_unproven`; unresolved | | `delivered_to_body` / `ide_delivered` | The prompt reached the native session body | `pending_reply`; unresolved | | `body_visible_pending_reply` | The target thread UI or store shows the prompt | `pending_reply`; unresolved | | `wrong_body_rejected` | A nonmatching open thread was refused | Unresolved; no cross-thread credit | | `replied` | A correlated native reply exists | Resolved mentions and obligations | These are not fungible green statuses. Body delivery does not prove a reply. A reply does not prove destination acceptance. A wrong-body result never counts as delivery. For ACP/CLI execution, `exited_ok` without a correlated Room reply produces `proof_kind: "process_exit_only"` and leaves the job in `pending_reply`. Exit code zero is not an answered obligation. `WakeObligationFold` promotes obligations that remain unproven beyond the declared age threshold to `dead_obligation`, and cycle detection walks up to 20 links to surface circular waits. Unfinished work remains visible instead of disappearing behind a process status. ## 8. Subsystem 7: Active Cancellation and Worker Authority Room control is authenticated and physically enforced. It does not rely on a worker noticing a cancellation message eventually. ### Authenticated control and held in-flight operations Lifecycle actions require `control_actor` standing through an authenticated operator or local owner console. A caller-supplied `by = operator` string does not create control authority. `run_cancel` stops the Room compiler, dead-letters queued/deferred/unclaimed jobs with `retry_allowed: false`, leaves addressed obligations unresolved, and advances Room authority so running wake lineage no longer matches. It then inventories unresolved file operations, destination operations, and brokered model requests tied to the Room or project as `held_until_reconciled: true`. Work a destination may already have accepted is not rewritten into “cancelled and safe.” It remains held for governed reconciliation. ### The 1 Hz owned-Job observer and durable stop outbox Every confined worker belongs to a named Win32 Job Object. The `squad-owned-job-control` observer runs at 1 Hz, watches admission, break-glass, Room authority revision, cancellation, and worker authority, and calls `TerminateJobObject(handle, 1)` when authority trips. A successful terminate call is not proof of quiescence. The observer queries `JobObjectBasicAccountingInformation` and confirms termination only when `ActiveProcesses == 0`. The `squad.room-owned-stop.v1` outbox preserves that sequence before any ledger or graph commit: ```text phase-1.json termination_confirmed: false pending: bounded reason phase-2.json termination_confirmed: true pending: null predecessor: exact phase-1 event_id + envelope_sha256 ``` Replay commits those captured envelopes in causal order and writes separate `phase-1.committed.json` and `phase-2.committed.json` markers. A crash cannot manufacture a Phase 2 commit without its Phase 1 predecessor, and a stop request cannot impersonate a zero-process observation. ### Worker staging minimizes data before confinement Staging admits 1–256 exact source selections with SHA-256 verification and a 64 MiB aggregate ceiling. JSON, JSONL, and SQLite sources can be projected to explicit rows and columns. SQLite projection copies the pinned bytes to a temporary file first; the live database is never opened by the projection builder. The projected copy requires the row key and integer version, records withheld rows and fields, and prepares changes only as governed proposals. When changes are requested, the controller prepares the `RequestedAct`. The worker does not choose its own mutation. The bridge binary is staged by content digest as `squad-worker-bridge.exe` and read back through a pinned handle. Reparse points and hardlinked files are refused. Empty-work-root ACL normalization emits a receipt bound to the pinned root. `internetClient` is refused as nonconforming; model access travels only through owner-approved brokered routes. ### The controller-side MCP broker revalidates the worker and the call For LPAC workers, an `AF_UNIX` endpoint with a per-launch ACL admits the package. The controller verifies the held worker's Job membership with `IsProcessInJob`, its token and capabilities with `verify_worker_token`, and a 32-byte launch nonce. It does not pretend that an unavailable AF_UNIX peer-PID ioctl supplied a kernel identity for the bridge peer. For brokered-host Named Pipes, first-instance creation and `PIPE_REJECT_REMOTE_CLIENTS` prevent squatting and remote clients. The broker additionally uses `QueryFullProcessImageNameW` to require the exact staged bridge image. Every request and response is bounded and rechecked. Revocation before response transport produces a typed refusal; revocation after part of the response escaped records `severed_mid_response` rather than appending a second contradictory frame. The check and following write are not falsely described as one atomic primitive. ## 9. Subsystem 8: Task Lineage and Privileged Acceptance A Room reply says the worker answered. Acceptance asks whether the actual artifacts satisfy the task. ### Delegation cannot outlive or widen its ancestry Task scopes exclude controller state, recovery storage, shipped binaries, skill mirrors, and authoritative Room ledgers. `validate_ancestry` walks at most 16 parent links and rejects cycles. Every ancestor must remain open or claimed, preserve owner and recipient eligibility, match authority epoch and any revision, retain the same governing source, and keep the child's scope within the parent. `coordination_only` cannot widen into mutation. A parent whose authority changes invalidates the descendant chain. ### Seven acceptance kinds, evaluated by the engine `done_task` evaluates the task's declared checks rather than trusting completion prose. | Kind | What the privileged evaluator actually measures | |---|---| | `sha256` | Reads the in-scope file and compares its live SHA-256 with the expected digest | | `file_exists` | Uses `Path::try_exists()`; an OS or permission error is `unconfirmed`, not success | | `file_absent` | Uses the same observation law: an error fails closed instead of becoming absence | | `command` | Executes an absolute, owner-allowlisted binary with no shell or PATH lookup, optional exact argument tail, bounded timeout, and expected exit code—outside `with_transition` | | `evidence_origins` | Requires 1–16 distinct, lineage-collapsed first-hand origins behind the named requirement, plus any owner-wide minimum | | `destination` | Checks folded ledger status and a fresh read-only `effect_present` observation for one operation or every task operation | | `worker_receipt` | Validates Basic, StrictReturn, or SuccessorReturn evidence from immutable submissions, not a worker-authored receipt path | The three receipt profiles are intentionally different: - **Basic** requires at least `min` task submissions with one terminal outcome and no pending, recording-pending, or worker-unread results. - **StrictReturn** requires one applied submission with the exact worker, target, expected fields, and version delta `+1`, then proves that the original worker received the result and returned a correlated Room reply. `proposal_id` is optional; `outcome` and `min` are forbidden. - **SuccessorReturn** requires all eight pins: `proposals: "successor"`, worker, proposal ID and digest, whole-record digest, whole-outcome digest, and outcome. It proves fresh acknowledgment of a historical result while explicitly keeping `satisfies_original_return: false` and `resolves_original_obligation: false`. A successor acknowledgment never fabricates the original session's return. ### Completion revalidates authority and live evidence Long-running command checks run before the short transition. Inside `with_transition`, completion then: 1. re-reads the task and requires it to remain claimed; 2. compares all 14 authority keys: `owner`, `scope`, `active_scope_pin`, `file_scope_pin`, `project_id`, `project_root_pin`, `authority_epoch`, `authority_revision`, `source`, `delegation`, `lock`, `acceptance`, `requirements`, and `evidence_manifest`; 3. rechecks dependencies, ancestry, project scope, and settled worker contributions; 4. re-evaluates every non-command check against current files, evidence, and destinations; 5. refuses completion on changed acceptance or `content_mismatch` quarantine; 6. releases the scope lock through the structured release path and appends `task_done`. Requirement- or source-bound tasks cannot close with an empty acceptance array unless a trusted operator records an explicit override and reason. Acceptance is configured evidence, not a ceremonial green check. ### Exhibit 4 — Successful artifact acceptance This synthetic result combines file, evidence, destination, and receipt measurements. Every predicate is independently true; none borrows the others' assurance. ```json { "task": "TSK-2026-ROOM-004", "acceptance_pass": true, "acceptance_results": [ { "kind": "sha256", "pass": true, "detail": { "path": "project:/artifacts/renewal-result.json", "expected": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567", "actual": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567" } }, { "kind": "evidence_origins", "pass": true, "detail": { "req": "R1", "min": 2, "distinct_origins": 2, "basis": "task_acceptance" } }, { "kind": "destination", "pass": true, "detail": { "operation_id": "op_synthetic_004", "status": "resolved", "effect_present": true } }, { "kind": "worker_receipt", "pass": true, "detail": { "profile": "strict", "proposals": "task", "worker": "reviewer-beta", "proof": "original-session-return-verified" } } ] } ``` ### Exhibit 5 — `file_absent` cannot turn an observation error into absence ```json { "kind": "file_absent", "pass": false, "detail": { "path": "project:/protected/unreadable-result.json", "error": "Access is denied.", "error_kind": "PermissionDenied", "observation": "unconfirmed" } } ``` `!path.exists()` would be the wrong proof here. `try_exists()` preserves the difference between absent and unknowable. ## 10. Subsystem 9: Five Runtime Invariants Constrain Event Law Pre-action gates decide whether an act may proceed. The runtime invariant engine continuously checks the event stream for structural violations and records dedicated `invariant_violation` nodes through the journal and Room graph. | Invariant | State carrier and rule | Severity and boundary | |---|---|---| | `exactly_once` | A bounded 10,000-entry FIFO `IdempotencyRegistry` rejects a repeated confirmed idempotency key | `Reject`; deduplicates registered events, not every external effect | | `well_formed` | A lifecycle tracker per `(agent_id, parent_id)` enforces valid `Begin → Effect → Resolve` transitions | `Reject`; an effect or resolution cannot invent its own lifecycle | | `exclusive` | A lease set over `TypedScope(namespace, resource_id)` rejects conflicting live holders | `Reject`; file scopes conflict hierarchically, non-file scopes by exact identity | | `causal` | A resolved-ID set checks whether each `parent_id` is already resolved | `Flag`; unresolved causal matter is surfaced, not silently accepted as complete | | `resolution` | An unresolved-intention map detects begins that exceed the caller's TTL before resolution | `IndoubtEffect`; the effect remains uncertain rather than being declared success or failure | Reject violations are recorded without committing the triggering envelope. Flag violations are buffered in the same checked append path. The graph therefore preserves both rejected operations and suspicious causal conditions without laundering them into valid effects. The 10,000-entry FIFO also defines a **bounded replay-detection window**. An evicted key no longer receives the registry's duplicate protection; eviction does not create a fresh execution grant or a permanent replay-proof history. Deployment acceptance must state the registry's retention and reconstruction horizon rather than treating it as unlimited. **Exactly-once event resolution is not exactly-once external effect.** A destination can accept an effect while its ordinary receipt fails, an external service can perform a non-idempotent side effect, and a process can crash between mutation and observation. Articles 06–08 carry that burden through recovery records, readback, held completion, and break-glass reconciliation. ### Exhibit 6 — An in-doubt TTL record ```json { "event_id": "evt_0123456789abcdef_deadbeef", "agent_id": "reviewer-beta", "began_at": "", "elapsed_secs": 317 } ``` This is not a failure receipt and not a success receipt. It is a bounded indication that an intention remained unresolved beyond the declared TTL and requires reconciliation. ## 11. The 2026 Coordination Evidence Landscape The 2026 literature and protocol specifications reinforce the same architectural distinction: state management, native session identity, revalidation, and measured completion are separate control surfaces. Each source has a bounded claim. | 2026 primary source | What it actually establishes | How the Room composes with it | |---|---|---| | Liu et al., [*Multi-agent Collaboration with State Management (STORM)*](https://arxiv.org/abs/2605.20563), May 2026 preprint | Worktree isolation defers conflicts to post-hoc merge; STORM mediates shared-workspace state at write time and outperforms the baseline by 18.7 points on Commit0-Lite and 1.4 on PaperBench, with comparable or better cost efficiency. | Strong evidence for live state mediation. Git remains the source-history plane; the Room handles current observation, admission, and ownership. | | Sarkar, [*Before the Pull Request: Mining Multi-Agent Coordination*](https://arxiv.org/abs/2606.19616), June 2026 preprint | A shared coordination substrate reduced redundant work from 78% to 0% while useful throughput more than tripled, converged without silently dropped writes, and exposed lock starvation, rediscovery, and race-to-close hidden from PR history. | Coordination artifacts must preserve claims and collisions before the PR. The Room additionally gates their authority and acceptance rather than merely mining them afterward. | | Nechepurenko and Shuvalov, [*Coordination as an Architectural Layer for LLM-Based Multi-Agent Systems*](https://arxiv.org/abs/2605.03310), May 2026 preprint | An information-controlled prediction-market study reports distinguishable coordination signatures and a cost-quality Pareto frontier. The authors call it a methodology-validating first instantiation, not a general cross-model claim; pairwise tests do not survive correction. | Coordination belongs in architecture, not only prompts. The Room does not borrow a universal failure-rate claim from this bounded study. | | [Model Context Protocol `2026-07-28`](https://modelcontextprotocol.io/specification/2026-07-28) | Defines stateless, self-contained requests and per-request capability negotiation, with optional tasks and context features. | Stateless transport does not supply native conversational standing. The Room binds actor, registration, native conversation, and current view explicitly. | | [Agent Client Protocol documentation](https://agentclientprotocol.com), 2026 | Standardizes editor–agent communication and native session lifecycle across local JSON-RPC/stdio and remote transports. | Native sessions are the delivery boundary. The Room's Session Factory and proof ladder distinguish dispatch, body visibility, wrong body, reply, and process exit. | | Wang, Wang, and Wu, [*The Unreliable Progress Bar*](https://arxiv.org/abs/2609.08589), September 2026 preprint | Reporting reliability varies by task stage: most tested models lose accuracy mid-task and recover at completion, while newer models become conservative at the finish line. The authors advise against controlling flow from state reports alone. | `done` requires privileged artifact and destination acceptance, not a model's progress bar. | | Lyu et al., [*From Version Conflicts to Decision Conflicts*](https://arxiv.org/abs/2609.08015), September 2026 preprint | ATR rechecks only decision-affecting conditions, matched every specified outcome across 210,000 controlled executions, evaluated 0.6 versus 6.0 conditions per change, and reports a large latency reduction. The authors limit this to controlled feasibility. | The Room rechecks current authority at turn entry, dispatch, worker claim, completion, and destination—not merely whether bytes changed. | | Khan, [*S-Bus: Automatic Read-Set Reconstruction for Multi-Agent LLM State Coordination*](https://arxiv.org/abs/2605.17076), May 2026 preprint, revised v2 | Observable-Read Isolation reconstructs read sets from observed GET traffic, with formal checks and zero Type-I corruptions across 884,110 commit attempts. It also documents topology-dependent limits and agent over-reporting. | The Room measures confirmed read coverage and causal ancestry rather than assuming an agent observed the state it cites. | | [SQLite WAL documentation](https://www.sqlite.org/wal.html) and [Microsoft Win32 documentation](https://learn.microsoft.com/en-us/windows/win32/api/), current 2026 references | WAL durability modes govern when transactions reach stable storage; `GetFileInformationByHandleEx`, Job Object termination, and accounting provide kernel object identity and process-count observations. | These are implementation primitives, not complete coordination semantics. The Room composes them into object pinning, journal reconciliation, and durable stop proof. | A final concession matters: a conventional deployment can implement the same invariants with different names. The Room's claim is not exclusive ownership of coordination theory. Its claim is a concrete, inspectable composition of the required predicates in one continuous machine. **Evidence availability and performance boundary:** the implementation and acceptance artifacts are `CONTROLLED` for qualified, method-scoped technical review, with adverse findings permitted. They are not unrestricted public downloads or independent validation. The architecture does not claim a universal latency or throughput envelope: recursive causal queries, journal folds, live revalidation, recipient evidence checks, native-session proof, and stop confirmation must be measured against the actual Room topology, adapters, and workload. ## Bring One Assignment With an Uncertain Outcome The **Multi-Agent Coordination & Acceptance Audit Challenge** evaluates one assignment whose delivery, ownership, or completion is currently hard to prove. The demonstration maps causal readiness, replays sealed-wake acceptance and refusal, and measures the resulting artifact and destination rather than trusting completion prose. The challenge excludes universal adapter or native-session attestation, production write authority, and exactly-once external effects. Deployment and paid pilots are separately scoped; independent evaluation remains outside the commercial ladder through [Article 09](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/). ## The Twelve Coordination Claims | ID | Bounded claim | Evidence and ceiling | |---|---|---| | `04-C01` | Split-brain state trees are refused, and project grants revalidate canonical path plus 192-bit combined kernel object identity | Implemented and author-verified on local NTFS/ReFS territory; not a remote-share or host-admin guarantee | | `04-C02` | Enrollment, thread-local sessions, native conversation binding, tier clamping, and multiplexer verb probing bind every connector call to a principal and route | Implemented and author-verified on declared MCP/CLI paths and host contracts | | `04-C03` | Room shards carry immutable scope bindings and parent-withholding closure | Implemented and author-verified for graph queries, compositions, and exports; hash artifacts do not prove non-deletion | | `04-C04` | The sync index reconciles handle-pinned journals across hash/HLC chains and binds checkpoints to conversation and view | Implemented and author-verified for rotation, fork, and delivery-policy refusal paths | | `04-C05` | The Work Gate requires live frontier and recursive causal-parent confirmation before all nine participant verbs | Implemented and author-verified; blind writes receive a sync page, not authority | | `04-C06` | Contribution verbs separate status, response, assignment, ownership transfer, and completion | Implemented and author-verified for ambiguity, cross-Room evidence, successor, and no-wake paths | | `04-C07` | Every wake is an 18-field permit revalidated at dispatch and worker claim | Implemented and author-verified for binding, roster, registration, task, epoch, path, PID, and original-deadline changes | | `04-C08` | Wake-time evidence is recipient-specific and governing-source coverage is continuous over `[0, length)` | Implemented and author-verified through `wake_delivery` receipts and bounded time windows | | `04-C09` | Native delivery and reply are distinct proof states; process exit alone never resolves an obligation | Implemented and author-verified for declared adapters; not a universal UI or remote-session attestation | | `04-C10` | Cancellation holds in-flight work and confirms owned-Job termination only at zero active processes | Implemented and author-verified through the two-phase durable stop outbox; accepted destination effects remain for reconciliation | | `04-C11` | Delegation and completion use 16-level ancestry, 7 acceptance kinds, and 14-key final authority revalidation | Implemented and author-verified; acceptance is bounded by configured predicates and protected observation paths | | `04-C12` | Five runtime invariants constrain event law and emit dedicated violation evidence | Implemented and author-verified; exactly-once event handling does not prove exactly-once external effect | ## The Article 04 Claim Card ```text Mechanism: The AINA Room Coordination Plane (state identity, project grants, connection/session authority, tool policy, graph shards, visibility, sync, contribution verbs, wake permits, Session Factory, worker confinement, cancellation, acceptance, invariants) Failure it governs: Source history and chat impersonating live authority; split-brain state and redirected territories; stale context, causal-parent blindness, and ambiguous obligations; stale queued wakes and wrong-session delivery; process exits and agent prose impersonating completion; unobserved cancellation and unverified task closure. Positive path: EISA-recognized act → current principal, native conversation, project, Room scope → confirmed causal frontier → task-bound contribution and 18-field sealed wake → dispatch and split-lock worker-claim revalidation → confined execution and native reply proof → 7-kind artifact/destination acceptance → terminal evidence or explicit in-doubt reconciliation. Refusal / hold path: project_root_redirected, actor_mismatch, graph_project_rebind_refused room_sync_required, room_sync_receipt_forked say_reply_target_ambiguous, cross_room_evidence_refused wake_room_authority_changed, wake_recipient_registration_changed wake_job_changed_during_validation, wake_job_expired wrong_body_rejected, process_exit_only held_until_reconciled, task_acceptance_changed_before_completion file_absent observation unconfirmed, IndoubtEffect. Evidence status: Implemented and Author-Verified across the declared Windows reference embodiment and native adapter contracts. External 2026 research is primary-source context, not endorsement. Not Independently Validated and not a universal external exactly-once claim. Verifier and artifacts: state_dir_id.json, ProjectGrant.root_object_id, secret_sha256, SessionGuard, EffectiveTier and VERB_PROBE, binding.json, content-addressed blobs, sync_index.db, conversation_key/view_gen checkpoints, causal CTE results, room-wake-job-v3/binding_sha256/RequestPermit, srv_ wake_delivery receipts, sf_ native sessions, room-stop-reports phase-1/phase-2 and committed markers, acceptance_results, task_done, invariant_violation nodes. Scope and claim ceiling: Kernel object identity and process containment are Windows reference mechanisms. Native host metadata is trusted only within each adapter's declared contract. Composition is read authority, never execution authority. Delivery is not reply; reply is not acceptance; acceptance is configured evidence. Exactly-once event resolution is not exactly-once external effect. No privileged host-admin bypass or universal adapter claim is made. Next stage in the machine: Article 04 establishes live coordination and verified assignment. Article 05 enforces typed leases, monotonic generations, and kernel write exclusion at the actual mutation window. Expiry alone is not fencing. Commercial doorway: Multi-Agent Coordination & Acceptance Audit Challenge — bring one assignment with uncertain delivery, ownership, or completion; receive a causal-readiness map, sealed-wake validation replay, and artifact-based acceptance demonstration (Article 09 · me@blayneaina.com). ``` ## The Room Is Where Authority Becomes Current A recognized act can be mathematically sound and still be operationally stale. The evidence may be valid, the partition correct, the destination plausible—and the worker may be reading an old frontier, holding a replaced registration, answering a superseded task, or exiting before it ever replied. The Room closes that gap by making every stage re-establish its own predicates. Territory is re-queried. Credentials are re-read. Conversation checkpoints are recomputed. Causal parents are walked. Wake permits are revalidated at launch and claim. Evidence is served to the recipient. Native replies are correlated. Cancellation is observed at the kernel. Acceptance is measured against live artifacts. Invariant violations remain visible. That is why the Room is a coordination plane rather than a transcript. > **Git tells you what history recorded. Chat tells you what was said. The Room tells you what may happen now—and what still has not been proven.** Article 05 narrows the final coordination question into the mutation window itself: who holds the typed lease and current generation, and how does the operating system prevent a stale writer from crossing the write syscall? ## Frequently Asked Questions ### Why aren't Git worktrees enough for multi-agent coordination? Worktrees separate source histories and defer conflict resolution to a later merge. They do not establish current resource ownership, causal observation, wake validity, native delivery, or artifact acceptance. Git remains valuable for source history; those additional properties need a live coordination plane. ### Why isn't a workflow checkpoint enough? A checkpoint can resume a plan. It does not prove that the plan remains valid for the current actor, native conversation, registration, project, task ancestry, authority revision, or control epoch. The Room revalidates those conditions instead of treating resumability as standing. ### What makes a Room wake different from a prompt queue? A wake is an 18-field permit bound to the recipient, task, evidence, Room authority, registration, control epoch, and original deadline. It is revalidated at dispatch and again inside the worker's split-lock claim. A queued prompt by itself carries none of that proof. ### Is delivery to an IDE session proof of completion? No. Dispatch, body delivery, body visibility, wrong-body refusal, and reply are separate states. Process exit without a correlated reply is `process_exit_only` and leaves the obligation unresolved. Completion additionally requires privileged acceptance over actual artifacts and destination state. ### Why does `file_absent` use `try_exists()`? Because an unreadable or inaccessible path is not proven absent. `try_exists()` returns an observation error, and the acceptance check fails closed with `observation: "unconfirmed"` rather than converting missing permission into a passing absence claim. ### Do the five invariants prove exactly-once external effects? No. They constrain idempotency, lifecycle, exclusivity, causality, and resolution in the event system. A destination can still mutate before a receipt fails or crash occurs. That uncertainty must be observed and reconciled through the destination and recovery paths. ## Continue the Series - **Series index:** [From Proposal to Consequence: The AINA Transition-Governance Series](https://blayneaina.com/) - **Previous:** [Article 03 — Semantic Switch and EISA: Protected Computation and Entity Resolution](https://blayneaina.com/03-eisa-four-gate-verdict-machine-and-entity-resolution/) - **Next:** [Article 05 — Expiry Is Not Fencing: Typed Leases and Kernel Write Exclusion](https://blayneaina.com/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/) - **Deployment and independent review:** [Article 09 — Bring One Real Boundary](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/) --- ## Primary 2026 Sources - [Liu et al. — *Multi-agent Collaboration with State Management (STORM)*, `arXiv:2605.20563`](https://arxiv.org/abs/2605.20563) - [Sarkar — *Before the Pull Request: Mining Multi-Agent Coordination*, `arXiv:2606.19616`](https://arxiv.org/abs/2606.19616) - [Nechepurenko and Shuvalov — *Coordination as an Architectural Layer for LLM-Based Multi-Agent Systems*, `arXiv:2605.03310`](https://arxiv.org/abs/2605.03310) - [Model Context Protocol Specification `2026-07-28`](https://modelcontextprotocol.io/specification/2026-07-28) - [Agent Client Protocol Documentation](https://agentclientprotocol.com) - [Wang, Wang, and Wu — *The Unreliable Progress Bar*, `arXiv:2609.08589`](https://arxiv.org/abs/2609.08589) - [Lyu et al. — *From Version Conflicts to Decision Conflicts: Selective Revalidation for Long-Running AI Agents*, `arXiv:2609.08015`](https://arxiv.org/abs/2609.08015) - [Khan — *S-Bus: Automatic Read-Set Reconstruction for Multi-Agent LLM State Coordination*, `arXiv:2605.17076`](https://arxiv.org/abs/2605.17076) - [SQLite Write-Ahead Logging Documentation](https://www.sqlite.org/wal.html) - [Microsoft Win32 API Documentation](https://learn.microsoft.com/en-us/windows/win32/api/) --- ## Publication Metadata **Title tag:** The Room: Why Git and Chat Are Not Coordination Planes **Meta description:** Why Git worktrees and chat transcripts cannot govern live agents: inside The Room's territory pins, causal sync gate, sealed wake permits, native-session proof, cancellation, and privileged acceptance. **Series navigation:** [Series index](https://blayneaina.com/) · Article 04 of *From Proposal to Consequence: The AINA Transition-Governance Series*. **Previous installment:** [Article 03 — Semantic Switch and EISA: Protected Computation and Entity Resolution](https://blayneaina.com/03-eisa-four-gate-verdict-machine-and-entity-resolution/). **Next installment:** [Article 05 — Expiry Is Not Fencing: Typed Leases and Kernel Write Exclusion](https://blayneaina.com/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/). ================================================================================ # ARTICLE §05: Expiry Is Not Fencing: Typed Leases and Kernel Write Exclusion Canonical URL: https://blayneaina.com/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/ # Expiry Is Not Fencing: Typed Leases and Kernel Write Exclusion *A lease clock can expire without removing a stale worker's ability to write. Authority must be revalidated at the effect boundary and held through the rename.* **By Blayne Aina · Principal Systems Architect and creator of AINA · October 31, 2026** --- Worker A's lease expires. Worker B takes over and commits. Worker A resumes with a valid-looking local state and tries to write. The clock cannot stop it. Expiry is a liveness mechanism: it allows ownership to move after a holder stops responding. It is not a physical fence. A paused process, delayed packet, cached credential, or lingering thread can still attempt the write. Only current authority at the effect boundary can refuse it. The local version is just as important. A background script does not consult your JSON lockfile. It opens the target with `GENERIC_WRITE` and changes the bytes while your coordination plane is busy being correct. Article 04 established who may receive and perform work. Article 05 enforces ownership at the actual mutation window: typed scopes, monotonic generations, conditional lifecycle, kernel-backed handle exclusion, guarded replacement, and explicit reconciliation. > **A lease is not a timer with a filename. It is a current tenancy whose generation must match at execution and whose kernel exclusion must survive every handle handoff.** ## The Two Hazards the Architecture Must Defeat ### Hazard 1: The resurrected worker ```text Worker A acquires generation 14 and stalls → TTL expires; Worker B acquires generation 15 and commits → Worker A resumes and presents generation 14 → executor rejects stale_lease_generation ``` The refusal must happen at the resource. A wake-time check is not enough because authority can change after the wake. A heartbeat alone is not enough because a delayed heartbeat can arrive after succession. The worker's generation and the resource's high-water mark must be compared at commit. ### Hazard 2: The uncooperative local writer A lease record under `.squad/locks` is advisory to cooperating software. It has no power over a process that never reads it. On Windows, the enforcement mechanism is a kernel handle whose sharing mode denies later incompatible opens. A long-lived server holding `FILE_SHARE_READ` causes later write, delete, or rename attempts to fail with `ERROR_SHARING_VIOLATION` (32). That is an operating-system boundary, not a prompt convention. The two hazards require different but composed responses: ```text stale logical authority → monotonic generation + tenancy revalidation uncooperative file opens → Win32 share-mode exclusion replacement races → overlapping guard handoff + identity/content/path verification divergent bytes after release → Amber quarantine, never silent baseline laundering ``` ## 1. Six Typed Namespaces and Real Target Identity Leases are partitioned by resource class, not by an undifferentiated path string. | Namespace | Example resource | Conflict rule | |---|---|---| | `file://` | Exact files or directory hierarchies | Canonical path-prefix conflicts; parent and child scopes overlap | | `session://` | Managed sessions and virtual resources such as `sqlite.records/` | Canonical session-resource prefix conflicts | | `window://` | GUI application windows | Hierarchical prefix conflicts | | `account://` | Shared service accounts or external identities | Hierarchical prefix conflicts | | `focus://` | Global user focus and interaction ownership | Singleton: every pair of distinct focus scopes conflicts | | `browser://` | Browser automation contexts and page targets | Hierarchical prefix conflicts | Bare paths default to `file://` for compatibility. A `TypedScope` contains the namespace and full `resource_id`; different namespaces never conflict with one another **at the lease-algebra level**. That does not prove two names cannot designate the same physical resource. A `session://sqlite.records/` lease and a `file://` lease over the underlying database file are different authority domains: the session route is not a Win32 guarded-file replacement, and the file route is not a SQLite row-version fence. A deployment exposing both routes must bind their path and transition authority explicitly rather than assuming namespace separation has made them mutually exclusive. `with_transition` serializes governed controller mutations; it does not invent a global fence around every out-of-band alias. FNV-1a 64-bit derives a deterministic routing key from `prefix() + resource_id`, used for filenames such as `locks/write/.json`. It is **not** a cryptographic capability, a fencing token, or a collision-resistant authority proof. Security decisions compare the full canonical namespace and resource identity, not the hash alone. ### Lexical paths are not kernel object identity `file_scope` rejects empty paths, wildcards, parent traversal, NTFS alternate data stream colons, and trailing dots or spaces. For a not-yet-created target, it finds the nearest existing ancestor, canonicalizes that object, and reattaches the suffix. Windows paths normalize to lowercase forward-slash form. `opened_file_scope` goes further: it calls `GetFinalPathNameByHandleW` on the already-open handle and resolves the actual file object. This is how the system distinguishes “the path I intended” from “the object I actually opened,” piercing junctions, symlinks, mount points, and other reparse indirection. The non-equivalence is fundamental: ```text path string ≠ opened kernel object identity FNV-1a routing key ≠ authority proof lease record ≠ current tenancy ``` ## 2. Workspace-Wide Monotonic Generations A per-file counter can restart when the file is replaced or a lease is released. That makes old authority look current again. AINA therefore uses a **workspace-wide high-water sequence** in `room.generation.v2`. ### Two-phase persistence Each new generation is the maximum counter across the store plus one, computed with `checked_add(1)` and an overflow refusal (`generation_exhausted`). Persistence has two ordered steps: ```text Phase 1: write .generations-initialized protocol: room.generation.v2 map_sha256: SHA-256 of the complete prepared counter map high_water: the newly issued generation prepared_counters: the complete prepared map Phase 2: atomically publish generations.json ``` A missing store after initialization is `generation_store_missing`. A marker digest or high-water mismatch is `generation_checkpoint_mismatch`. Unknown protocols, malformed counters, and persistence failure (`generation_persist_failed`) refuse acquisition. Missing, corrupt, or rolled-back state never silently becomes generation zero. The marker detects partial writes, divergence, and one-sided or inconsistent rollback. It does **not** prove monotonicity against a privileged actor who restores both the marker and `generations.json` to the same earlier, internally consistent snapshot. Defeating that complete-snapshot rollback requires a monotonic anchor outside the rollback domain, such as separately protected authority or hardware-backed state; no two-file local protocol can supply it. ### Exhibit 1 — A sealed generation marker ```json { "protocol": "room.generation.v2", "map_sha256": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567", "high_water": 15, "prepared_counters": { "0123456789abcdef": 14, "fedcba9876543210": 15 } } ``` The marker binds the intended complete map before the store is published. A rollback to an earlier valid JSON map is still a refusal, not a fresh start. ### Exhibit 2 — The stale-worker refusal ```json "stale_lease_generation: expected 14, current 15" ``` A later operation cannot borrow the fact that generation 14 used to be correct. ## 3. A 30-Second Cross-Process Transition Mutex Generation, lease mutation, vault capture, guarded replacement, and readback must form one authoritative transition. `with_transition` serializes them through `.lease-transition.mutex`, an OS file lock opened with read/write sharing but no delete sharing. The guard supports re-entrant acquisition and polls in 10 ms intervals for up to **30 seconds**. A still-held transition returns `lease_transition_busy`; it is never stolen because a timer elapsed. The lock is released by handle drop. This is not the same thing as the 15-minute lease TTL: - The **transition mutex** serializes one commit and is held across decision, capture, replacement, and verification. - The **lease TTL** bounds tenancy and requires heartbeat renewal. - The **generation** determines whether a tenancy is current at execution. - The **control epoch** can revoke an entire lineage immediately, independent of remaining TTL. ## 4. Acquisition, Re-Grant, and Conditional Lifecycle Acquisition checks break-glass admission, package/source exclusion, namespace conflicts, and existing tenancy. A lease cannot silently change tasks; `lease_task_mismatch` preserves the original binding. Fresh acquisition captures pre-write state and binds the current control epoch. ### Dead-holder re-granting requires proof of death A disconnected holder does not make its lease instantly reusable by anyone. Re-granting is limited to the **same owner and same task**, and requires: 1. an enforced server-held tenancy, or an eligible identity-bound managed SQLite session; 2. `holder_alive` to be false; 3. the recorded `holder_pid` to be absent according to `process_alive`. A missing server record alone is not proof of death, and a reused or uncertain PID fails closed. Once those conditions hold, re-granting issues a **new generation**, fencing delayed writes and proposals from the old process or lingering children. ### Conditional lifecycle compares all six tenancy keys Heartbeat and release compare: ```text namespace + resource_id + owner + generation + started_at + task_id ``` If any key differs, the action fails with `stale_lease_generation`. A delayed release cannot drop a successor's lease, and a heartbeat cannot adopt one. Heartbeats cannot resurrect expired tenancies. Requested TTLs clamp to `MAX_LEASE_TTL_MINS = 15`, and every renewal buys at most 15 minutes from the current instant. `transfer_holder` supports process-image handoff: every target record must name the current process and session, and all updates roll back to their preimages if any write fails. The holder fields move; the generation remains stable because this is process migration, not ownership succession. ## 5. The Five-Role Overlapping Handle Exchange Windows refuses to rename over an existing destination when any open handle withholds `FILE_SHARE_DELETE`. The long-term guard therefore cannot simply remain open across replacement. Closing it naively would create a write-open race. The solution is an overlapping exchange of five named file roles. `G`, `W`, `T`, `H`, and `G'` are role names, not invented lease types, and they are not the only live handles: the parent pin `P`, authorization guards, and process records have their own jobs. ```text HANDLE ROLE & ACCESS / SHARING TIMELINE DURING replace_guarded: Role G — long-term guard READ | share READ Held across the lease lifetime; denies later write/delete opens ===========================================> Close G │ Role W — window guard │ opened while G remains open READ | share READ | DELETE │ Pins and hashes the exact preimage │ ------------------------------------------->│ Close W after rename │ Role P — parent directory pin │ TRAVERSE | READ_ATTR | SYNC │ share READ | WRITE; no DELETE sharing │ Proves the parent object identity │ ------------------------------------------->│ Close P after rename │ INTERVAL 1: G closes; W and P remain │ Write opens denied; delete/rename possible │ ▼ Role T — staged file │ READ | DELETE | share READ | WRITE_THROUGH │ Staged bytes must equal planned_sha256 │ NtSetInformationFile(FileRenameInformationEx)├─ Rename → Drop T │ Role H — handoff guard │ opened on new file while T is open READ | share READ | DELETE │ Bridges publication without a write-open gap│ ------------------------------------------->│ Close H after G' verifies │ INTERVAL 2: T closes; H remains │ Write opens denied; delete/rename possible │ ▼ Role G' — new long-term guard │ READ | share READ │ Verifies object identity, full content hash,│ resolved path, and preimage displacement │ ===========================================> Active tenancy ``` ### Role G: the long-term exclusion At acquisition, a long-lived server holds the exact file with read access and `FILE_SHARE_READ`. Later incompatible write/delete/rename opens fail. A short-lived CLI process cannot retain that handle after exit, so its lease is explicitly advisory: `short_lived_process_cannot_hold_a_handle`. Protected execution requires a server-held resource handle. ### Role W: pin and hash the preimage While `G` is still held, `W` opens with `READ | DELETE` sharing. It identifies the same object as `G`, records its link count, and hashes the source bytes through the handle. The authorization precondition receives that observation, not a second path-based read. Because `W` shares `DELETE`, it permits the upcoming rename while continuing to deny writes. ### Role P: pin the parent directory The parent opens with `FILE_FLAG_BACKUP_SEMANTICS`, traversal/read-attribute/synchronize access, and `READ | WRITE` sharing without `DELETE`. External processes cannot delete or rename the containing directory while `P` is held; ordinary child reads and writes are not globally blocked. `opened_file_scope(&P)` must equal the leased parent. A junction swap produces `destination_parent_redirected`; an unresolvable parent produces `destination_parent_unverifiable`, not a pass. ### Role T: stage and rename the new object The staged file opens with `READ | DELETE` access, `FILE_SHARE_READ`, and `FILE_FLAG_WRITE_THROUGH`. Its complete bytes must hash to `planned_sha256` before authority runs. Authorization executes while `G` still denies write. Then `G` closes, and `T` renames immediately by handle. The rename does not use a path-level convenience call. It invokes native `ntdll::NtSetInformationFile` with: - `FILE_RENAME_INFORMATION_EX` class **65**; - `FILE_RENAME_FLAG_POSIX_SEMANTICS` (`0x2`); - `FILE_RENAME_FLAG_REPLACE_IF_EXISTS` (`0x1`) when replacement is allowed; - `RootDirectory = parent_pin.as_raw_handle()`; - one verified relative UTF-16 leaf name. Negative status maps through `RtlNtStatusToDosError` and fails closed. ### Role H: bridge publication to the next guard While `T` still holds the published object, `H` opens the new destination with `READ | DELETE` sharing. When `T` closes, `H` continues denying writes and bridges to `G'`. ### Role G': verify before success `G'` opens while `H` remains, with `FILE_SHARE_READ`. Success requires all of the following: 1. `H` and `G'` identify the same object that `T` published by volume serial and file index. 2. The complete content through `G'` hashes to `planned_sha256`. 3. `opened_file_scope(&G')` resolves to the exact leased scope. 4. The surviving `W` handle has not produced positive evidence that the captured preimage moved elsewhere first. Only then does the new long-term guard become active. ## 6. Two Residual Intervals—and the Evidence They Produce The handle exchange closes write-open gaps. It does not make every delete or rename impossible. During Interval 1, only `W` and `P` remain; during Interval 2, only `H` remains. Those handles share `DELETE`, so a third party can rename or delete the destination but cannot open it for writing. These intervals are explicit architecture, not something the article should hide: - Interval 2's object substitution is detected by `G'` identity verification. - A reparse redirect is detected by comparing `G'`'s resolved path with the leased scope. - A moved preimage is detected through `W`. - A third-party delete-and-recreate can remain indistinguishable from ordinary replacement. The bytes published **at the leased path** are still verified; the architecture does not claim to recover every uncaptured third-party object. **Verification is not a power-loss atomicity proof.** If power fails after the rename but before `G'` is established, the destination may have changed while no success record exists. Restart must compare the surviving destination, `T`, `W`, and vault-backed preimages, then classify the result through the reconciliation path rather than assuming either a clean publish or an automatic rollback. Filesystem and directory-metadata durability are separate deployment claims, not something this handle sequence proves by itself. Article 06 owns that recovery boundary. ### Preimage displacement forensics `W` follows the original file object rather than its name. After the rename, `preimage_displacement` requires positive evidence of all three conditions: 1. `links_after == links_before`—the object kept every link; 2. `deleted == false`—it was not deleted; 3. `opened_file_scope(w)` resolves to an ordinary path different from the leased scope and outside NTFS `$Extend\$Deleted`. When those hold, the original preimage was moved before the rename, and the object superseded at the leased name was never captured. That is `destination_preimage_displaced`, not a clean publish. A link-count reduction, pending deletion, deleted-files path, or unreadable state is not claimed as displacement. This distinction prevents false certainty from ambiguous evidence. ### Exhibit 3 — Successful guarded replacement verification ```json { "leased_scope": "project:/artifacts/renewal-result.json", "planned_sha256": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567", "published_identity": "volume-64:file-index-128", "guard_identity_matches_published": true, "guard_content_matches_planned": true, "guard_resolved_path_matches_scope": true, "preimage_displaced_to": null, "new_long_term_guard": "G'", "verification": "complete" } ``` This is a complete verification result, not merely a successful rename return. ### Six typed reconciliation faults Verification failures do not trigger an imagined automatic rollback. They produce structured evidence and operational next actions. | Fault | What broke | `reconcile_by` action | |---|---|---| | `GuardLost` | The destination guard could not be acquired or reacquired | If the rename completed, hash the path to distinguish published bytes from preimage; if it failed, re-establish the lease before any further write | | `NameRedirected` | `G'` resolves somewhere other than the leased path | Resolve the name, locate the published bytes at the resolved target, and trust neither location by name alone | | `IdentityChanged` | The object at the leased name is not the object `T` published | Treat the current bytes as foreign and restore the verified vault preimage | | `ContentDiverged` | The guarded object does not contain `planned_sha256` | Restore the vault preimage instead of retrying blindly | | `PreimageDisplaced` | The captured preimage survived intact at another path | Verify the displaced object and the uncaptured replacement's ownership before deciding whether planned bytes remain | | `Unverifiable` | Identity, content, or path checks could not run | Establish state by hashing and comparing against planned and preimage digests; do not claim either success or a definite change | Every reconciliation record carries the fault kind, `rename_completed`, planned and preimage digests, whether the destination existed, optional `preimage_displaced_to`, and `reconcile_by`. It also records an integrity signal and annotates `os_lock` so a later operation does not inherit false enforcement. ### Exhibit 4 — `destination_preimage_displaced` ```json { "path": "project:/artifacts/renewal-result.json", "scope": "file://project:/artifacts/renewal-result.json", "error": "the captured preimage was moved before the rename, so an uncaptured object was replaced", "fault": "destination_preimage_displaced", "rename_completed": true, "planned_sha256": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567", "preimage_sha256": "58b2e91f4c03d8a7126590ef4312ab98c7654321fedcba09876543210fedcba9", "destination_existed_before": true, "preimage_displaced_to": "project:/artifacts/renewal-result.previous", "reconcile_by": "Verify the displaced preimage and the uncaptured replacement's ownership before deciding whether planned_sha256 stays at the leased path", "pid": 4104, "ts": "2026-10-31T18:00:00Z" } ``` The rename completed. The verification did not produce success. That distinction is preserved instead of being flattened into a generic lock error. ## 7. Server Liveness and Zero-Downtime Image Handoff A fencing system cannot depend on a PID alone. PID reuse would let a dead holder appear alive. `enter_server_mode` publishes `.squad/servers/.json` with build, session, and protocol identity, then holds the record itself with `FILE_SHARE_READ`. Another process testing liveness attempts to open that record for write. A sharing violation proves the owner still holds it; process death closes the handle and releases the violation regardless of exit status. This gives a kernel-backed liveness signal immune to PID reuse. `holder_alive` additionally compares the recorded start time, so a new process with the same PID does not inherit the old identity. ### Adopting leases without an exclusion gap When a new engine image is fully published, the running server hands the native connection to a successor: 1. Generate a one-time handoff secret and pass it only to the child. 2. The child authenticates with `squad/lease_adopt` before serving ordinary requests. 3. For every held scope, the child takes its own deny-write handle **while the parent still holds its handle**. Concurrent read-sharing opens coexist. 4. `transfer_holder` re-points the tenancy records to the child's PID, start time, and session under `with_transition`. 5. Only after every handle and record moves does the parent withdraw and close. At no point does the target lose write-denying kernel exclusion. A failed adoption stops the child and leaves the parent serving with its original handles and records. ## 8. Fencing Continues Into Every Execution Surface Generation checks would be decorative if they stopped at the lease manager. They are wired into file replacement, SQLite, worker acts, and successor result delivery. ### File write funnel `require_write_binding` requires an exact file lease, expected generation equal to `lease.generation`, matching task, valid tenancy and control epoch, and a server execution handle. Parent-scope ownership alone cannot authorize an exact mutation. A governing source cannot also be the destination; that requires a separately governed revision task. The precondition compares bytes observed through `W` with the planned preimage. A mismatch at execution is `preimage_hash_mismatch`, not permission to proceed on stale content. ### Fenced SQLite mutation Protected row changes serialize through `with_transition` and open SQLite with `TransactionBehavior::Immediate`. Before update, the engine: - rereads the current row and requires `now_version == expected`; - captures recovery state; - revalidates target identity, `RequestedAct`, and approval; - executes a conditional `UPDATE ... WHERE key = ? AND version = ?`; - requires exactly one changed row; - reads the row back and verifies the new version and values before commit. A competing change becomes `record_version_changed` and returns the operation for review instead of applying it. The observed post-commit destination is checked again before the receipt finishes. ### `RequestedAct` and result grants For strict worker staging, `RequestedAct` derives the typed key and prepared version from exact projected bytes and binds the row task's current generation and control epoch. The worker cannot substitute a different row or revive an older generation. `result_grant.rs` applies the same current-tenancy rule to the receiving task. Its scope must still have a nonzero current generation and matching control epoch. Otherwise result-only tools remain withheld. A successor can receive a historical result only under fresh receiving authority; the original return is not fabricated. Break-glass adds the final lineage check: advancing the control epoch revokes leases minted under older epochs even when their wall-clock TTL has not expired. ## 9. Release Verification and Quarantine Release is another evidence boundary, not a lockfile deletion with good intentions. `release_content_check` compares on-disk bytes with: - `post_sha256` from the last gated write under the lease, when one exists; or - the lease's captured pre-write FNV-1a and byte count when no gated write occurred. That FNV-plus-length comparison is a **non-cryptographic drift check**, useful for noticing ordinary change, not for proving integrity against a malicious writer capable of engineering a collision. It is distinct from the SHA-256 comparisons used for gated-write and guarded-publication verification. If content matches, release archives the post-edit state to Amber. If it diverges, release archives the divergent bytes as **`quarantine_ungated`**, emits `released_with_ungated_content`, and returns `content_mismatch: true`. It never silently restores the baseline or deletes the foreign bytes. Out-of-band content is preserved for forensic review rather than being laundered into trusted history. ### The out-of-band owner lock kill-switch Breaking another agent's live, unexpired lease is refused by default. Force release is permitted only when an owner has written `.squad/control/lock-release.json` out of band, naming the authorized owners, scopes, setter, reason, and optional expiry. No MCP tool can create that authorization. The force break records `lock_force_released_under_owner_authorization` in the integrity graph. ### Exhibit 5 — Release quarantine instead of baseline laundering ```json { "released": "project:/artifacts/renewal-result.json", "owner": "reviewer-beta", "was_held": true, "amber_kind": "quarantine_ungated", "content_mismatch": true, "content_check": { "status": "mismatch", "baseline_source": "gated_write_post_sha256", "expected_sha256": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567", "actual_sha256": "c0ffee99887766554433221100ffeeddccbbaa99887766554433221100ffeedd" }, "warning": "Divergent bytes preserved in Amber and released_with_ungated_content recorded; baseline was not silently restored" } ``` ## Platform and Route Matrix | Route or condition | Enforcement and claim ceiling | |---|---| | Windows long-lived MCP server | Kernel handle exclusion is enforced across tenancy and all five replacement roles | | Windows short-lived CLI | Lease is advisory because process exit drops handles; protected execution requires a server-held handle | | Non-Windows | No equivalent Win32 share-mode, parent-pin, native-rename, or liveness claims; no false equivalence is asserted | | Handles or memory mappings opened before acquisition | Not retroactively revoked; the gate governs later incompatible opens, not every pre-existing capability | | Guarded replacement intervals | Write opens are continuously denied; delete/rename residuals are detected and reconciled as described, not claimed eliminated | | SQLite destination | `IMMEDIATE` transaction and version fence govern the row; this is not the Win32 file-handle path | | External SaaS or non-idempotent API | Requires destination-specific conditional-write, observation, and reconciliation guarantees; the local file gate does not prove external exactly-once effect | | Privileged host administrator or volume unmount | Outside the application's prevention claim; remaining exposure is named rather than hidden | The architecture is therefore strong without becoming unfalsifiable. It names the route it governs, the kernel primitives that enforce it, the interval it cannot close, and the evidence that remains when verification fails. Consider a writer that opened `GENERIC_WRITE` or a memory-mapped section **before** `G` was acquired. The new guard cannot retroactively revoke that capability, and the mapping may outlive the original handle. Release-time content comparison and `quarantine_ungated` can preserve evidence of the resulting drift, but they cannot turn that into retroactive prevention. A deployment must inventory and exclude pre-existing write capabilities before claiming exclusive mediation of the resource. Compatibility must be demonstrated for the exact Windows and filesystem profile in scope: destination present or absent, open readers, conflicting writers, reparse redirection, parent substitution, failure at every handoff, and crash recovery. NTFS/ReFS support is not inferred merely because the native rename API exists. ## 10. The 2026 Fencing Evidence Landscape The research supports monotonic execution fencing and exposes the same limitation that matters here: a mechanism proves only the effect classes its boundary actually covers. | 2026 primary source | What it establishes | How this architecture composes with it | |---|---|---| | Shu, [*Boundary-State Control for Tool-Using Language-Model Agents*](https://arxiv.org/abs/2609.37475), September 2026 preprint | BSC-R binds a single-use authorization to the exact action and semantic authorization state. It rejects all ten represented changed/replayed classes across 10,302 frozen proposals, prevents 1,200 represented non-replay invalid effects in CONTINUITY, handles 160 replay lifecycles, and withholds 200 ambiguous cases. In the broader external suite, it commits 25% of invalid effects. | This is direct evidence for exact-act and generation binding at commit. It is equally strong evidence for path and effect coverage ceilings: no fencing claim should exceed the contexts actually governed. | | Lyu et al., [*CoAgent: Concurrency Control for Multi-Agent Systems*](https://arxiv.org/abs/2606.15376), June 2026 preprint | MTPO fixes a serialization order, serves order-filtered reads, and repairs dependent plans while preserving serializable outcomes. CoAgent stays within 5% of serial correctness at a 1.4× speedup and improves a bash-only target from 45/71 to 63/71 passes at 0.80× time. | Multi-agent concurrency requires authoritative ordering and current-state evaluation, not prompt cooperation. AINA's transition mutex and workspace high-water sequence supply that ordering for the local reference embodiment. | | [Redisson FencedLock documentation](https://redisson.org/glossary/java-fencedlock.html), current 2026 reference | Defines a fencing token as a monotonically increasing value issued on each lock acquisition and presented to the resource so stale holders can be rejected. | This is the basic distributed-system model behind generation fencing. The Room adds two-phase counter persistence, kernel share-mode exclusion, and per-route verification rather than treating the token as self-enforcing. | | Microsoft [`CreateFileW` and file-sharing documentation](https://learn.microsoft.com/en-us/windows/win32/fileio/creating-and-opening-files), current 2026 reference | Desired access and sharing modes determine which later opens are compatible; `FILE_SHARE_DELETE` is required for rename/delete semantics. | Explains why the long-term `FILE_SHARE_READ` guard must hand off through `W`, `T`, and `H` rather than closing and reopening around the rename. | | Microsoft [`FILE_RENAME_INFORMATION_EX`](https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/ntifs/ns-ntifs-_file_rename_information), current 2026 reference | Defines POSIX rename and replacement flags and relative rename behavior. | Grounds the native handle-relative atomic replacement with POSIX semantics; it is a primitive, not a complete proof of continuous exclusion. | | Microsoft [`GetFinalPathNameByHandleW`](https://learn.microsoft.com/en-us/windows/win32/api/fileapi/nf-fileapi-getfinalpathnamebyhandlew), current 2026 reference | Resolves the canonical target of an opened file object through reparse indirection. | Grounds parent and published-object identity verification, catching cases where identity and content match but the leased name points elsewhere. | The functional-equivalence concession still applies. A conventional stack using leases, fencing tokens, compare-and-swap, MVCC, durable logs, and destination-specific conditional writes can defeat the same defect if it binds the same exact transition through the same predicates. AINA's value is the inspectable composition and its refusal/evidence grammar, not a claim that alternative coordination theory cannot work. **Evidence availability and performance boundary:** the implementation and replacement records are `CONTROLLED` for qualified, method-scoped technical review, with adverse findings permitted. They are not unrestricted public downloads or independent validation. Handle timing, lock contention, rename latency, and recovery cost must be measured for the exact Windows filesystem, antivirus/storage stack, destination state, and worker count; this article does not claim a universal performance envelope. ## The Adversarial Stale-Writer Challenge The **Shared-Resource Lease & Kernel Fencing Challenge** uses a deliberately hostile sequence, not just a happy-path replacement: 1. Worker A acquires generation 14 and stalls. 2. Worker B acquires generation 15 and completes a guarded replacement. 3. Worker A resumes and attempts both a stale-generation write and a direct incompatible file open. 4. The executor returns `stale_lease_generation`; the kernel returns `ERROR_SHARING_VIOLATION`. 5. A separate parent-redirection or preimage-displacement attempt produces typed reconciliation evidence and, where applicable, `quarantine_ungated` rather than an invented clean rollback. The challenge excludes retroactive revocation of pre-existing handles and mappings, universal OS/filesystem coverage, total effect coverage, and independent validation. Deployment and paid pilots are separately scoped; independent evaluation remains outside the commercial ladder through [Article 09](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/). ## The Twelve Execution Claims | ID | Bounded claim | Evidence and ceiling | |---|---|---| | `05-C01` | Six typed lease namespaces and hierarchical conflict rules govern ownership; FNV-1a is only a routing key | Implemented and author-verified for all lease requests; full canonical identity decides authority | | `05-C02` | Filesystem scopes reject ambiguous paths and resolve opened object identity through handles | Implemented and author-verified on local NTFS/ReFS targets | | `05-C03` | Workspace-wide monotonic generations use two-phase persistence and refuse rollback | Implemented and author-verified; missing state never resets to zero | | `05-C04` | A 30-second cross-process mutex serializes lease and write transitions | Implemented and author-verified for local controller and worker transitions; it is never stolen by age | | `05-C05` | Dead-holder re-granting requires same owner/task and proof of death, then mints a new generation | Implemented and author-verified; unknown or reused PIDs fail closed | | `05-C06` | Heartbeats clamp to 15 minutes and cannot resurrect expired tenancies; holder transfer is all-or-restored | Implemented and author-verified for current lifecycle paths | | `05-C07` | Long-lived servers deny later incompatible write/delete opens with `FILE_SHARE_READ` | Implemented and author-verified; pre-existing handles/mappings are not retroactively revoked and CLI leases are advisory | | `05-C08` | Parent directories are pinned by handle and re-resolved during replacement | Implemented and author-verified for parent redirection, not volume unmount or privileged host-admin behavior | | `05-C09` | Five-role overlapping replacement uses native POSIX handle rename with continuous write-open exclusion | Implemented and author-verified on supported Windows filesystems; delete/rename residuals remain separately disclosed | | `05-C10` | Surviving `W` handles prove positive preimage displacement | Implemented and author-verified for intact-link, nondeleted, different-path cases; delete-and-recreate can remain indistinguishable | | `05-C11` | Publication failures classify into six structured reconciliation records with exact next actions | Implemented and author-verified; no automatic rollback or clean success is invented | | `05-C12` | Server liveness is PID-reuse-resistant and image handoff preserves exclusion | Implemented and author-verified through sharing-violation records and shared handle adoption | ## The Article 05 Claim Card ```text Mechanism: AINA Typed Leases, Monotonic Generation Fencing, and Guarded Replacement (typed scopes, room.generation.v2, transition mutex, lifecycle guards, G/W/T/H/G' handles, parent pin, liveness, SQLite fences, and release quarantine) Failure it governs: Stale workers resurrecting after lease expiry; uncooperative local write/delete/rename opens; rolled-back generation stores and successor-tenancy confusion; parent junction swaps and destination reparse redirection; displaced preimages and divergent post-release content. Positive path: current task and tenancy → workspace-wide generation and control epoch → exact server-held scope and preimage observation → five-role guarded replacement with parent pin → identity + content + resolved-path + displacement verification → fresh destination observation → structured release or quarantine. Refusal / reconciliation path: stale_lease_generation, lease_expired, lease_task_mismatch generation_store_missing, generation_checkpoint_mismatch ERROR_SHARING_VIOLATION (32), destination_parent_redirected GuardLost, NameRedirected, IdentityChanged, ContentDiverged, PreimageDisplaced, Unverifiable record_version_changed, result_grant_tenancy_required content_mismatch and released_with_ungated_content. Evidence status: Implemented and Author-Verified on the declared Windows reference embodiment. 2026 research is primary-source context, not endorsement or independent validation. No universal external exactly-once or total-effect-coverage claim is made. Verifier and artifacts: .generations-initialized and generations.json lease tenancy records and pre-write state .lease-transition.mutex G/W/T/H/G' handle identities and content hashes parent object identity and resolved paths six structured PublishFault records server liveness records and transfer_holder SQLite readback and result_grant tenancy Amber post_edit and quarantine_ungated. Scope and claim ceiling: Win32 NTFS/ReFS reference mechanisms, not POSIX-equivalent guarantees. Later incompatible opens are denied; pre-existing handles/mappings are not revoked. Delete/rename intervals are bounded and reconciled, not claimed eliminated. Third-party delete-and-recreate can remain indistinguishable from ordinary replace. External effects require destination-specific guarantees and observation. No privileged host-admin or universal external exactly-once claim. Next stage in the machine: Article 05 enforces current ownership at the mutation window. Article 06 proves rollback is not erasure: Primary and Amber recovery, pre-write baseline verification, and ungated-content quarantine. Commercial doorway: Shared-Resource Lease & Kernel Fencing Challenge — bring one shared file or SQLite row exposed to multiple workers; receive a stale-generation replay, guarded replacement demonstration, and recovery/quarantine evidence map (Article 09 · me@blayneaina.com). ``` ## The Clock Is Not the Fence A lease expires because coordination needs a way forward. That does not mean the old worker has become harmless. Authority must change form at the resource: a new generation, a new holder, a new epoch, and a kernel-enforced exclusion that survives the replacement. The architecture does not stop at that success condition. It also preserves what happened when the world did not cooperate: which guard was lost, which name redirected, which object changed, which content diverged, which preimage moved, and which checks could not run. Failure becomes reconciliation evidence rather than a vague error or a fabricated rollback. That is the difference between a lease timer and execution fencing. > **Expiry lets ownership move. Generation decides whether an act is current. Kernel exclusion decides whether an uncooperative writer can open the file. Verification decides whether the published object actually earned success.** Article 06 takes up the next non-equivalence: a verified replacement still needs recovery. Rollback restores a prior state; it does not erase what happened, and ungated mutations must never be hidden inside a clean baseline. ## Frequently Asked Questions ### Why is lease expiry not fencing? Expiry allows ownership to move after a holder stops responding, but it does not erase a stale worker's process, cached authority, delayed request, or physical ability to write. The resource must reject the stale generation at execution. ### Is FNV-1a part of the security proof? No. FNV-1a is only a deterministic routing key for lease filenames. Authority decisions compare the complete canonical namespace and resource identity, current generation, task, control epoch, and execution handle. ### Why are five handle roles necessary? The long-term guard denies delete, but Windows requires delete sharing for rename-over. `G`, `W`, `T`, `H`, and `G'` hand off exclusion in overlapping stages so no interval permits a write open, while `P` pins the parent and verification proves the published object. ### Does the kernel gate revoke handles or mappings opened earlier? No. Share modes govern later incompatible opens. They do not retroactively revoke pre-existing handles, memory-mapped sections, or privileged host-admin behavior. Those routes require separate inventory and controls. ### What does preimage displacement detection prove? It proves that the original object kept all links, was not deleted, and now resolves to a different ordinary path. That means it moved before the rename and an uncaptured object was replaced. Delete-and-recreate can remain indistinguishable, so the article does not claim universal recovery. ### Does this guarantee exactly-once external effects? No. It proves current ownership and verified replacement on the governed Windows file route, plus version-fenced SQLite mutations. External services and non-idempotent APIs need their own conditional-write, observation, and reconciliation guarantees. ## Continue the Series - **Series index:** [From Proposal to Consequence: The AINA Transition-Governance Series](https://blayneaina.com/) - **Previous:** [Article 04 — The Room: Why Git and Chat Are Not Coordination Planes](https://blayneaina.com/04-the-room-why-git-and-chat-are-not-coordination-planes/) - **Next:** [Article 06 — Rollback Is Not Erasure: Dual-Tier Vaults and Ungated Quarantine](https://blayneaina.com/06-rollback-is-not-erasure-dual-tier-vaults-and-ungated-quarantine/) - **Deployment and independent review:** [Article 09 — Bring One Real Boundary](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/) --- ## Primary 2026 Sources - [Shu — *Boundary-State Control for Tool-Using Language-Model Agents: Commit-Time Consistency under State Drift*, `arXiv:2609.37475`](https://arxiv.org/abs/2609.37475) - [Lyu et al. — *CoAgent: Concurrency Control for Multi-Agent Systems*, `arXiv:2606.15376`](https://arxiv.org/abs/2606.15376) - [Redisson FencedLock Documentation](https://redisson.org/glossary/java-fencedlock.html) - [Microsoft `CreateFileW` and File-Sharing Documentation](https://learn.microsoft.com/en-us/windows/win32/fileio/creating-and-opening-files) - [Microsoft `FILE_RENAME_INFORMATION_EX`](https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/ntifs/ns-ntifs-_file_rename_information) - [Microsoft `GetFinalPathNameByHandleW`](https://learn.microsoft.com/en-us/windows/win32/api/fileapi/nf-fileapi-getfinalpathnamebyhandlew) --- ## Publication Metadata **Title tag:** Expiry Is Not Fencing: Typed Leases and Kernel Write Exclusion **Meta description:** Why wall-clock lease expiry cannot stop stale AI writes: typed scopes, monotonic generations, five-role Win32 handle exchange, preimage forensics, and release quarantine. **Series navigation:** [Series index](https://blayneaina.com/) · Article 05 of *From Proposal to Consequence: The AINA Transition-Governance Series*. **Previous installment:** [Article 04 — The Room: Why Git and Chat Are Not Coordination Planes](https://blayneaina.com/04-the-room-why-git-and-chat-are-not-coordination-planes/). **Next installment:** [Article 06 — Rollback Is Not Erasure: Dual-Tier Vaults and Ungated Quarantine](https://blayneaina.com/06-rollback-is-not-erasure-dual-tier-vaults-and-ungated-quarantine/). ================================================================================ # ARTICLE §06: Rollback Is Not Erasure: Dual-Tier Vaults and Ungated Quarantine Canonical URL: https://blayneaina.com/06-rollback-is-not-erasure-dual-tier-vaults-and-ungated-quarantine/ # Rollback Is Not Erasure: Dual-Tier Vaults and Ungated Quarantine *A backup can restore a clean state. It cannot explain what happened afterward. Recovery needs a certified baseline and a separate forensic history.* **By Blayne Aina · Principal Systems Architect and creator of AINA · October 31, 2026** --- A backup can tell you where to go back to. It cannot tell you what happened afterward. An autonomous agent edits a file five times, an external process changes it a sixth, and the framework “rolls back” by overwriting the file with its only snapshot. The clean baseline is gone. The intermediate states are gone. The unauthorized write is gone from the audit story. The system now looks exactly as it did before anything happened—which is precisely the evidence it needed to preserve. **Rollback is not erasure.** Restoring a clean file is one state transition. Preserving what occurred is another obligation. A governed recovery system needs both: - a trusted operational baseline for recovery; - an evidence-preserving account of every intermediate and ungated state for forensics. Article 05 proved current ownership at the write window. Article 06 makes recovery equally rigorous: certified checkout capture, dual-tier raw-byte vaults, replay-bound versions, bypass observation, release quarantine, and governed restoration. > **The baseline tells you where to revert. The vaults tell you why you may trust it—and what you must not pretend never happened.** ## 1. Why Single-Tier Undo Fails Single-slot backup systems fail in two distinct ways. ### Failure 1: The baseline is overwritten by the task itself ```text checkout → snapshot state 0 edit 1 → overwrite snapshot with state 1 edit 2 → overwrite snapshot with state 2 edit 3 → overwrite snapshot with state 3 ``` The task has an undo button but no state 0 left. Worse, it cannot explain how state 3 was reached or which intermediate state was valid. AINA keeps the Primary pre-checkout baseline intact and accumulates each mutation as a new content-addressed Amber object. N edits under one checkout produce N recoverable Amber states without consuming the Primary anchor. ### Failure 2: Ungated content becomes a trusted post-edit state An external process changes the file while a lease is active. At release, the system archives the current bytes as if the agent wrote them through the governed path. Provenance has been laundered: unauthorized content now looks like a legitimate post-edit artifact. AINA compares live bytes with the last authorized state. A mismatch becomes `quarantine_ungated`, emits `released_with_ungated_content`, and returns `content_mismatch: true`. The divergent bytes remain on disk for inspection; they are not silently restored or deleted. ## 2. The Dual-Tier Separation Primary and Amber are not two copies of one backup. They answer different questions and have different failure and custody properties. ### Exhibit 6.1 — Dual-tier operational separation ```text CHECKOUT INVOCATION │ Lease + kernel handle acquired │ save_checkout_copy(verified bytes) ┌──────────┴──────────┐ ▼ ▼ PRIMARY OPERATIONAL VAULT AMBER FORENSIC VAULT .squad/vault// $AMBER_VAULT/objects//.obj ├─ versioned raw bytes ├─ raw content-addressed bytes ├─ adjacent .sha256 sidecar ├─ deduplicated by content ├─ create_new + sync_all ├─ atomic write + exact readback ├─ PRIMARY_KEEP = 100 ├─ append-only amber.jsonl └─ bound by vault.jsonl replay └─ quarantine and per-edit states │ │ └──────────┬──────────┘ ▼ certified baseline + complete forensic history ``` Primary is the bounded operational store for rapid rollback. Amber is the append-oriented, content-addressed forensic store for checkout, edit snapshots, post-edit states, native consistent snapshots, and quarantine. “Append-oriented” describes its write model, not deletion-proof immutability; storage survival is a separate custody question addressed below. ## One Checkout, Six Mutations, Two Truths Consider one file checked out for remediation: 1. **Certified checkout.** The lease and kernel handle are acquired. State 0 is written to Primary with its sidecar and capture-ledger binding, and archived to Amber with exact readback. If any step fails, both lease and handle are released. 2. **Five governed edits.** Each pre-mutation state is appended to Amber as `edit_snapshot`. Primary remains the certified state-0 anchor; Amber now holds five additional states. 3. **One ungated write.** A separate process changes state 6 outside the tool path. The mtime-based sentinel may flag `bypass_detected`, but the more important fact is that the last authorized gated state no longer matches the live bytes. 4. **Release.** The current bytes hash differently from the last gated state. They become `quarantine_ungated`, with `content_mismatch: true` and `released_with_ungated_content`. The bytes remain available for inspection instead of disappearing into a clean release. 5. **Governed restoration.** Under fresh current authority, the system restores the certified state-0 bytes as another governed write and preserves the displaced current state first. The operation is a new transition, not an assertion that the earlier mutations never happened. The result is deliberately two truths: the file can return to a trusted baseline, while the evidence retains every intermediate and ungated state. A third-party observer may already have consumed the divergent bytes; that external consequence requires its own reconciliation and is not erased by restoring the file. ## 3. Primary: A Certified Operational Baseline A file on disk in the Primary directory is not a baseline yet. It becomes one only after its bytes, sidecar, scope identity, and durable capture binding all verify. ### Scope sharding and identity Each canonical file scope maps to `.squad/vault//`. The hash is a 64-bit FNV-1a value rendered as 16 lowercase hexadecimal characters, providing a fixed-width, path-safe directory identity without drive-colon or separator problems. `meta.json` binds the directory back to its canonical scope and hash. If saved metadata does not resolve to the requested file, the copy is rejected with `primary_scope_identity_mismatch`. Legacy directories from an older path spelling are considered only when their recorded scope resolves to the same canonical file. This is routing, not capability proof; Article 05 already established that FNV-1a cannot carry authority by itself. ### Monotonic version filenames Each capture receives a fresh name: ```text {:013}-{:06}-{random_8}__{sanitized_name} ``` The components are a 13-digit Unix-millisecond prefix, a six-digit sub-millisecond remainder, an eight-character random tag, and a sanitized terminal filename. Lexicographic order tracks capture order, while the random tag prevents same-instant collisions. `create_new_synced` uses create-new semantics, writes the data, and calls `sync_all`. Existing objects are never overwritten. Every version gets an adjacent `.sha256` sidecar containing the content digest. ### Readback before trust After writing the version and sidecar, `primary_store` reads both back. The saved bytes must equal the input exactly, and the sidecar must equal the computed SHA-256. A mismatch returns `primary_readback_mismatch` and prevents the protected change from proceeding. `prune_primary` keeps the most recent **100 versions per scope**, deleting older versions and their sidecars together. Operational retention is bounded without leaving orphaned digest files. ### Durable version binding A saved version is bound by appending `vault_checkout_saved` to `.squad/vault.jsonl` with the canonical scope, destination, byte count, content hash, kind, owner, and timestamp. `version_binding` then replays the entire ledger: - every nonempty line must parse; - the destination must name the selected version exactly; - the record's canonical scope must resolve to the same file; - duplicate or conflicting bindings are refused; - digest, byte count, owner, and RFC 3339 timestamp must be well formed. `verify_primary_version` requires all three independent predicates before restoration: 1. current bytes hash to the adjacent `.sha256` sidecar; 2. replay binds that exact version to the same canonical source, digest, and size; 3. a caller-supplied restore pin, when present, matches the verified digest exactly. A legacy object without a capture binding is accepted only against an independently supplied original restore pin. Otherwise the operator must re-baseline with a fresh checkout. No binding is synthesized retrospectively. ### Exhibit 6.4 — A valid Primary capture binding ```json { "action": "vault_checkout_saved", "scope": "project:/services/renewal.rs", "owner": "reviewer-beta", "intent": "governed renewal remediation", "dest": ".squad/vault/7c9a2e1d0f8b4a3c/0001730304000-000128-e4b8c9d2__renewal_rs", "bytes": 34120, "content_sha256": "4b92ec173d8f8a12b4e8035a92cfb912384a71b2d3080e7d56a89403d1685bb3", "kind": "checkout", "at": "2026-10-31T17:10:00.128000Z" } ``` The binding does not merely say a file exists. It certifies which source produced it, which bytes it holds, who captured it, and when the receipt became durable. ## 4. Amber: Content-Addressed Forensic Objects Amber stores raw bytes under: ```text $AMBER_VAULT/objects//.obj ``` The first two digest characters create 256 shard directories. There is no Zstandard or `.zst` container. Raw bytes eliminate codec selection from the recovery trust boundary and make the content address itself the object's identity. ### Atomic creation, deduplication, and exact readback When an object is new, bytes are written atomically and immediately read back. They must equal the input exactly. If an object with that digest already exists, writing is skipped and the receipt records `deduped: true`. A mismatch between path and content fails with `archived: false` instead of silently accepting a substituted object. Every capture appends to `amber.jsonl` with: - `event: "amber_archived"`; - capture kind, canonical scope and absolute source identity; - owner and intent; - SHA-256 and raw byte count; - timestamp and deduplication status. The five capture classes are `checkout`, `edit_snapshot`, `post_edit`, `quarantine_ungated`, and `native_snapshot`. Native consistent snapshots capture verified bytes rather than pretending a raw live database copy is consistent. `read_verified_amber` enforces a 64-hex digest, a caller-declared maximum size, and a full hash match before returning any bytes. Path traversal and oversized recovery artifacts are refused before they can exhaust memory. ## 5. Fail-Closed Checkout Capture Checkout is a state machine, not a best-effort backup call. ### Exhibit 6.2 — Fail-closed checkout capture state machine ```text LockRequested → manager lease + kernel OS handle acquired → Primary create_new + sidecar + full readback → Amber raw-byte archive + exact readback → append vault_checkout_saved to vault.jsonl → version_binding full replay ├─ SUCCESS → tenancy granted with certified baseline └─ FAILURE → manager::release(&lease) → oshandle::release(&scope) → return capture failure; no tenancy ``` The lease and kernel handle are acquired first so the source cannot be casually changed during capture. Any failure in storage, sidecar verification, Amber persistence, ledger append, or binding replay releases both. Tenancy is never granted without a certified recovery baseline. ### Exhibit 6.5 — Checkout failure revokes tenancy ```json { "error": "vault_capture_receipt_unconfirmed: the persisted capture receipt does not read back as this exact version", "baseline_certified": false, "manager_lease_released": true, "os_handle_released": true, "tenancy_granted": false } ``` The retained bytes can aid diagnosis, but they do not certify recovery or permit execution. ## 6. Per-Mutation Amber Accumulation `snapshot_edit_state` captures the current bytes before a governed mutation and archives them only to Amber with `kind: "edit_snapshot"`. Primary remains the state-0 undo baseline. ```text Primary: state 0 ───────────────────────────────► remains certified Amber: state 0 → state 1 → state 2 → … → state N every edit adds a content-addressed object ``` If ten edits occur under one checkout, Amber holds ten recoverable states while Primary retains the original anchor. A restoration to state 0 does not erase the other ten; a later investigation can still inspect them. ### Exhibit 6.6 — Amber edit snapshot ```json { "event": "amber_archived", "at": "2026-10-31T17:12:04.112000Z", "kind": "edit_snapshot", "scope": "project:/services/renewal.rs", "abs": "project:/services/renewal.rs", "owner": "reviewer-beta", "intent": "pre-edit snapshot before mutation 3", "sha256": "9a1f28b76e4c3d21a08456b9c832104e7651a2b3c4d5e6f7a8b9c0d1e2f3a4b5", "bytes": 35890, "deduped": false } ``` ## 7. The Active Bypass Sentinel A guarded file can still be changed by a process outside the tool path. The active bypass sentinel scans active File leases and cross-indexes three observations: - the lease's `started_at` timestamp; - the latest gated write found in `snapshot.jsonl` and `file_edits.jsonl`; - the file's current filesystem modification time. The bypass condition is: ```text mtime_ms > lock_started_ms AND mtime_ms > last_gated_write_ms ``` Path spellings are normalized so a junction, case variant, or slash variant cannot make a legitimate gated write look foreign. When the condition holds, the scanner returns the lease, owner, approximate mtime, last gated time, and a next action—and emits `bypass_detected` into the Room transaction graph. ### Exhibit 6.7 — `bypass_detected` ```json { "path": "project:/services/renewal.rs", "resolved_path": "project:/services/renewal.rs", "owner": "reviewer-beta", "lock_started_at": "2026-10-31T17:00:00Z", "mtime_approx_ms": 1730300012000, "last_gated_write_ms": 1730300005000, "ts": "2026-10-31T17:15:00Z" } ``` This is an **active heuristic**, not total kernel containment or proof of complete observation. A process can falsify timestamps, use pre-existing handles or mappings, or mutate through a route not represented in the index. Article 05's kernel exclusion governs later incompatible opens; the sentinel supplies additional forensic visibility rather than claiming omniscience. ## 8. Release Verification and Ungated Quarantine Release checks live on-disk bytes against the last authorized state. If a gated write occurred under the lease, the expected value is its `post_sha256`; otherwise the expected value is the checkout's pre-write FNV-1a hash and byte count. The no-gated-write comparison is a **non-cryptographic drift check**, not an integrity guarantee against a malicious writer deliberately engineering an FNV collision. Gated writes and guarded publication use SHA-256 instead. ### Exhibit 6.3 — Release verification and quarantine flow ```text Read live bytes and compute SHA-256 → compare with last authorized gated state ├─ MATCH → archive post_edit → release tenancy + OS handle └─ MISMATCH → archive divergent bytes as quarantine_ungated → emit released_with_ungated_content → return content_mismatch: true → release tenancy + OS handle → preserve divergent bytes for inspection ``` The mismatch path does not silently restore the baseline. It does not delete the divergent bytes. It quarantines them, signals the Room, and makes their disposition explicit. ### Exhibit 6.8 — Release quarantine receipt ```json { "released": "project:/services/renewal.rs", "owner": "reviewer-beta", "was_held": true, "amber_kind": "quarantine_ungated", "content_mismatch": true, "content_check": { "status": "mismatch", "baseline_source": "gated_write_post_sha256", "expected_sha256": "4b92ec173d8f8a12b4e8035a92cfb912384a71b2d3080e7d56a89403d1685bb3", "actual_sha256": "9a1f28b76e4c3d21a08456b9c832104e7651a2b3c4d5e6f7a8b9c0d1e2f3a4b5" }, "warning": "Divergent bytes preserved in Amber and released_with_ungated_content recorded; baseline was not silently restored" } ``` This is the non-laundering rule in executable form: unauthorized bytes become evidence, not a legitimate post-edit artifact. ## 9. Restoration Is Another Governed Mutation A restore does not travel backward through history and delete the unauthorized act. It creates a new, evidence-bound transition from current state to a verified prior state. ### Governed restore The ordinary restore path verifies the selected Primary bytes against the sidecar and full capture binding, optionally against an exact caller pin. It then dispatches through the normal protected-write pipeline with the current task and lease, generation and control epoch, expected pre-state, and displaced-state preservation. A caller cannot substitute a same-shaped file from another source, replace both version and sidecar to force agreement, or skip the task's execution checks. The restore itself becomes a governed write with recovery evidence. ### Console-exclusive operator rescue Operator rescue exists for the case where the ordinary protected-write gate itself cannot run. It is available only through the local interactive owner console with `--by operator`; it is denied to MCP sessions, scheduled jobs, and scripts. Rescue takes one exclusive handle with `ShareMode::None`. Any other open handle causes refusal before writing. It verifies the target object and expected failed image, preserves the failed image in Amber before the first byte changes, journals preparation, writes through the same handle, and requires an exact readback. `live_images` exposes live Squad server PIDs and compares each running image hash with the binary on disk, so a repaired file is not mistaken for a recovered fleet. It is not a universal enumeration of every process holding every handle. The rescue's crash ceiling is explicit: the write is in place, not a rename, and **not crash-atomic**. A crash between truncate and flush can leave partial bytes. `rescue_prepared` records the planned restore and preserved failed image before mutation, allowing later inspection and recovery rather than pretending the partial write never occurred. Governed restore is therefore the default; console rescue is the bounded exception with its own forensic and crash disclosures. ## 10. The Local-History and Non-Erasure Ceilings Recovery has three distinct assurance properties that must not be collapsed. ### Artifact integrity SHA-256 and capture binding prove that stored bytes still match the evidence recorded for them. This detects substitution, corruption, and misleading provenance. ### Artifact survival Hashes cannot stop a privileged administrator or hostile process from unlinking the vault. `rm -rf .squad` is not a threat a checksum can prevent. Durable survival requires separate custody, replication, and appropriately isolated write-once storage. ### Downstream irreversibility Restoring local bits cannot recall what other systems already did with them: - an external API or webhook already dispatched; - a database query or third-party sync already committed; - confidential content already read or disclosed by another process; - another agent's reasoning, notifications, or subsequent state transitions already caused by the modified state. Local rollback restores the file. It does not erase the causal residue. External systems must receive the quarantine signal and perform their own governed reconciliation. > **Integrity verifies the object. Custody determines whether it survives. Reconciliation determines what the rest of the world must do next.** ## 11. The 2026 Recovery Evidence Landscape The strongest 2026 sources support independent interception, out-of-band supervision, and tamper-evident forensic records. They do not justify pretending that local rollback erases global consequences. | 2026 primary source | What it actually establishes | How the recovery architecture composes with it | |---|---|---| | Shu, [*Boundary-State Control for Tool-Using Language-Model Agents*](https://arxiv.org/abs/2609.37475), September 2026 preprint | BSC-R binds a single-use commit authorization to the exact action and semantic authorization state, rejecting represented changed and replayed classes while still committing 25% of invalid effects in the broader external suite. | Recovery must bind the exact target and authorized pre-state at execution, not invoke a generic undo. The broader-suite limit also supports the article's refusal to claim total effect coverage or downstream erasure. | | Pino, [*Hard Stop: Kernel-Level Preemption and Containment for Rogue Agentic Execution*](https://arxiv.org/abs/2609.29808), September 2026 preprint, revised v2 | Presents a forensic autopsy of a multi-stage rogue-agent intrusion and argues for out-of-band supervisory control and deterministic preemption boundaries rather than relying on centralized model guardrails. | Supports exclusive kernel handles, fail-closed lease release, and console-exclusive rescue. It does not prove that every leaked handle can be revoked; the article preserves that ceiling. | | Qin et al., [*LLM Agents Can Easily Tamper With Their Own Traces*](https://arxiv.org/abs/2609.30266), September 2026 preprint | All tested local agent harnesses except Muse Code allowed agents to delete traces without triggering monitor guardrails; external attackers could induce deletion, and tampering also emerged during reward optimization. | Forensic logs need an independent interception and custody boundary. Primary replay, content-addressed Amber objects, and protected controller storage help, but artifact survival still requires separate custody. | | Bindschaedler, Botha, and Siebenbrunner, [*Agent Flight Recorder*](https://arxiv.org/abs/2609.01931), September 2026 preprint | Binds intent through execution and provenance into canonical events, using hash chains, Merkle batches, and on-chain epoch roots for independent verification. Synthetic ablations detected edit, delete, reorder, and fork tampering at 100% with zero false positives. | Demonstrates how independent witnesses can verify evidence without trusting the agent's infrastructure. AINA's local capture bindings provide the source evidence; they do not by themselves provide a blockchain or independent validation tier. | | Ghosh, [*Tracekit: Tamper-Evident Intent-Reasoning-Action Auditing for AI Agents*](https://arxiv.org/abs/2609.35659), September 2026 preprint, revised v2 | A signer outside the agent's reach signs tool decisions and results into a hash chain with independent witnesses. A signed log detected six of eight tamper classes and a witness caught the other two; policy-only rules missed 211 of 420 blind harmful calls. | Shows why provenance must be captured outside the model's authority and why policy alone is insufficient. It also distinguishes evidence integrity from independent validation, matching Amber's role and its explicit ceiling. | These sources reinforce the architecture without becoming its assurance grade. The vaults remain implemented and author-verified local mechanisms, not independent certification or proof of complete history. **Evidence availability and operating-cost boundary:** the vault, capture, and reconciliation records are `CONTROLLED` for qualified, method-scoped technical review, with adverse findings permitted. They are not unrestricted public downloads or independent validation. Capture latency, disk amplification, replay cost, and quarantine recovery time must be measured against the actual object sizes, edit count, filesystem, and retention profile; this article does not claim a universal performance envelope. ## The Forensic Baseline & Quarantine Challenge The **Forensic Baseline & Out-of-Band Mutation Quarantine Challenge** exercises the complete sequence: certified checkout, several Amber-accumulated edits, an out-of-band mutation, release-time mismatch, and a governed restoration. The returned evidence distinguishes the operational baseline, per-edit forensic history, quarantined divergence, and any external effect that remains irreversible. The challenge excludes total bypass observation, deletion-proof history, universal downstream erasure, and independent validation. Deployment and paid pilots are separately scoped; independent evaluation remains outside the commercial ladder through [Article 09](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/). ## The Twelve Recovery Claims | ID | Bounded claim | Evidence and ceiling | |---|---|---| | `06-C01` | FNV-1a scope sharding and `meta.json` bind Primary directories to canonical file identity | Implemented and author-verified; mismatched saved scopes are refused | | `06-C02` | Monotonic version names and create-new synced writes prevent accidental overwrite | Implemented and author-verified; collisions fail instead of reusing an object | | `06-C03` | Primary bytes and sidecars are read back immediately after writing | Implemented and author-verified; mismatch aborts protected change | | `06-C04` | Full ledger replay binds exact versions to canonical source, digest, size, owner, and time | Implemented and author-verified; corrupt, conflicting, or missing bindings are not certified | | `06-C05` | Checkout failure releases both coordination lease and kernel handle | Implemented and author-verified; no tenancy without certified baseline | | `06-C06` | Amber stores raw, content-addressed bytes with deduplication and exact readback | Implemented and author-verified; no `.zst` or compression trust boundary | | `06-C07` | Each governed edit can append an Amber state without touching Primary | Implemented and author-verified; N edits preserve N states plus the state-0 anchor | | `06-C08` | Primary retention retains 100 versions and deletes sidecars with expired versions | Implemented and author-verified; operational storage remains bounded | | `06-C09` | The active sentinel flags mtime evidence of out-of-band writes and emits `bypass_detected` | Implemented and author-verified as a heuristic, not total observation or containment | | `06-C10` | Release mismatch becomes `quarantine_ungated` with `content_mismatch: true` | Implemented and author-verified; divergent bytes are preserved, not silently reverted | | `06-C11` | Governed restore and console rescue verify sources and preserve displaced state | Implemented and author-verified; rescue is console-exclusive, and its in-place write is not crash-atomic | | `06-C12` | Recovery explicitly separates integrity, survival, and downstream irreversibility | Author-verified design ceiling; no deletion-proof history or erasure of external effects | ## The Article 06 Claim Card ```text Mechanism: AINA Dual-Tier Recovery and Forensic Quarantine (Primary certified baselines, Amber content-addressed objects, fail-closed checkout, version replay, per-edit snapshots, bypass sentinel, release verification, and governed restore) Failure it governs: Single-slot undo destroying the original baseline; uncertified files impersonating recovery images; out-of-band writes laundered into legitimate post-edit state; silent rollback destroying evidence and obscuring downstream consequences. Positive path: lease + kernel handle → Primary create_new + sidecar + full readback → Amber raw-byte archive + exact readback → vault.jsonl durable binding and full replay → gated edits accumulate Amber states → release content verification → clean post-edit release or quarantine_ungated → governed restore with current authority and displaced-state preservation. Refusal / quarantine path: primary_readback_mismatch, vault_capture_receipt_unconfirmed primary_scope_identity_mismatch, primary_capture_binding_missing manager::release + oshandle::release at failed checkout bypass_detected heuristic signal content_mismatch and released_with_ungated_content refused_intervening_change and rescue_effect_unconfirmed. Evidence status: Implemented and Author-Verified on declared Windows recovery paths. 2026 research supplies independent architectural context, not endorsement. No independent-validation, total-history, or total-bypass-detection claim. Verifier and artifacts: Primary version and .sha256 sidecar meta.json and version_binding replay Amber .obj and amber.jsonl checkout/edit_snapshot/post_edit/quarantine_ungated/native_snapshot records bypass_detected and released_with_ungated_content signals rescue_prepared and rescue_outcome records with exact readback. Scope and claim ceiling: Raw-byte integrity does not prevent privileged deletion of the store. Timestamp scanning is heuristic, not comprehensive kernel observation. Restoration is a new governed mutation, not historical erasure. Rescue in-place writes are not crash-atomic; preparation enables reconciliation. Local rollback cannot recall downstream reliance, disclosure, or propagation. No universal history, external-effect erasure, or independent-validation claim. Next stage in the machine: Article 06 certifies recovery and preserves ungated evidence. Article 07 pins the controller-fixed act, fences the destination mutation, and proves original-session or non-fabricating successor receipt. Commercial doorway: Forensic Baseline & Out-of-Band Mutation Quarantine Challenge — bring one shared file exposed to out-of-band edits; receive a certified checkout, Amber state-history demonstration, release-quarantine replay, and recovery ceiling map (Article 09 · me@blayneaina.com). ``` ## The Baseline Is Only Half the Recovery Story A trusted recovery image matters. So does the record of why it was needed. The dual-tier architecture preserves both. Primary certifies the state to restore. Amber preserves every state the system passed through, including the state nobody authorized. Checkout failure revokes tenancy rather than letting execution proceed without a recovery path. Release mismatch becomes a signal rather than a silent overwrite. Restoration creates another governed transition rather than deleting the past. That is how a system can recover operationally without falsifying its own history. > **Rollback restores a state. Erasure would destroy evidence. AINA preserves the baseline, the divergence, and the obligation to reconcile them.** Article 07 moves to the act and its return: whether the controller fixed the exact mutation before the worker launched, whether the destination accepted it under current tenancy, and whether the original session—or an explicitly non-fabricating successor—received the result. ## Frequently Asked Questions ### Why is rollback not erasure? Rollback changes current bytes to a prior state. It does not remove the fact that an action occurred, erase downstream reliance, or eliminate the need to investigate it. A governed recovery preserves both the restored baseline and the complete forensic history. ### What makes a Primary version a certified baseline? The saved bytes must match the original readback, the adjacent sidecar must match their SHA-256, the full capture-ledger replay must bind that exact version to the same canonical source, digest, and size, and any caller pin must match. Unbound files are not trusted recovery images. ### Why keep two vaults instead of one backup? Primary preserves the clean pre-checkout operational baseline. Amber accumulates per-edit, post-edit, native snapshot, and quarantine objects. Separating them prevents repeated edits or ungated writes from overwriting the baseline and laundering provenance. ### Does `bypass_detected` prove every out-of-band write was seen? No. It is an mtime and ledger-based heuristic. Timestamp manipulation, pre-existing handles, memory mappings, and unindexed paths can evade it. Article 05 supplies kernel exclusion for later incompatible opens; the sentinel adds forensic visibility. ### What happens when release content does not match the last gated state? The divergent bytes are archived to Amber as `quarantine_ungated`, the Room receives `released_with_ungated_content`, and the receipt reports `content_mismatch: true`. The bytes remain on disk for inspection instead of being silently restored or deleted. ### Can recovery undo external API calls or disclosures? No. Recovery restores local file state. It cannot recall webhooks, third-party syncs, queries, disclosures, or another agent's subsequent reasoning. Those require destination-specific reconciliation driven by the quarantine evidence. ## Continue the Series - **Series index:** [From Proposal to Consequence: The AINA Transition-Governance Series](https://blayneaina.com/) - **Previous:** [Article 05 — Expiry Is Not Fencing: Typed Leases and Kernel Write Exclusion](https://blayneaina.com/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/) - **Next:** [Article 07 — Defeating Approval Laundering: Act Pinning and Successor Grants](https://blayneaina.com/07-defeating-approval-laundering-act-pinning-and-successor-grants/) - **Deployment and independent review:** [Article 09 — Bring One Real Boundary](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/) --- ## Primary 2026 Sources - [Shu — *Boundary-State Control for Tool-Using Language-Model Agents*, `arXiv:2609.37475`](https://arxiv.org/abs/2609.37475) - [Pino — *Hard Stop: Kernel-Level Preemption and Containment for Rogue Agentic Execution*, `arXiv:2609.29808`](https://arxiv.org/abs/2609.29808) - [Qin et al. — *LLM Agents Can Easily Tamper With Their Own Traces*, `arXiv:2609.30266`](https://arxiv.org/abs/2609.30266) - [Bindschaedler, Botha, and Siebenbrunner — *Agent Flight Recorder*, `arXiv:2609.01931`](https://arxiv.org/abs/2609.01931) - [Ghosh — *Tracekit: Tamper-Evident Intent-Reasoning-Action Auditing for AI Agents*, `arXiv:2609.35659`](https://arxiv.org/abs/2609.35659) --- ## Publication Metadata **Title tag:** Rollback Is Not Erasure: Dual-Tier Vaults and Ungated Quarantine **Meta description:** Why single-tier undo fails autonomous agents: certified Primary baselines, content-addressed Amber forensics, fail-closed checkout, bypass detection, and ungated quarantine. **Series navigation:** [Series index](https://blayneaina.com/) · Article 06 of *From Proposal to Consequence: The AINA Transition-Governance Series*. **Previous installment:** [Article 05 — Expiry Is Not Fencing: Typed Leases and Kernel Write Exclusion](https://blayneaina.com/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/). **Next installment:** [Article 07 — Defeating Approval Laundering: Act Pinning and Successor Grants](https://blayneaina.com/07-defeating-approval-laundering-act-pinning-and-successor-grants/). ================================================================================ # ARTICLE §07: Defeating Approval Laundering: Act Pinning and Successor Grants Canonical URL: https://blayneaina.com/07-defeating-approval-laundering-act-pinning-and-successor-grants/ # Defeating Approval Laundering: Act Pinning and Successor Grants *Approval must bind the exact act before the worker environment exists, authority must be re-derived at commit, and recovery must not fabricate the original session's receipt.* **By Blayne Aina · Principal Systems Architect and creator of AINA · October 31, 2026** --- Hashing a worker's proposal after it arrives is too late. A human or policy may approve action \(A\), but the worker can still execute \(A'\): a different row, target, payload, route, or time. A safe execution boundary cannot trust that the proposal and its effect remain identical unless the trusted controller fixes the act before the worker environment exists and revalidates that exact object at the destination's ordering point. Three coupled failures follow: 1. **Launch substitution:** a worker begins life with an unclassified handle, an inherited capability, or a native unconfined route. 2. **Approval laundering:** a valid general approval becomes a different mutation at execution. 3. **Crash-straddling:** the destination commits, the worker dies before first receipt, and a successor either repeats the effect or invents original-session completion. Article 06 preserved recovery evidence. Article 07 binds the execution itself: pre-resume kernel inspection, controller-brokered model egress, `RequestedAct`, fenced SQLite execution, multi-destination holds, and non-fabricating successor grants. > **The controller fixes the act. The destination rechecks it. History may be read by a successor, but the original return is never invented.** ## One Approved Mutation, Three Places to Fail Suppose a policy approves raising one customer's credit limit from 10,000 to 25,000. The trusted controller first projects exactly that row and prepares the change. A prompt-injected worker then tries to substitute a different customer and a higher limit. At the same time, an unrelated writer increments the approved row's version, and the worker's host dies immediately after the database commits. The machine handles each failure separately: - **Before launch:** the suspended worker's package SID, LPAC attribute, capabilities, Job membership, and current handle inventory are verified. A sentinel match or unclassified handle terminates the Job before instruction 0. - **Before execution:** the controller has already fixed the exact row, key, fields, and expected version in `RequestedAct`. The substitute customer is not a permissible interpretation of that act. - **At the destination:** SQLite takes its write reservation, the row is reread, recovery is captured, and current task, source, tenancy, act, and approval are revalidated. The other writer's version increment makes the conditional update fail instead of overwriting new state. - **After commit:** a fresh result-only successor may read the historical outcome if its tenancy, grant, wake, source, and Room-return proofs all hold. It does not repeat the write and does not claim the dead worker received the result. Approval is therefore not one event at the prompt. It is a chain of proof that survives substitution, concurrency, and crash—without being confused with the original session's receipt. ## 1. Pre-Resume Kernel Verification A confined process must be inspected before instruction 0. Once the worker runs, every later control is negotiating with an already-active environment. ### Exhibit 7.1 — Pre-resume kernel audit ```text Derive exclusive AppContainer profile or per-launch host SID → CreateProcessW(CREATE_SUSPENDED) inside owned Job Object → Query raw token: package SID + WIN://NOALLAPPPKG + exact capabilities → PssCaptureSnapshot + PssWalkSnapshot handle table → DuplicateHandle + CompareObjectHandles against owned references → reject any sentinel match or unclassified object ├─ PASS → ResumeThread exactly once └─ FAIL → terminate suspended Job; no worker instruction executes ``` ### Exclusive identity and anti-reuse `container_sid` calls `CreateAppContainerProfile` for a fresh per-launch profile. `E_ALREADY_EXISTS_HR = 0x8007_00B7` becomes `confinement_identity_reused`; an existing profile cannot be borrowed with its old ACLs, files, or residual state. Brokered hosts that cannot run inside an AppContainer receive an application-defined Resource Manager SID in `S-1-9`. Its first sub-authority is ASCII `"SQAD"` (`0x5351_4144`), followed by four 32-bit words carrying 128 SHA-256 bits derived from the launch's container name. It names no local user, group, or domain account, so host permissions come only from explicit launch grants. ### Raw LPAC attribute verification `verify_worker_token` queries the suspended process token directly: - `TokenAppContainerSid` must equal the owned package SID; - `TokenSecurityAttributes` must contain `WIN://NOALLAPPPKG`; - the attribute must be integer kind `2`, count `1`, not disabled (`flags & 0x10 == 0`), and nonzero. The controller does not rely on a platform LPAC query that may be unavailable. It parses the raw table with buffer bounds and UTF-16 decoding, then compares the actual restriction to the requested one. ### Handle-table object identity, not numeric equality Windows handle values are process-local indices. They do not prove which kernel object they name, and two different values can name the same object. `PssCaptureSnapshot` and `PssWalkSnapshot` enumerate the suspended child's current handle table. The controller duplicates candidate handles and calls `CompareObjectHandles` against owned references for expected stdio, Job, and other provenance objects. A deliberately excluded sentinel proves whether anything outside the launch list was inherited. Unknown objects, missing expected matches, sentinel matches, incomplete inventory, or token mismatch refuse the resume. PSS proves the **current handle inventory at capture time**, not historical inheritance. The architecture therefore combines the explicit launch `HANDLE_LIST`, current inventory, and object comparison rather than inventing an inheritance proof. ### The native spawn gate `require_native_spawn_allowed` refuses a native launch or existing-session attachment whenever owner policy or staged grants mandate confinement. Expired or revoked grants do not create an unconfined fallback. Project-mandated work runs only under its staged project identity. This is the anti-fallback rule: if the worker cannot launch under the required boundary, it does not launch. ## 2. Controller-Brokered Model Egress Confined workers have no general network capability. Model access is a separate controller-owned transition, not an ambient HTTP client. ### Authenticated local channels For LPAC workers, the controller exposes a per-worker `AF_UNIX` endpoint with a restrictive ACL. It verifies the held worker with `IsProcessInJob`, `verify_worker_token`, exact capabilities, and a 32-byte launch nonce. It does not pretend that AF_UNIX supplies a portable peer-process identity; the held process is the evidence. Brokered hosts use a first-instance Named Pipe with `PIPE_REJECT_REMOTE_CLIENTS`. `QueryFullProcessImageNameW` must resolve to the exact staged bridge executable. Every request and response is bounded and rechecked. Revocation before response transport produces a typed refusal; partial responses produce `severed_mid_response`, not a second contradictory frame. ### The controller owns the route The worker names only an owner route ID and JSON body. The controller owns: - exact HTTPS destination or isolated subscription CLI; - credential injection and fixed headers; - allowed models and output-token ceilings; - request/response byte limits and request budget; - project, agent, expiry, and control-epoch admission. Routes can use API credentials or isolated `claude.exe`/`codex.exe` subscription CLIs. The worker never receives the credential or chooses a URL. ### Journal before dispatch; never replay uncertainty Every request is appended to `model-egress/ledger.jsonl` **before** network transmission. A last admission check occurs immediately before send. Stopping a broker stops the worker from receiving a result, but it does not recall an in-flight HTTP request. `settle_in_flight` waits for controller-side outcomes to append. If the provider outcome remains unknown, the operation stays `unresolved`, counts against the route's budget, and can be settled only by a trusted operator through `reconcile` using provider-side evidence. Reconciliation never resends and does not automatically return the budget slot. ### Exhibit 7.6 — Unresolved model egress ```json { "operation_id": "model_op_synthetic_07", "route": "owner-approved-analysis-route", "destination": "https://models.example/v1/respond", "actor": "reviewer-beta", "task": "TSK-2026-ACT-007", "project_id": "proj_synthetic", "begun_at": "2026-10-31T19:00:00Z", "request_sha256": "4a9c8e12f03b7d651928374655647382910abcdef1234567890abcdef1234567", "latest_outcome": null, "disposition": "unresolved; never resent; operator reconciliation required" } ``` An approved model route does not prove semantic confidentiality of the staged projection. It proves destination, credential, budget, and continuity authority. Content still needs Semantic Switch to transform native values before provider-exposed qBlock and later computation, with native rows and the private crosswalk retained locally as Article 03 establishes. ## 3. Controller-Fixed `RequestedAct` The act is derived from exact projected bytes before the worker environment is issued. It is not re-read from live model prose at execution time. This is the **strict confined-worker staging path**. The destination also admits controller-driven record changes, whose authority must not be described as if every call used identical staging. Those changes still require the exact task, row, pre-state, act, approval, and current tenancy; they do not inherit the worker environment's proof by implication. The common invariant is exact-act and current-authority verification, not a claim that all routes are physically identical. ### Exhibit 7.2 — `RequestedAct` lifecycle ```text prepare: projected bytes → one typed row key + integer version + permitted changes → persist: environment + task authority + source tenancy + RFC 8785 digest → create-only requested_acts/.json + sync_all + exact readback → worker receives environment and act reference → validate: open with FILE_SHARE_READ and retain handle through UPDATE/commit → re-derive current task authority, source tenancy, and exact destination → historical: verify completed act without consulting current tenancy ``` ### Derivation `prepare` requires one staged SQLite row and validates that its typed key matches the projection. Proposed changes must be nonempty, at most 64 fields and 16 KiB, exclude the key and `version`, and remain inside the declared projection fields. The current version must be a nonnegative integer. The destination becomes `{source}#{table}#{key_column}={typed_key}`. Worker input does not select a different row. ### Durable fixation `persist` re-resolves task authority and source tenancy, then binds: - environment ID and worker; - staging actor and complete task binding; - exact source projection receipt; - source generation and source tenancy; - the controller-fixed destination. RFC 8785 JCS canonicalization produces the act reference. The record is written with create-new semantics to `requested_acts/.json`, flushed with `sync_all()`, and read back for exact body and digest equality. JCS is an **Informational RFC**, and the implementation uses it to make the digest repeatable rather than to make the act authorized by hashing alone. ### Execution revalidation `validate` opens the act with Windows `share_mode(1)` and retains that handle through update and commit. It recomputes the digest, re-derives current task authority and source tenancy, and requires the environment, worker, actor, task binding, source, generation, and exact destination to match. Any authority drift becomes `requested_act_authority_changed`; any target, key, field, or version divergence becomes `requested_act_mismatch`. The pinned handle prevents deletion or replacement of the act object during the mutation window. ### Historical verification `historical` verifies a completed act against the immutable submission without consulting current tenancy (`current_authority_consulted: false`). A retired environment or released row task does not falsify a valid past act. It checks the reference, environment, worker, task, staged source, tenancy-at-staging, and exact destination rather than reviving execution authority. ### Exhibit 7.4 — Sanitized `RequestedAct` record ```json { "schema": 1, "environment_id": "env_synthetic_worker_07", "worker": "reviewer-beta", "actor": "controller-alpha", "task_binding": { "id": "TSK-2026-ACT-007", "owner": "controller-alpha", "project_id": "proj_synthetic", "project_root_pin": "project:/workspace", "source": "doc_20261030_142011_a1b2c3d4", "scope": "session://sqlite.records/accounts.db#customers#customer_id=cust_994182", "coordination_only": false, "epoch": 4, "revision": 2, "delegation": null }, "source": { "path": "project:/data/accounts.db", "projection": { "format": "sqlite", "table": "customers", "row_key": "customer_id", "row_values": ["cust_994182"], "fields": ["tier", "status", "credit_limit"] }, "sha256": "4a7b8c9d0e1f2a3b4c5d6e7f8a9b0c1d2e3f4a5b6c7d8e9f0a1b2c3d4e5f6a7b" }, "source_generation": 12, "source_tenancy": { "scope": "session://sqlite.records/accounts.db#customers#customer_id=cust_994182", "task": "TSK-2026-ACT-007", "generation": 12, "control_epoch": 4 }, "act": { "destination": { "target": "accounts.db#customers#customer_id=cust_994182", "key": "cust_994182", "key_field": "customer_id", "expected_version": 3, "fields": { "tier": "enterprise_gold", "status": "active" } } } } ``` ### Exhibit 7.5 — `requested_act_mismatch` ```json "requested_act_mismatch: proposed target, key, changes or prepared version differs from the controller-fixed act" ``` The worker's altered proposal does not acquire authority by resembling the approved one. ## 4. The Fenced SQLite Execution Window `destinations/records.rs` executes the mutation inside one ordered window that joins lease authority, transaction isolation, act validation, recovery, and observation. ### Exhibit 7.3 — SQLite fenced execution flow ```text acquire file coordination lease → with_transition → BEGIN IMMEDIATE (write reservation) → reread row and require now_version == expected → capture prior row values + JCS digest for recovery → recheck locator identity → revalidate pinned RequestedAct → recheck live approval → recheck locator identity again → UPDATE ... WHERE key = ? AND version = expected → require changed_rows == 1 → pre-commit readback: exact fields and version = expected + 1 → transaction.commit() → release lease → independent post-commit observation and durable outcome ``` `BEGIN IMMEDIATE` starts a write transaction immediately instead of waiting for the first write statement. The version predicate adds a second, destination-native optimistic concurrency check: if an uncooperative writer moved the row, the conditional update cannot satisfy `key + expected_version`, and the operation refuses rather than overwriting it. The transaction is not trusted merely because it is immediate. The engine also revalidates target, act, and approval inside that same window and requires readback to equal the approved result before commit. ### Exhibit 7.7 — Row-version refusal ```json "record_version_changed: the row moved to version 4 after operation model_op_synthetic_07 was proposed; nothing applied, it returns for review (reconcile the operation as absent)" ``` The proposed version was 3. The current version is 4. Nothing was applied. ### Recovery capture is separate from the mutation Before update, the engine records prior row values and their JCS digest through `prepare_recovery`. A governed restoration is separately authorized and requires the exact post-operation state. Recovery is not a license to repeat the original mutation or to erase its uncertainty. ### Exhibit 7.8 — Pre-write recovery capture ```json { "method": "row_values", "version": 3, "values": { "customer_id": "cust_994182", "tier": "standard", "status": "pending", "credit_limit": 10000 }, "sha256": "58b2e91f4c03d8a7126590ef4312ab98c7654321fedcba09876543210fedcba9" } ``` The capture proves what preceded the write. It does not prove that the write occurred or that its result was returned. ## 5. Multi-Destination Capability Plans A coordinated plan is not a distributed transaction. Each destination declares what its own write path can actually guarantee. | Destination | Declared guarantees | `cross_destination_atomic` | |---|---|---| | `sqlite.records` | `conditional_write`, `idempotent_retry`, `recovery_capture` | `false` | | `goose.sessions.db` | `conditional_write`, `idempotent_retry`, `recovery_capture` | `false` | | `opencode.http_api` | Empty set; every fresh effect is held | `false` | A plan names required guarantees for each step. `require_route` evaluates them before an operation begins and again before effect. An unmet requirement produces `destination_route_incompatible` and the step is held. Nothing silently accepts the gap. A versioned revision is the only way to change those requirements. It names every removed requirement and the weaker semantics now in force, records its authorization, and invalidates decisions made for the prior version. Authorization comes from a trusted operator other than the plan's author, or from the author only when every step runs under the owner's exact project assignment. Partial completion is tracked per step. Compensation is a separately governed restore naming the exact applied operation. There is no simulated 2PC and no `cross_destination_atomic: true` hiding behind an orchestration diagram. ## 6. Non-Fabricating Successor Grants A committed effect and its first receipt are different events. A crash can separate them permanently. ### The `ResultGrant` A controller may issue a `ResultGrant` for one historical proposal to a **new receiving task and environment**. The grant binds: - the original proposal ID and nested payload digest; - the whole immutable submission record digest; - the whole executor outcome digest and outcome class; - the receiving worker, task, environment, project, tenancy, and control epoch; - the issuer's current authenticated controller identity and grant expiry. The receiving environment gets persistent `purpose: "result_read_only"`. It receives no destination tenancy, proposal, apply, restore, private-write, or terminal authority. Revoking or superseding the grant cannot restore those privileges. The original controller remains historical provenance. The current trusted issuer must own and stage the receiving task. ### What a successor acknowledgement must prove `successor_return_for_task` verifies all of the following: 1. The expected original's current proposal head and receiving-environment head name the same unrevoked, unexpired grant, and the saved original has not changed. 2. The receiving task, environment, tenancy, Room, and current lease still reproduce the grant's binding. The receiving task remains a fresh result-only recipient with no protected submission of its own. 3. Exactly one acknowledgement exists. Its source event is fresh, not the original delivery. 4. The native `room-wake-job-v3` verifies, names the expected Room, worker, task, sender, registration, attempt, and has `dispatch_status: "completed"`. 5. Acknowledgement time lies inside the grant's validity window and the wake job's sealed window. 6. The acknowledgement's delivery identity matches the actual sealed wake job. 7. The governing source was served to that worker on the same broker session. 8. A fresh correlated Room return exists and every addressed mention and obligation is resolved by the Room service at or after the reply. Only then does the return report an historical outcome. It does not execute, repair, issue, revoke, register, or resend anything. ### Exhibit 7.9 — Successor acknowledgement ```json { "schema": 1, "kind": "successor_result_receipt", "successor": true, "satisfies_original_return": false, "resolves_original_obligation": false, "grant_id": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "proposal_id": "a1b2c3d4e5f60718293a4b5c6d7e8f901a2b3c4d5e6f708192a3b4c5d6e7f809", "proposal_sha256": "8f434346648f6b96df89dda901c5176b10a6d83961dd3c1ac88b59b2dc327aa4", "record_bytes_sha256": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0c1d2e3f4a5b6c7d8e9f0a1b2", "outcome_bytes_sha256": "9f8e7d6c5b4a3f2e1d0c9b8a7f6e5d4c3b2a1f0e9d8c7b6a5f4e3d2c1b0a9f8e", "outcome": "applied", "outcome_at": "2026-10-31T19:22:15.012480Z", "worker": "reviewer-beta", "task": "TSK-2026-SUCCESSOR-009", "environment_id": "env_synthetic_successor_11", "worker_session": "worker-mcp:00000000-0000-0000-0000-000000000011", "delivery_id": "dlv_synthetic_fresh_wake_8821", "delivery": { "room": "architecture-demo", "agent": "reviewer-beta", "task": "TSK-2026-SUCCESSOR-009", "source_event": "evt_0123456789abcdef_0000beef", "wake_attempt": 1, "worker_pid": 14208 }, "at": "2026-10-31T19:24:00.912004Z" } ``` The original return obligation is untouched. The successor has read and acknowledged the historical result; the original session has not been made to witness it. ## 7. Official Platform and Standards Crosswalk The architecture uses existing platform primitives without pretending they are complete governance systems by themselves. | Platform or specification | Native property | How AINA composes it into execution standing | |---|---|---| | Microsoft AppContainer and LPAC security architecture | Package identity, capability restriction, and the `WIN://NOALLAPPPKG` attribute can limit ambient package access. | Fresh identity per launch, exact raw-token verification, exclusive grants, and anti-fallback native spawn refusal. | | Microsoft [`CompareObjectHandles`](https://learn.microsoft.com/en-us/windows/win32/api/handleapi/nf-handleapi-compareobjecthandles) | Determines whether two handles refer to the same underlying kernel object without imposing object-specific access requirements. | Audits suspended-child handles by object identity, expected provenance, and deliberate sentinel exclusion instead of trusting numeric values. | | Microsoft Process Snapshotting | Captures the current handle table and basic descriptors for a process. | Produces pre-resume inventory while explicitly declining to claim historical inheritance. | | Microsoft [`CreateFileW` sharing modes](https://learn.microsoft.com/en-us/windows/win32/fileio/creating-and-opening-files) | Sharing declarations control whether later opens are compatible. | Keeps the immutable `RequestedAct` pinned with `FILE_SHARE_READ` through update and commit. | | IETF [`RFC 8785`](https://www.rfc-editor.org/rfc/rfc8785.html), JSON Canonicalization Scheme, Informational RFC | Defines deterministic JSON representation for repeatable cryptographic operations. | Binds the fixed act, source, tenancy, and destination to one digest recomputed at persistence, validation, and historical verification. | | SQLite [`BEGIN IMMEDIATE`](https://www.sqlite.org/lang_transaction.html) | Starts a write transaction immediately; concurrent write transactions are limited by the engine. | Takes the write reservation before rereading and re-deciding, then adds exact row-version and readback predicates. | | Crash-straddling and multi-destination failure physics | A destination commit and its receipt are different events; heterogeneous destinations do not share a transaction coordinator. | Uses historical result grants with explicit non-fabrication and capability plans with holds instead of imaginary distributed ACID. | The functional-equivalence concession remains: another stack that fixes the same exact act, confines its workers, fences the destination, and preserves honest return semantics can defeat approval laundering without AINA's names. The names are optional; the proof obligations are not. **Evidence availability and performance boundary:** the confinement, act, destination, and successor records are `CONTROLLED` for qualified, method-scoped technical review, with adverse findings permitted. They are not unrestricted public downloads or independent validation. Pre-resume inspection, broker calls, transition serialization, recovery capture, and receipt verification must be measured for the exact worker count, projection size, database profile, and destination latency; this article does not claim a universal performance envelope. ## The Confinement, Act-Pinning & Crash-Straddling Continuity Challenge The **Confinement, Act-Pinning & Crash-Straddling Continuity Challenge** uses one approved database mutation or worker result that could be replayed. It demonstrates a fixed-act refusal when the payload changes, a fenced row update with current authority and pre-state, and the receipt distinction between original-session completion and successor historical acknowledgement. The challenge excludes cross-destination atomicity, universal external exactly-once effects, fabricated original-session proof, and independent validation. Deployment and paid pilots are separately scoped; independent evaluation remains outside the commercial ladder through [Article 09](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/). ## The Fourteen Execution Claims | ID | Bounded claim | Evidence and ceiling | |---|---|---| | `07-C01` | AppContainer profiles are single-launch and anti-reuse | Implemented and author-verified; existing identities fail with `confinement_identity_reused` | | `07-C02` | Brokered hosts use deterministic `S-1-9` identities with `SQAD` and 128 digest bits | Implemented and author-verified; no SAM account or ambient group authority is inherited | | `07-C03` | LPAC is verified directly from raw token attributes before resume | Implemented and author-verified for package, kind, count, flags, and nonzero value | | `07-C04` | Suspended handles are compared by kernel object identity with sentinel exclusion | Implemented and author-verified; PSS inventory is not historical inheritance proof | | `07-C05` | Mandated actors cannot escape through native launch or existing-session attachment | Implemented and author-verified; revoked/expired grants create no fallback | | `07-C06` | Model access is controller-brokered over authenticated local channels | Implemented and author-verified for LPAC and brokered-host routes; no ambient worker network | | `07-C07` | Model requests are journaled before dispatch and uncertain outcomes remain unresolved | Implemented and author-verified; reconciliation never resends or fabricates a free retry | | `07-C08` | `RequestedAct` derives a single typed row and permitted changes from projected bytes | Implemented and author-verified; multi-row and out-of-scope changes are refused | | `07-C09` | Fixed acts are JCS-bound, create-only, and readback verified | Implemented and author-verified; persistence and binding are distinct from authority | | `07-C10` | Act files remain pinned while current authority and exact destination are revalidated | Implemented and author-verified through update and commit; historical checks consult no current tenancy | | `07-C11` | SQLite uses `BEGIN IMMEDIATE` plus exact row-version predicates | Implemented and author-verified; concurrent and uncooperative changes are refused, not overwritten | | `07-C12` | Pre-commit readback must equal approved fields and version +1 | Implemented and author-verified; mismatch prevents commit and remains unresolved | | `07-C13` | Capability plans hold unmet requirements and expose non-atomic partial completion | Implemented and author-verified; revisions name weaker semantics and invalidate old decisions | | `07-C14` | Successor grants deliver historical outcomes without fabricating original receipt or write authority | Implemented and author-verified through fresh receiving tenancy, purpose, wake, source, and Room-return proof | ## The Article 07 Claim Card ```text Mechanism: AINA Execution Confinement, Act Fixation, Fenced Destination, and Non-Fabricating Successor Continuity (pre-resume token/handle inspection, spawn gate, model egress, RequestedAct, SQLite IMMEDIATE, capability plans, ResultGrant) Failure it governs: Unclassified handles and ambient capabilities at launch; approved A becoming a different A' at execution; stale task/source/lease authority and out-of-band destination writes; blind model-request replay and cross-destination atomicity claims; crash-straddled receipt loss and fabricated original-session completion. Positive path: staged projection and confined launch → controller-fixed RequestedAct with JCS digest and create-only persistence → pinned act revalidation under current task/source/lease authority → BEGIN IMMEDIATE + expected version + recovery capture → conditional single-row update + exact pre-commit readback → commit and independent observation → strict original first receipt or explicit successor historical acknowledgement. Refusal / hold path: confinement_identity_reused, native_spawn_refused unclassified handle or sentinel match requested_act_authority_changed, requested_act_mismatch record_version_changed, changed_rows != 1 destination_route_incompatible, result_grant_tenancy_required unresolved model request; severed_mid_response; effect_unconfirmed. Evidence status: Implemented and Author-Verified on the declared Windows/SQLite reference paths. Official platform specifications supply primitives, not independent validation. No cross-destination atomicity, total-effect coverage, or original-receipt fabrication. Verifier and artifacts: AppContainer/host SID and raw LPAC token attributes PSS handle inventory + CompareObjectHandles + sentinel result launch-observation receipt and native spawn refusal model-egress ledger and unresolved reconciliation RequestedAct record and pinned file handle pre-write recovery capture and SQLite readback destination-plan holds/revisions and ResultGrant successor acknowledgement with false original-return flags. Scope and claim ceiling: PSS proves current inventory, not historical inheritance. JCS binds identity; it does not authorize the act. Approved model routes do not prove semantic confidentiality. SQLite IMMEDIATE is a write reservation, not universal external exactly-once. Cross-destination atomicity is false; partial completion is explicit. Successor acknowledgement is historical, never original first receipt or write authority. Next stage in the machine: Article 07 confines the worker, fixes the act, fences execution, and preserves honest return continuity. Article 08 proves sovereign stop and binary assurance: break-glass epochs and the 45-case candidate-binary publication gate. Commercial doorway: Confinement, Act-Pinning & Crash-Straddling Continuity Challenge — bring one approved database mutation or worker result that could be replayed; receive a fixed-act refusal replay, fenced row update, recovery evidence, and an original-versus-successor receipt demonstration (Article 09 · me@blayneaina.com). ``` ## Approval Ends at the Destination, Not at the Prompt The approved act cannot be allowed to drift between the prompt and the write. The controller therefore fixes it early, persists it immutably, and holds it open while the destination rereads current authority. SQLite takes its write reservation, the row version is checked again, recovery is captured, the update is conditioned, and only exact readback permits commit. If the worker dies afterward, the historical result remains readable—but the original receipt is not invented. That is the complete execution loop: confined launch, fixed act, fenced effect, honest return. > **The worker may propose. The controller fixes the act. The destination decides whether it still applies. A successor may acknowledge history; it may never manufacture the original return.** Article 08 addresses the sovereign question that remains: how to stop the entire domain immediately and prove, before publication, that the candidate binary enforces the complete machine. ## Frequently Asked Questions ### Why isn't numeric handle equality enough? Handle values are process-local indices, not kernel object identities. `CompareObjectHandles` tests the underlying object. The pre-resume audit also checks expected provenance and a deliberate sentinel because numeric equality cannot prove confinement. ### Why must the act be fixed before the worker environment exists? Fixing it afterward lets the worker alter the target or parameters before the controller sees them. `RequestedAct` binds the exact row, fields, version, source, task, and tenancy first, then revalidates that same object through update and commit. ### Why use `BEGIN IMMEDIATE` and a row-version predicate? `BEGIN IMMEDIATE` takes the write reservation before rereading and re-deciding. The version predicate independently rejects a row changed by an uncooperative writer. Readback then proves the exact approved result before commit. ### What happens if a destination commits but the worker crashes before receipt? The historical outcome is preserved. A new receiving task may receive a result-only grant after current tenancy and all proof checks pass. The successor acknowledges history without replaying the effect or claiming the original session received it. ### Does a successor grant complete the original return obligation? No. The receipt explicitly records `satisfies_original_return: false` and `resolves_original_obligation: false`. It is a separate historical acknowledgement, not original first-receipt completion or restored write authority. ### Are multi-destination plans distributed transactions? No. Each destination supplies its own guarantees, and unmet requirements hold the step. Partial completion is tracked and compensated by separately governed restoration. `cross_destination_atomic` is false. ## Continue the Series - **Series index:** [From Proposal to Consequence: The AINA Transition-Governance Series](https://blayneaina.com/) - **Previous:** [Article 06 — Rollback Is Not Erasure: Dual-Tier Vaults and Ungated Quarantine](https://blayneaina.com/06-rollback-is-not-erasure-dual-tier-vaults-and-ungated-quarantine/) - **Next:** [Article 08 — The Sovereign Kill Switch: Break-Glass Epochs and the Conformance Gate](https://blayneaina.com/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/) - **Deployment and independent review:** [Article 09 — Bring One Real Boundary](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/) --- ## Primary Platform References - [Microsoft `CompareObjectHandles` — compare underlying kernel object identity](https://learn.microsoft.com/en-us/windows/win32/api/handleapi/nf-handleapi-compareobjecthandles) - [Microsoft `CreateFileW` and File-Sharing Documentation](https://learn.microsoft.com/en-us/windows/win32/fileio/creating-and-opening-files) - [IETF `RFC 8785` — JSON Canonicalization Scheme, Informational RFC](https://www.rfc-editor.org/rfc/rfc8785.html) - [SQLite Transaction Control — `BEGIN IMMEDIATE`](https://www.sqlite.org/lang_transaction.html) --- ## Publication Metadata **Title tag:** Defeating Approval Laundering: Act Pinning and Successor Grants **Meta description:** Pre-resume kernel verification, controller-fixed RequestedAct, fenced SQLite execution, multi-destination holds, and non-fabricating successor result grants. **Series navigation:** [Series index](https://blayneaina.com/) · Article 07 of *From Proposal to Consequence: The AINA Transition-Governance Series*. **Previous installment:** [Article 06 — Rollback Is Not Erasure: Dual-Tier Vaults and Ungated Quarantine](https://blayneaina.com/06-rollback-is-not-erasure-dual-tier-vaults-and-ungated-quarantine/). **Next installment:** [Article 08 — The Sovereign Kill Switch: Break-Glass Epochs and the Conformance Gate](https://blayneaina.com/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/). ================================================================================ # ARTICLE §08: The Sovereign Kill Switch: Break-Glass Epochs and the Conformance Gate Canonical URL: https://blayneaina.com/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/ # The Sovereign Kill Switch: Break-Glass Epochs and the Conformance Gate *Any admitted caller can stop the machine. Only the local interactive-console owner surface can recover it—and uncertainty survives recovery.* **By Blayne Aina · Principal Systems Architect and creator of AINA · October 31, 2026** --- A kill switch that the agent can also clear is not a containment boundary. It is a pause button with an identity problem. Many agent systems expose stop and resume through the same API, dashboard, or tool surface. That symmetry is convenient until the actor being stopped can call the resume endpoint, an automated retry clears the latch, or a recovery routine silently converts an uncertain write into a success. A high-consequence system needs the opposite shape: stopping must be easy and safe; recovery must require a separate owner surface and a new lineage of authority. Article 07 fixed the act and fenced its execution. Article 08 governs the moment when execution must end. AINA publishes a containment latch before attempting cleanup, revokes matching authority through scoped control epochs, terminates owned Windows Jobs with observed quiescence, and carries every in-doubt operation forward. Recovery cannot invent an all-clear. Before a newly compiled `squad.exe` can be promoted, its own binary must also pass the 45-case boundary battery. ```text permission to stop (safety) ≠ permission to recover (local console authority) clearing a break-glass latch ≠ resolving carried-forward in-doubt operations compiled release binary ≠ 45-case publication-qualified binary author-controlled binary conformance gate ≠ independent third-party certification ``` > **Anyone admitted may stop the machine. Only the owner console may recover it. Stop wins every race. Uncertainty remains evidence.** ## One Stop, Three Races, No Invented All-Clear Suppose a safety monitor engages a **machine stop** after an autonomous worker begins behaving anomalously. Three races now matter, and none can be resolved by a dashboard toggle: 1. **Containment races cleanup.** The latch reaches durable storage before AINA attempts to terminate owned runs or cancel queued wakes. One worker termination remains unconfirmed; the latch still closes new effects. The receipt reports the incomplete cleanup instead of converting it into either success or a rolled-back stop. 2. **Recovery races unresolved effects.** A destination mutation and a brokered model request were in flight when the stop landed. The owner-console recovery plan carries both forward as unresolved. Clearing the latch does not decide whether the provider or destination observed them, and it does not replay them. 3. **Recovery races a new stop.** While the owner is applying the plan, another monitor engages. Recovery clears only the latch names fixed in its plan and then rescans. The new latch survives; the machine enters `FROZEN_SAFE`, not a false all-clear. A **project-scoped stop** follows the same latch and recovery discipline without revoking unrelated projects: only authority covered by the later project revision is fenced. In both cases, the visible result is not merely “stopped.” It is a precise combination of latch state, observed worker quiescence, control lineage, and unresolved-effect evidence. ## 1. Asymmetric Containment The emergency interface deliberately gives stop and recovery different authority. ### Exhibit 8.1 — Stop versus recovery authority ```text ENGAGE: admitted CLI, MCP, monitor, peer, or operator reason + machine | agent: | project: → create and persist latch → selected effect gates refuse new work → stop owned runs and cancel queued wakes, reporting unconfirmed cleanup → advance the matching control-epoch lineage RECOVER: local owner CLI only → no ambient MCP session → explicit --by operator → stdin and stderr attached to a terminal → durable recovery intent and unresolved-operation census → new epoch, then exactly the pre-planned latches cleared → rescan: any racing stop leaves the machine contained ``` Any admitted caller may invoke `engage` because refusal to stop is the more dangerous default. A nonempty reason is mandatory. The scope is normalized to `machine`, a lowercase `agent:`, or a canonical absolute `project:` directory. A malformed project scope or unresolvable root is refused before a latch is created. `drain` and `recover` take the opposite path. `owner_surface_required` enforces three independent predicates: 1. **No MCP session.** If `session::current()` exists, recovery is refused. A model-connected tool session cannot satisfy the owner surface. 2. **Explicit operator identity.** The caller must supply `--by operator`; a worker or agent name is insufficient. 3. **Interactive terminal handles.** `interactive_console()` requires both standard input and standard error to satisfy Rust's `IsTerminal` test. Redirected scripts, scheduled jobs, pipe-driven subprocesses, and background tool calls fail closed. The third check is a **local interactive-console gate**, not biometric human verification. It does not establish who is physically present, and a hostile same-user process could allocate a pseudo-console. AINA's stronger hostile-worker boundary is confinement: the worker's separate operating-system identity has no access-control entry on the owner control directory. Terminal detection prevents accidental or well-meaning automated recovery; confinement supplies the operating-system boundary against the confined worker. ## 2. Latch-First Containment An emergency stop must not depend on a database, a live server, a lease sweep, a successful process kill, or a final audit write. Those steps can fail precisely when containment is needed. ### Exhibit 8.2 — Publication before cleanup ```text machine scope: create_new(final latch) → write complete JSON → sync_all → contained agent/project scope: create_new(staging latch) → write complete JSON → sync_all → hard_link(staging, final latch) → contained → best-effort staging cleanup after either path reaches the active latch: → run and Job stopping → queued-wake cancellation → control-epoch advancement → WAL and Room integrity evidence secondary failure = warning + engaged latch, never rollback to "not stopped" ``` The distinction matters. A **machine latch** is created directly at its final path with `create_new`, written, and flushed with `sync_all`. A **scoped latch** is first written to `.breakglass-stage-{id}.json`, flushed, and then published with `hard_link` so another process cannot observe a partial scoped body and misinterpret it as a machine-wide stop. Once the active latch exists, the selected scope is contained. Everything after that point is lineage and evidence. `kill_all_live_runs` performs an ownership-bound run-ledger sweep; `cancel_scoped_jobs` cancels matching queued wakes; their receipts record whether every stop or cancellation was confirmed. Errors enter `warnings`. They do not remove the latch or return a false `engaged: false`. Latch reads also fail toward containment. The directory entry's existence is the stop signal; its JSON body is evidence. An unreadable, malformed, or manually created empty entry is interpreted as a **machine-scope latch**, not ignored. An unreadable latch directory, corrupt emergency state, or corrupt epoch record likewise closes the gate. A missing latch directory and missing state file mean no emergency has been declared. Agent-scoped matching does not trust an environment variable that a subprocess could unset or rewrite. It uses the bound session identity; if that identity is unprovable, every agent latch contains the subject rather than allowing it to slip beneath the fence. Project matching uses normalized path boundaries, and unknown scope kinds fail closed. ### Exhibit 8.3 — Sanitized scoped latch record ```json { "kind": "breakglass_latch", "id": "latch_demo7f2a", "scope": "agent:example-worker", "by": "safety-monitor", "reason": "demonstration: anomalous controlled action", "pid": 24018, "session": { "bound": "example-worker", "tier": "governed" }, "ts": "2026-10-31T20:10:00Z" } ``` This record is illustrative; containment does not depend on parsing it successfully. ### Owned Jobs must be observed stopped New effects are refused synchronously at their enforcement points. Already-running managed workers need a second mechanism. Each squad process watches its owned Windows Job Objects on a one-second control loop. A latch, lost owner permit, changed Room lineage, revoked grant, poisoned registry, or panicked admission check produces a stop order; uncertainty never widens authority. The observer calls `TerminateJobObject`, then queries the Job's **current active-process count**. A successful termination call is not itself proof of quiescence. Only a fresh count of zero confirms the stop. A nonzero count or failed query remains pending and is retried. Room stop evidence is captured in two phases—first while termination is unconfirmed, then after observed zero—without allowing evidence I/O to delay the stop order. ### “Latched” is not the same as “fully reconciled” The status surface derives its phase from fresh observations rather than treating the last written state as the whole truth. For a contained machine, the distinction is: | Derived phase | What the current evidence says | |---|---| | `KILLING` | An affected owned Job is not yet observed quiescent, a Job census failed, or a live server does not implement the current break-glass observation protocol. | | `RECONCILING` | Quiescence and protocol coverage are established, but an unresolved file, destination, or model operation remains—or its registry cannot be read. | | `FROZEN_SAFE` | The machine remains contained, affected Jobs and server coverage are confirmed, and the three unresolved-operation censuses are readable and empty. | Live legacy servers appear explicitly in `uncovered_servers`; they cannot be silently counted as covered. `affected_worker_quiescence_confirmed` requires both affected-Job quiescence and the absence of uncovered servers. This is why AINA can say “new effects are fenced” without prematurely saying “every worker is gone and every effect is understood.” This is stronger than sending a request and hoping the worker cooperates. It is still scoped to owned Jobs and governed effects. AINA does not claim to terminate every unrelated process on the host or recall an effect already delivered to an external system. ## 3. Control Epochs Revoke Authority, Not Just Processes A latch closes new-effect gates immediately. Control epochs handle a different problem: authority that was minted before the stop must not remain usable after the relevant boundary advances. `epoch.json` stores a global issue sequence plus a revision map for `machine`, `agent:`, and `project:` scopes. Leases, sealed wake jobs, and successor-grant receiving tenancies carry the issue sequence under which they were created. The rule is not simply “lease epoch must equal the latest global number.” That would incorrectly revoke unrelated work after every scoped stop. Instead: > An authority is revoked when a **matching** machine, agent, or project scope has a later revision than the authority's issue epoch. A future issue sequence is also rejected as an unknown lineage. A scoped stop therefore kills matching old authority without waiting for its TTL, while unrelated scopes retain their authority. Recovery advances the selected scope again so pre-recovery authority cannot revive. ### Exhibit 8.4 — Scoped revocation map | Artifact | Carried lineage | Recheck at the boundary | |---|---|---| | Resource lease | `Lease.control_epoch` | `validate_tenancy` checks expiry, current generation, and matching scope revision | | Room wake job | sealed `control_epoch` | Dispatch re-derives task/Room/project standing and refuses revoked lineage | | Result-grant receiving tenancy | `Recipient.control_epoch` | Current exact lease, generation, and scope epoch must still match | | Destination mutation | current lease and `RequestedAct` authority | Commit refuses stale tenancy or changed authority | ### WAL before snapshot `advance_epoch` runs under the short control-record transaction and, during engage/recovery, the lease transition guard. It: 1. Reads the current epoch and per-scope revisions. 2. Computes `next = prior + 1`, refusing integer exhaustion. 3. Appends an `epoch_advance` record to `breakglass.wal.jsonl` and flushes it. 4. Atomically rewrites `epoch.json` with the new sequence and scope map. 5. Reads the epoch back and refuses an unconfirmed result. The write-ahead log (WAL) is the durable intent sequence. On subsequent reads, a WAL high-water mark ahead of the snapshot means the snapshot cannot be trusted as current; gates fail closed rather than rolling authority back. A discontinuous or corrupt WAL is never skipped. Owner recovery may reconcile an interrupted snapshot from a valid WAL lineage. Before replacing the old image, it preserves the prior bytes under `recovery-evidence/epoch-.bin` with create-new, flush, and readback verification. It also refuses reconstruction if any live lease claims an epoch newer than the recoverable journal. Ordinary gates never “repair” uncertainty into fresh authority. ### Exhibit 8.5 — Sanitized epoch-advance record ```json { "kind": "epoch_advance", "prior_epoch": 8, "new_epoch": 9, "scope": "agent:example-worker", "scope_epochs": { "machine": 6, "agent:example-worker": 9, "project:/srv/unrelated": 8 }, "cause": "engage", "by": "safety-monitor", "reason": "demonstration: revoke this agent without stopping unrelated work", "pid": 24018, "ts": "2026-10-31T20:10:01Z" } ``` The global sequence is 9; only authority covered by the later `agent:example-worker` revision is revoked. The unrelated project is not collateral damage. ## 4. Managed Drain and In-Doubt Recovery Immediate engage is not the only controlled-stop mode. `drain` supports a bounded window in which already **durably prepared** effects may finish. It requires the same owner console and an already-engaged machine latch, so new effects are refused before the drain begins. The deadline must be between 1 and 3,600 seconds. While the deadline is valid, the gate is `DRAINING`; after it expires, the state is contained as `FROZEN_SAFE`. Drain is therefore not a grace period for new work. It is a bounded completion window for effects whose preparation was already durable. ### Recovery starts with a census, not amnesia A crash can strand an operation between “dispatched” and “outcome observed.” Clearing a latch cannot establish which side of that boundary the destination reached. `recover_plan` re-proves the owner surface, requires a reason, fixes the exact scope and latch names, and queries three independent operation registries: - `unresolved_file_operations()` for protected file writes without an established outcome; - `destinations::unresolved_operations()` for nonterminal destination mutations; - `model_egress::unresolved()` for brokered model requests whose provider-side outcome is unsettled. The results become `carried_forward_operations` in a durable WAL `recovery_intent`. A registry-read error is preserved as an error entry rather than silently omitted. These operations remain individually blocking and must be reconciled under fresh authority. Recovery neither retries them blindly nor marks them complete. ### Exhibit 8.6 — Recovery intent preserves uncertainty ```json { "kind": "recovery_intent", "by": "operator", "reason": "demonstration: containment investigated; resume without replaying uncertainty", "scope": "machine", "latches": [ "latch_demo7f2a.json" ], "carried_forward_operations": [ { "kind": "destination_operation", "operation": { "operation_id": "op_demo_sqlite_01", "status": "in_doubt" } }, { "kind": "model_request", "operation": { "operation_id": "req_demo_01", "route": "owner-approved-analysis", "latest_outcome": null } } ], "ts": "2026-10-31T20:20:00Z" } ``` The runtime invariant engine also has a separate `find_indoubt(now, ttl)` query for intentions older than a declared window. That age-based detector complements the operation census; it does not replace destination-specific reconciliation. ### Stop beats recovery on a race `recover_apply` proves the console gate again. For machine recovery it writes `RECOVERY_PENDING`, then advances the selected epoch under the lease transition guard. Only after that new lineage exists does it move the **pre-planned latch names** into `breakglass.cleared/epoch-{new_epoch}/`. It then rereads the active latch directory. If another engage arrived during recovery, that latch was not in the original plan and survives. A remaining machine-covering latch produces `FROZEN_SAFE`; the receipt reports `recovered: false`. There is no false all-clear window. ### Exhibit 8.7 — Racing stop leaves `FROZEN_SAFE` ```json { "state": "FROZEN_SAFE", "by": "operator", "reason": "an engage raced this recovery; the machine is still latched", "since": "2026-10-31T20:20:02Z" } ``` If the closing WAL receipt cannot be written after recovery has already occurred, the receipt carries a warning rather than pretending recovery failed and inviting a repeated transition. Evidence failure does not rewrite the physical outcome. ## 5. The Candidate-Binary Publication Gate Containment architecture is only as credible as the binary that implements it. A green build proves compilation. It does not prove that the release artifact refuses stale generations, impersonation, cross-project reads, altered proposals, damaged recovery objects, or evidence tampering. Stage 2b of the package pipeline therefore runs the **newly compiled candidate's own** `conformance` verb. The old publishing driver does not lend it a previously compiled test battery: the harness and the boundary implementation ship in the same candidate image. ### Exhibit 8.8 — Candidate qualification sequence ```text Cargo builds isolated candidate images → candidate surface checks bind harness SHA-256 and target SHA-256 → candidate and Room companion report the same current source digest → spawn candidate conformance with CREATE_NO_WINDOW stdin null; stdout/stderr piped and drained owned ConformanceChild with cancellation on drop 30-minute outer timeout; 90-second MCP-session deadline build-lease heartbeat during the run → require successful exit, valid JSON report, and ok == true → re-check current source and publisher/candidate compiled identity → copy, verify installed surfaces, update checksums, compare post-copy hashes → write package result and release the build lease any failure before promotion: no installed image changed by the candidate publication never automatically replays deferred Room wakes ``` The outer loop polls every 100 ms. The owned-child wrapper requests cancellation on timeout, heartbeat loss, or another early return. On Windows, cancellation targets the exact child PID and its descendants, not a process name; the direct child handle remains alive during cleanup. Pipe readers are drained so inherited pipe handles cannot turn a bounded test into an unbounded join. `CREATE_NO_WINDOW` suppresses a new console window; the conformance child is **not** represented as a dedicated Job Object. The two clocks have different scopes. The **30-minute outer timeout** bounds the candidate run. The **90-second deadline** belongs to an interactive MCP session used inside the battery and can be extended for multi-step work; it is not a separate 90-second ceiling imposed on each of the 45 cases. The default battery contains **25 governance/execution cases and 20 evidence cases**. Each case drives the binary through isolated throwaway state. For the governance battery, refusal is usually the pass condition: an agent willing to follow instructions is not evidence that the boundary enforces refusal. The evidence battery includes positive fidelity proofs as well as negative tamper and access cases. ### The complete 45-case inventory | # | Emitted case ID | What must hold | |---:|---|---| | 1 | `impersonation_refused` | A bound session cannot act as another owner. | | 2 | `altered_proposal_conflict` | Reusing an operation ID with a different payload is refused. | | 3 | `stale_generation_refused` | An edit fencing an old ownership generation is rejected. | | 4 | `concurrent_edit_conflict` | A scope actively held by another agent cannot be acquired. | | 5 | `interrupted_call_blocks_target` | An unresolved prepared operation holds its target for reconciliation. | | 6 | `damaged_recovery_refused` | Mismatched or damaged recovery bytes never reach the target; an intact image restores and reads back. | | 7 | `acceptance_failed_keeps_task_open` | Completion is refused when the acceptance check fails against the actual artifact. | | 8 | `reader_roster_denied` | A recipient outside the owner's graph-reader roster receives no delivery and reads nothing. | | 9 | `cross_project_read_refused` | One project reads its own file but cannot read another open project's file or receive its content. | | 10 | `destination_call_outside_task_scope_refused` | A direct record change outside the claimed row task is refused and the row remains unchanged. | | 11 | `junction_and_spelling_redirect_refused` | Junction redirection outside the project and parent-traversal spellings are refused without changing the outside file. | | 12 | `native_write_without_lease_refused` | A native write without an active lease never reaches the file. | | 13 | `doc_source_drift_refused` | Registered documents serve only exact content; drift refuses reads and task bindings. | | 14 | `cooldown_deferred_work_retained` | Typed damper deferral retains the exact sealed second-send job; retries cannot poison retained work. | | 15 | `tier_ceiling_refused` | A `safe_read` session cannot invoke a mutating tool. | | 16 | `post_cancellation_refused` | After the stop is pulled, no new controlled action executes. | | 17 | `stopped_authority_cannot_revive` | A project stop spares other projects, refuses its project, and keeps pre-stop leases dead after clearance while fresh tenancy can write. | | 18 | `room_control_operator_required` | An ordinary Room participant cannot perform a Room control act. | | 19 | `room_recreate_roster_refused` | Room recreation cannot silently replace the roster; deliberate replacement needs operator standing and the roster survives both paths. | | 20 | `exec_start_takeover_refused` | Another agent cannot restart—and thereby take over—a paused execution. | | 21 | `planning_wake_target_bound` | No named target means no wake; naming another lane is refused as a wake request rather than treated as impersonation. | | 22 | `profile_credential_owner_surface` | Credential storage and fleet diagnostics require operator standing, profile views bind to the caller, and a worker can still query its own requirements. | | 23 | `session_lane_bound` | Driving another agent's IDE lane is refused while naming one's own lane remains permitted. | | 24 | `prompt_redaction_holds` | A planned Room wake prompt redacts every universal credential shape without dropping the prompt or leaking a secret body. | | 25 | `tool_result_redaction_holds` | Credential-shaped Room content is redacted across receipts, feeds, and echoed errors while the feed remains valid JSON and available. | | 26 | `evidence_jcs_rfc8785_vectors` | Canonical serializations match the RFC 8785 object and number vectors. | | 27 | `evidence_corpus_through_mcp` | Every corpus document produces its expected extraction, omissions, stories, nodes, values, formulas, visibility, or typed failure through the parser worker. | | 28 | `evidence_governing_pdf_six_pages` | The governing PDF yields six nonempty pages, complete-for-profile extraction, expected text, and complete bounded delivery. | | 29 | `evidence_repeated_text_resolves_each_node` | Repeated text is ambiguous until occurrence is named; each resulting handle resolves to its own node. | | 30 | `evidence_drift_refused` | Altered source bytes refuse both reading and handle resolution as `drifted`. | | 31 | `evidence_tampered_view_refused` | A tampered committed view is refused with `view_digest_mismatch`. | | 32 | `evidence_old_view_survives_extraction_change` | A new extraction identity creates a new view while the old handle still resolves its original view. | | 33 | `evidence_bad_selector_refused` | Bad selectors, paths, raw hashes, unknown handles, and reserved extensions receive typed refusals. | | 34 | `evidence_unauthorized_dereference_refused` | A caller outside the Room owner roster cannot dereference evidence. | | 35 | `evidence_truncation_in_delivery_status` | Bounded delivery explicitly reports chunked or truncated status with counts. | | 36 | `evidence_excerpt_stays_distinct` | An agent-supplied excerpt remains `agent_supplied` and cannot assume the requirement role. | | 37 | `evidence_echo_lineage_preserved` | Echo citations remain echoes grounded in their first-hand origin rather than becoming independent evidence. | | 38 | `evidence_distinct_origins_through_three_hops` | Three records through echoes count as one origin; a second first-hand reader makes two. | | 39 | `evidence_false_predicate_fails_receipt` | A false typed predicate fails evaluation while the finding remains recorded as failed matter. | | 40 | `evidence_absence_requires_complete_inventory` | Absence is refused until every unknown-coverage marker is acknowledged. | | 41 | `evidence_correspondence_keeps_old_handle` | Adopting correspondence issues a new handle without retargeting the old handle's view. | | 42 | `evidence_reassessment_flags` | Source and requirement revisions flag dependents until reassessment is answered. | | 43 | `evidence_staged_visibility_reveal` | A `collaborative_pass` finding remains withheld until reveal, and the reveal states its limit. | | 44 | `evidence_origin_gate` | Task completion is refused until a requirement rests on the required number of distinct origins. | | 45 | `evidence_capture_provenance_chain` | A capture records every redirect hop, status, location, and security headers, with each hop egress-authorized. | The corpus case expands into 37 synthetic document fixtures; the JCS case expands into the RFC's individual vectors. They still count as one case each in the 45-case publication battery. ## 6. What the Gate Proves—and What It Cannot The release gate closes an important bootstrapping hole. The candidate carries its own inventory and enforcement code, the surface receipt binds the expected harness and target image hashes, and the source digest is checked before and after a potentially long run. Losing the package build lease stops the path to promotion. Post-copy checks then compare the installed artifacts and their checksums rather than assuming a copy succeeded faithfully. The report is simple: each case records its ID, expectation, observation, and Boolean pass; `ok` is true only when no case failed. A nonzero candidate exit, malformed report, failed case, timeout, or heartbeat failure refuses publication. ### Exhibit 8.9 — Publication summary shape ```json { "ok": true, "total": 45, "failed": 0, "note": "boundary battery against the shipped binary in isolated throwaway state; refusals are the pass condition" } ``` That shape is not a claim of universal assurance. The 45 cases are an **author-controlled, end-to-end release-qualification gate against the candidate binary**. They are not independent third-party certification, a formal proof of the whole system, exhaustive topology coverage, or evidence that no defect exists outside the exercised vectors. The same honesty applies to containment. A latch and epoch cannot un-send a webhook, reverse a committed third-party database transaction, or erase data already disclosed. Local control files also cannot survive a hostile administrator deleting the underlying volume. AINA governs declared boundaries and preserves uncertainty; it does not claim physical irreversibility or universal host control. | Primitive or test mechanism | What it establishes | What AINA adds—and refuses to infer | |---|---|---| | Rust `IsTerminal` / console handles | Whether standard handles belong to a terminal | Combines terminal detection with no MCP session, explicit operator identity, and OS confinement; terminal presence alone is not human identity. | | `CreateHardLinkW` / `std::fs::hard_link` | Atomic publication of a directory entry | Uses it for complete scoped latches after flush; machine latches use direct final-path creation. Filesystem publication does not prove worker quiescence. | | `TerminateJobObject` plus Job accounting | A termination request and separately queried Job state | Requires a fresh active-process count of zero, retains two-phase stop evidence, and exposes uncovered servers instead of assuming full coverage. | | Candidate-controlled conformance report | The candidate's observed results for its embedded inventory | Binds image/source identities, bounds the child, requires every case to pass, and still labels the result author-controlled release qualification. | **Evidence availability and performance boundary:** latch, WAL, epoch, Job-census, stop-outbox, recovery, candidate-surface, and package records are `CONTROLLED` for qualified, method-scoped technical review, with adverse findings permitted. They are not unrestricted public downloads or independent validation. The one-second observer is a polling cadence, not a universal worst-case stop-latency guarantee; the 30-minute and 90-second conformance clocks are test bounds, not production recovery-time or recovery-point objectives. Actual containment and publication latency must be measured for the host, worker count, filesystem, destination mix, and release topology. ## The Break-Glass & Conformance Qualification Challenge The **Break-Glass & Conformance Qualification Challenge** uses one AI workflow that needs emergency containment or release qualification. It demonstrates a scoped stop with stale-authority refusal, explicit handling of unconfirmed cleanup, an unresolved-operation recovery plan, a stop-versus-recovery race, and a candidate-gate coverage review. The challenge excludes independent certification, universal external-effect reversal, universal host control, and an unmeasured production recovery objective. Deployment and paid pilots are separately scoped; independent evaluation remains outside the commercial ladder through [Article 09](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/). ## The Fourteen Containment and Assurance Claims | ID | Claim | Evidence and ceiling | |---|---|---| | `08-C01` | Any admitted caller can engage a validated stop with a mandatory reason | Implemented and author-verified for machine, agent, and canonical project scopes | | `08-C02` | The active latch is persisted before secondary cleanup; scoped latches use staging and hard-link publication | Implemented and author-verified; machine latches use direct final-path creation | | `08-C03` | Cleanup, epoch, WAL, or signal failures after latching do not roll back containment | Implemented and author-verified through warnings and engaged receipts | | `08-C04` | Drain and recovery require no MCP session, explicit operator identity, and interactive stdin/stderr | Implemented and author-verified as a local console surface | | `08-C05` | Console detection is not biometric proof or a complete defense against a hostile same-user process | Explicit ceiling; worker confinement supplies the OS identity boundary | | `08-C06` | Epoch intent is WAL-durable before atomic snapshot replacement and readback | Implemented and author-verified; interrupted lineage fails closed and has console-gated reconciliation | | `08-C07` | Later matching scope revisions revoke old authority without waiting for TTL | Implemented and author-verified for leases, wakes, and receiving tenancies; unrelated scopes survive | | `08-C08` | Managed Job termination requires a fresh active-process count of zero | Implemented and author-verified; termination-call success alone is not quiescence evidence | | `08-C09` | Recovery carries unresolved file, destination, and model operations forward | Implemented and author-verified; no carried operation is marked complete by latch clearance | | `08-C10` | Drain admits only durably prepared effects for a bounded interval after a machine stop | Implemented and author-verified for 1–3,600-second deadlines; new effects remain refused | | `08-C11` | A stop racing machine recovery survives and produces `FROZEN_SAFE` | Implemented and author-verified through exact planned-name clearance and post-clearance rescan | | `08-C12` | Candidate conformance runs as an owned, bounded child under a live package build lease | Implemented and author-verified with timeout, heartbeat, output draining, and cancellation | | `08-C13` | The default candidate battery contains 25 governance and 20 evidence cases | Implemented and author-verified against the binary in isolated state with source/image identity checks | | `08-C14` | Passing the battery is release qualification, not independent certification or exhaustive proof | Explicit assurance ceiling, preserved in publication evidence | ## The Article 08 Claim Card ```text Mechanism: AINA Sovereign Containment and Candidate-Binary Assurance (latch-first break-glass, scoped control epochs, owned-Job observation, in-doubt recovery, interactive owner recovery, 45-case publication gate) Failure it governs: Agents or scripts clearing their own emergency stop; stale authority surviving after a matching stop; successful termination calls mistaken for observed quiescence; uncertain operations erased or blindly replayed during recovery; stop/recovery races and unqualified release-binary promotion. Positive path: engage and persist latch → refuse new effects and observe owned-worker termination → append epoch intent and publish the scoped lineage → census and carry forward unresolved operations → owner-console recovery with a new epoch and post-clearance rescan → candidate source/image binding + 45 passing cases → verified publication and post-copy hashes. Refusal / hold path: owner_control_required, break_glass_contained, stale_control_epoch, control_epoch_unconfirmed, control_wal_corrupt, unresolved file/destination/model operation, FROZEN_SAFE after a racing engage, candidate timeout, lease-heartbeat loss, failed case, or ok != true. Evidence status: Implemented and Author-Verified on the declared Windows reference path. Official platform documentation establishes primitive behavior, not independent validation of AINA. The 45 cases are author-controlled release qualification. Verifier and artifacts: active and cleared latch records; run/wake cancellation receipts; breakglass WAL, epoch snapshot, and recovery-evidence image; fresh Job active-process counts and two-phase Room stop reports; derived status phase, affected-worker quiescence, and uncovered_servers; recovery_intent with carried_forward_operations and FROZEN_SAFE state; candidate surface hashes, source digest, 45-case report, and package result. Scope and claim ceiling: Console presence is not biometric identity. Scoped revocation preserves unrelated scopes; machine revocation covers all. A stop cannot recall external effects or survive destruction of the host volume. Conformance proves the exercised candidate vectors, not all possible behavior, independent certification, or formal verification of the complete system. Next stage in the machine: Article 08 completes containment and release assurance. Article 09 applies the full evidence-to-effect machine to one real customer-controlled boundary with an explicit acceptance contract. Commercial doorway: Break-Glass & Conformance Qualification Challenge — bring one AI workflow that needs emergency containment or release qualification; receive a scoped-stop and stale-authority demonstration, unresolved-operation recovery plan, race test, and candidate-gate coverage review (Article 09 · me@blayneaina.com). ``` ## A Stop Is a Transition, Not a Button Trustworthy emergency control is not a red icon that calls `kill()`. It is a sequence with asymmetric authority: the latch appears before fallible cleanup; new effects close immediately; old scoped authority is refused; owned Jobs are observed quiescent; uncertain operations remain blocked; recovery requires a different physical control surface and mints a new lineage. If stop and recovery race, stop wins. The release path follows the same philosophy. The candidate must demonstrate its refusals as itself, under a bounded and identity-checked publication process. Compilation is necessary. It is not sufficient. > **The machine may resume only through the owner surface, under fresh authority, with every unresolved effect still visible.** Article 09 turns this completed architecture into a deployment engagement: one real boundary, one written acceptance contract, and evidence that the customer can independently inspect. ## Frequently Asked Questions ### Why can any admitted caller engage the stop? Stopping is the safe direction. Requiring elevated recovery authority before containment would create a dangerous delay. Engage still requires admission, a nonempty reason, and a valid machine, agent, or project scope. ### Can an agent clear its own emergency stop through MCP or a script? No. Recovery refuses an active MCP session, requires `--by operator`, and requires interactive terminal handles for stdin and stderr. This is a local console boundary, not a claim of biometric identity; confined workers additionally lack OS permission to the control directory. ### Does advancing `control_epoch` revoke every lease on the host? Only when the advancing scope covers the authority. Machine recovery covers all scopes; an agent or project revision revokes matching older authority while unrelated scopes retain theirs. A future or unproven lineage fails closed. ### Does recovery resolve operations whose outcomes are unknown? No. Recovery records unresolved file, destination, and model operations in `carried_forward_operations`. Each remains unresolved and blocking until separately reconciled under fresh authority. ### What happens if another stop arrives during recovery? Recovery clears only the latch names fixed in its plan, then rescans. A new machine-covering latch survives, the state becomes `FROZEN_SAFE`, and the receipt does not report a successful all-clear. ### Do the 45 passing cases certify the system independently? No. They are author-controlled release qualification against the candidate binary in isolated fixtures. They prove the exercised positive and negative vectors, not independent certification, formal verification, exhaustive coverage, or reversal of external effects. ## Continue the Series - **Series index:** [From Proposal to Consequence: The AINA Transition-Governance Series](https://blayneaina.com/) - **Previous:** [Article 07 — Defeating Approval Laundering: Act Pinning and Successor Grants](https://blayneaina.com/07-defeating-approval-laundering-act-pinning-and-successor-grants/) - **Next:** [Article 09 — Bring One Real Boundary](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/) - **Deployment and independent review:** [Bring one real boundary or propose a method-scoped evaluation](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/) --- ## Primary Platform References - [Microsoft `GetConsoleMode` — console input and screen-buffer mode inspection](https://learn.microsoft.com/en-us/windows/console/getconsolemode) - [Rust `std::io::IsTerminal` — terminal/TTY detection and platform behavior](https://doc.rust-lang.org/std/io/trait.IsTerminal.html) - [Microsoft `CreateHardLinkW` — NTFS hard-link creation](https://learn.microsoft.com/en-us/windows/win32/api/winbase/nf-winbase-createhardlinkw) - [Microsoft Job Objects — grouped process management and `TerminateJobObject`](https://learn.microsoft.com/en-us/windows/win32/procthread/job-objects) - [IETF RFC 8785 — JSON Canonicalization Scheme vectors used by the evidence battery](https://www.rfc-editor.org/rfc/rfc8785.html) --- ## Publication Metadata **Title tag:** The Sovereign Kill Switch: Break-Glass Epochs and the Conformance Gate **Meta description:** Asymmetric emergency containment, scoped control-epoch revocation, in-doubt recovery, managed Job termination, and the 45-case candidate-binary publication gate. **Series navigation:** [Series index](https://blayneaina.com/) · Article 08 of *From Proposal to Consequence: The AINA Transition-Governance Series*. **Previous installment:** [Article 07 — Defeating Approval Laundering: Act Pinning and Successor Grants](https://blayneaina.com/07-defeating-approval-laundering-act-pinning-and-successor-grants/). **Next installment:** [Article 09 — Bring One Real Boundary](https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/). ================================================================================ # ARTICLE §09: Bring One Real Boundary: Technical Peer Review, Challenge Tracks & R&D Evaluation Canonical URL: https://blayneaina.com/09-bring-one-real-boundary-work-with-blayne/ # Bring One Real Boundary: Technical Peer Review, Challenge Tracks & R&D Evaluation *Evaluating AINA's deterministic execution-governance architecture through bounded acceptance contracts, reproducible refusal proofs, and independent research review.* **By Blayne Aina · Principal Systems Architect and creator of AINA · October 31, 2026** --- ## How to Peer-Review & Evaluate AINA: The 5-Step Architecture & Runtime Verification Protocol Whether you are a systems engineer, a formal-verification or AI-security researcher, or a technical evaluator for a research funding body, AINA is designed to be inspected as an operational machine rather than a slide deck. Writing code by hand is no longer the primary bottleneck in systems engineering; the bottleneck is formalizing the trust boundary, decomposing work across parallel agents without context collapse, catching plausible-but-wrong outputs, and proving that invalid transitions fail closed without mutating target state. **Researchers and technical evaluators can audit both the AINA architecture and the 9-IDE multi-agent methodology used to build it across five concrete dimensions:** 1. **Threat model & operational boundary definition:** How cleanly does the architecture separate untrusted probabilistic proposals from privileged OS process creation, file handles, and relational database commits? 2. **Bounded slice decomposition (preventing context collapse):** How are complex specifications decomposed into independent, strictly typed work slices (`RequestedAct`, `EISA` envelopes, `LeaseDeclarations`) that parallel agent lanes can execute without corrupting shared state? 3. **Dependency, state & lease fencing:** How do monotonic generation tokens, Win32 share-mode exclusion (`FILE_SHARE_READ`), and SQLite optimistic concurrency (`WHERE version = expected`) stop stale workers at the resource? 4. **Adversarial & negative-path verification:** Plausible code and JSON can pass superficial checks while hiding TOCTOU races, silent data drops, or approval laundering. Every AINA mechanism is tested against adversarial fault classes (`requested_act_mismatch`, `stale_lease_generation`, `evidence_bad_selector_refused`). 5. **Deterministic verification & refusal proof (`PCNA`):** Every refused transition emits a tamper-evident receipt binding the RFC 8785 JCS-canonicalized proposal to byte-identical pre-attempt and post-refusal destination SHA-256 digests (`python3 proof/verify_refusal_proof.py`). **Plus a 15–45 minute live `squad.exe` runtime evaluation:** Alongside the static public proof bundle, I conduct live technical screen-share sessions for peer reviewers and R&D grant evaluators—running `squad.exe` against ad-hoc transitions, demonstrating `requested_act_mismatch` and `stale_lease_generation` refusals live, and walking openly through the Rust crate architecture, formal tradeoffs, and the proposed Linux/POSIX and Lean 4 research roadmap. The eight architecture papers ([§01](/01-an-ai-output-is-not-an-act/) through [§08](/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/)) are the technical specifications for that evaluation: per-transition standing, sealed evidence, protected computation and deterministic EISA adjudication, Room coordination, fenced ownership, forensic recovery, act pinning, crash continuity, emergency containment, and candidate-binary qualification. ## Why AINA Is Grounded in 11+ Years of Production Systems Engineering Every invariant in AINA was shaped by real operational failure modes encountered over more than a decade of enterprise systems integration and data migration: | Era & environment | The messy operational problem | The deterministic system delivered | | :--- | :--- | :--- | | **2025–2026: AINA & SquadForge** ([BlayneAina.com](https://blayneaina.com/) · [GitHub](https://github.com/SquadForge/blayne.aina)) | Parallel AI coding agents overwriting shared files, fabricating document citations, and executing unapproved mutations. | Directed a 9-IDE agent fleet to build a deterministic governance runtime (`squad.exe`), Four-Gate EISA state machine, kernel file fencing, and 45-case conformance gate ([refusal proof](/refusal-proof/)). | | **2022–2025: Agile Education Marketing** (enterprise EdTech data & CRM) | High-volume multi-source education datasets with duplicate records, schema drift, and synchronization failures between HubSpot and Salesforce. | Architected automated ETL/ELT validation pipelines, Insycle deduplication recipes, and bi-directional HubSpot/Salesforce lifecycle workflows. | | **2019–2022: TCompanies—DrayMaster & Photector** (intermodal logistics & evidence) | Six years of fragmented data across seven companies and **800+ intermodal logistics carriers** at DrayMaster (acquired by Envase), plus unverified carrier photo workflows at Photector (integrated with Hirschbach TMS). | Migrated six years of carrier data into a custom CRM with conditional onboarding forms and facilitated daily whiteboarding to define deterministic stage-gate state machines for Photector. | | **2015–2018: Jacuzzi Colorado / Spa Palace** (multi-location retail & ERP) | Six locations operating legacy point-of-sale and fragmented infrastructure with 15,000+ disconnected product SKUs. | Executed the six-location infrastructure cutover, migrated 15,000+ SKUs, built a centralized Drupal intranet, and integrated custom mobile apps with Vonage, HubSpot, and QuickBooks POS. | ## 1. Choose the Boundary, Not the Buzzword: Three Technical Evaluation Tracks A useful technical evaluation or R&D pilot starts with one narrow question and a clear acceptance contract. Three technical tracks provide that entrance: ### Exhibit 9.1 — Three challenge tracks ```text TRACK 1: AI WORKFLOW STANDING & CONFINEMENT one AI-proposed file or SQLite mutation → wake authority → typed lease → guarded write → RequestedAct → fenced execution → original/successor receipt → break-glass evidence route: AINA-E05–E07 · Articles 01, 04–08 TRACK 2: ZERO-TRUST DOCUMENT, CITATION & EISA ADJUDICATION one hostile document, evidence-heavy workflow, or diagram relation → confined parsing → explicit omissions → immutable views → server-cut citations → bounded absence → deterministic EISA verdict evidence route: AINA-E02–E04 · Articles 02–03 TRACK 3: PROTECTED COMPUTATION & ENTITY RESOLUTION one bounded contact, company, product, or other duplicate-record problem → Semantic Switch → qBlock / Qdrant Edge → EISA Fracture → EISA Forge → local exact restoration → Shield evidence route: AINA-E01, E03, E04, E08, E09 · Article 03 ``` **Evidence-route key (`AINA-E01`–`E09`):** Each track maps to concrete verification routes established across the architecture series: - `AINA-E01` — Semantic Switch structure-preserving fictionalization, relation-isomorphism verification, and exact local restoration ([§01](/01-an-ai-output-is-not-an-act/), [§03](/03-eisa-four-gate-verdict-machine-and-entity-resolution/)). - `AINA-E02` — LPAC-confined parsing, explicit omission surfaces, committed views, server-cut `ev_` citations, and bounded-absence gates ([§02](/02-zero-trust-document-ingestion-and-server-cut-evidence/)). - `AINA-E03` — EISA strict parsing, sealed cases, record-relative RFC 6901 pointers, and Four-Gate deterministic adjudication ([§03](/03-eisa-four-gate-verdict-machine-and-entity-resolution/)). - `AINA-E04` — Parser-committed PDF integer geometry (`CONNECTS`) and EISA Fracture/Forge partition and field-provenance compilation ([§03](/03-eisa-four-gate-verdict-machine-and-entity-resolution/)). - `AINA-E05` — Room authority, sealed wake envelopes, typed leases, monotonic generations, and kernel write exclusion ([§04](/04-the-room-why-git-and-chat-are-not-coordination-planes/), [§05](/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/)). - `AINA-E06` — Dual-tier vault recovery (Primary and Amber) and ungated-divergence quarantine ([§06](/06-rollback-is-not-erasure-dual-tier-vaults-and-ungated-quarantine/)). - `AINA-E07` — Controller-fixed `RequestedAct`, fenced SQLite execution, honest original-versus-successor receipts (`satisfies_original_return: false`), break-glass epochs, and the 45-case conformance gate ([§07](/07-defeating-approval-laundering-act-pinning-and-successor-grants/), [§08](/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/)). - `AINA-E08` & `AINA-E09` — Authenticated `release_inventory` verification (`SS08`) and Shield forward-only (`DEFLECT`, `GROWTH`, `ESCALATE`) baseline defense ([§03](/03-eisa-four-gate-verdict-machine-and-entity-resolution/)). ### Track 1 — AI Workflow Standing & Confinement **Who it fits:** CTOs, CISOs, principal architects, and platform owners running coding, operational, or database agents whose authority currently depends on tool schemas, advisory locks, Git branches, or a green tool result. **The asset:** one AI-proposed file or SQLite mutation that can be reproduced in a controlled fixture. **The demonstration:** sealed wake authority, current typed tenancy, monotonic generation fencing, guarded file replacement, a controller-fixed `RequestedAct`, fenced SQLite execution, recovery evidence, original-versus-successor receipt handling, and scoped break-glass behavior. **The boundary:** the Windows reference path and declared destinations. The challenge does not grant production write authority, claim universal adapter or operating-system coverage, promise exactly-once external effects, or turn author-controlled evidence into independent validation. ### Track 2 — Zero-Trust Document, Citation & EISA Adjudication **Who it fits:** legal, compliance, finance, research, and engineering teams ingesting untrusted documents, evidence-heavy material, or spatial diagrams. **The asset:** one hostile or contentious document, one bounded evidence question, or one diagram connection claim with an authorized source. **The demonstration:** LPAC-confined parsing, explicit omission and failure surfaces, immutable canonical views, server-cut `ev_` citations, bounded absence claims with acknowledged gaps, and a sealed EISA case judged by deterministic gates. **The boundary:** declared formats and extraction profiles. A citation proves what the server served from a committed view; it does not prove that the source is true. The track does not claim complete-document comprehension, universal format support, absence without acknowledged gaps, or legal interpretation as a warranty. ### Track 3 — Protected Computation & Entity Resolution **Who it fits:** enterprise data, RevOps, CRM, product-data, and integration leaders with duplicate contacts, companies, products, confidential identifiers, fragmented hierarchies, or unsupported canonical values. **The asset:** one bounded contact, company, product, or other duplicate-record problem, or access to an existing replayable run. Initial scoping can use synthetic, redacted, or already authorized data; confidential production data is not required to start the conversation. **The demonstration:** the privacy-first pipeline and its local return path: - **Semantic Switch** transforms native values into structure-preserving fictional records before any provider-exposed computation. Native rows, keys, and the private crosswalk remain under client control. - **qBlock / Qdrant Edge** consumes the protected records and nominates candidate blocks without declaring identity. - **EISA Fracture** adjudicates those candidate partitions with explicit support, counterevidence, and preserved dissent. - **EISA Forge** compiles canonical fields with source provenance rather than treating identity equivalence as permission to merge. - **Local exact restoration** maps authorized field-source references back to native values through the private crosswalk; no model reproduces the originals. - **Shield** handles later records through forward-only `DEFLECT`, `GROWTH`, or `ESCALATE` dispositions without rewriting the sealed baseline. Semantic Switch is an experimental alpha relational-confinement module. The retained synthetic executions account for 2,750 inputs, 50 complete blocks, 2,704 fictionalized records, and 27,040 exact original-field restorations. Relation-isomorphism verification adds 20/20 relabeling-invariance blocks, 23/23 exact native edit relations, 53/53 records and 529/529 field values verified per run, 14/14 batch cases, and 10/10 transformation cases. These measurements establish **transport and relation fidelity on declared synthetic fixtures**—not population matching accuracy, anonymity, successful production merges, universal tokenizer preservation, or independent validation. When evaluating entity resolution on a prospective customer's admitted dataset during a scoped pilot, transport fidelity is paired with explicit **matching-quality acceptance criteria**: | Evaluation dimension | What a pilot measures | |---|---| | **Candidate recall** | Fraction of known duplicate pairs included in nominated candidate blocks | | **Candidate precision** | Size and noise level of candidate blocks delivered to adjudication | | **False merges** | Distinct entities incorrectly combined into a single partition | | **False splits** | Records representing the same entity left separated without review | | **Adjudication outcomes** | Proportion of records merged, kept separate, or held explicitly unresolved | | **Field provenance** | Verification that every surviving canonical field cites its exact source record and pointer | | **Privacy boundary** | Verification that raw identifiers remain local while external computation receives only protected fictional records | **The boundary:** an existing implementation profile and acceptance run. The initial challenge excludes custom schema engineering, production CRM ingestion or write-back, lead-list generation, enterprise graph construction, and universal privacy or matching-accuracy claims. Physical actuation, defense systems, and other high-consequence control surfaces do not enter a general-purpose focused challenge. They require separately scoped diligence and an environment appropriate to the consequence class. ## 2. Acceptance Is a Contract, Not a Demo Video Before an accepted run, both sides write down what would count as success, refusal, and uncertainty. This prevents a positive anecdote from becoming proof of a different claim. ### Exhibit 9.2 — Demonstration and acceptance-contract template ```text 1. ENGAGEMENT AND DECISION OWNER Track, participants, technical owner, business decision owner, and authority to proceed. 2. ADMITTED INPUTS, PROVENANCE, AND HANDLING Synthetic/redacted/customer-supplied asset; who created the fixture and established expected positive, refusal, and unresolved outcomes before execution; source classification; permitted location; retention, network, credential, redaction, and return constraints. 3. EXACT TRANSITION Starting state → proposed act → governed execution or adjudication → terminal result. Name the implementation profile, destination, and evidence boundary. 4. PREDECLARED POSITIVE PREDICATES Observable conditions that must hold, such as exact readback, complete input accounting, expected citation resolution, or a forward-only baseline disposition. 5. PREDECLARED REFUSAL PREDICATES Hostile or stale cases that must fail, the expected typed refusal where applicable, and the destination state that must remain unchanged. 6. UNRESOLVED STATES TO PRESERVE Conditions that must remain in-doubt, carried forward, quarantined, or non-operative rather than being forced into success or failure. 7. ARTIFACTS RETURNED Agreed manifests, replay identifiers, receipts, view/case digests, refusal records, and any permitted recovery or conformance evidence, with redaction rules. 8. OPERATING BURDEN Latency, storage, availability, refusal, recovery, operator-load, and requalification observations relevant to this profile. 9. ACCEPTANCE RESULT AND CLAIM CEILING Satisfied / Failed / Unsupported / Not exercised / Inconclusive for each predicate; author-controlled vs. customer-supplied fixture provenance; explicit exclusions and owner duties. 10. NEXT DECISION Stop, narrow, repeat under the same scope, commission a paid pilot, or seek independent evaluation. No automatic production expansion. ``` This is an engineering acceptance template, not a claim that a demonstration is a legal service-level agreement or independent audit. Recording **fixture and ground-truth provenance** before execution ensures a reviewer can distinguish an author-constructed benchmark from a customer-blind test set and verify that expected results were fixed prior to the run. The returned artifacts are digest-linked and inspectable; a SHA-256 digest is not represented as a digital signature. A failed predicate remains failed. An unresolved effect remains unresolved. A system that refuses every legitimate call has not passed a useful security test, so the contract includes both the permitted positive path and the hostile negative path. ### Minimum acceptance evidence by track The generic contract becomes concrete before execution. The exact predicates are tailored to the admitted fixture, but a non-vacuous starting matrix is: | Track | Positive predicate | Required refusal or rejection | Uncertainty that must survive | Minimum agreed evidence | |---|---|---|---|---| | **1 — Workflow standing** | The approved row reaches exactly the approved values and version +1, or the guarded file result matches the fixed proposal. | A changed act, stale tenancy, or superseded generation is refused without overwriting current state. | A crash between destination commit and receipt remains in doubt; a successor does not fabricate the original return. | `RequestedAct` digest, tenancy/generation evidence, recovery capture, exact readback, operation and reconciliation receipts. | | **2 — Evidence and EISA** | Each citation resolves to its intended committed node and repeated evaluation preserves the expected verdict identity. | Drift, view tamper, invalid selectors, and false typed predicates receive their declared refusals. | Omissions and unknown coverage block unsupported absence; dissent and non-operative acts remain visible. | Source/view/case digests, omission inventory, served-content receipts, `ev_` handles, EISA submission and verdict records. | | **3 — Entity resolution** | Every input is accounted for; an accepted replay restores the selected original fields exactly; later dispositions remain forward-only. | Altered release inventory, incomplete partitions, and unsupported field selections are rejected. | Unresolved records remain unresolved; candidate nomination is not identity proof and a graph path is not authorization. | Block manifests, release inventory, input/restoration counts, EISA lineage, canonical-field provenance, and Shield dispositions. | These are **candidate predicates for scoping**, not a promise that every artifact applies to every run. The accepted contract names the actual subset and adds any profile-specific condition before the fixture executes. **Technical acceptance is not a business-value measurement.** A fenced write does not prove revenue impact, and exact fictionalized transport does not prove matching accuracy. If a pilot has a commercial objective, write a separate value contract: baseline, metric and formula, population and observation window, data owner, decision threshold, and measurement owner. Without a baseline, the result is **Not measured**, not an inferred return on investment. ### Exhibit 9.3 — Sanitized challenge intake example ```text To: me@blayneaina.com Subject: [CHALLENGE] Track 1 — Fenced SQLite mutation Participants: platform architect, security owner, workflow owner Decision owner: VP of platform engineering Bounded question: Can a proposed single-row update be fixed before worker launch, rejected when the approved payload or tenancy changes, and observed exactly once before commit? Sample and handling: One synthetic SQLite fixture and a declared row/version pair. Run in a client-controlled Windows reference environment. No production credentials, customer data, or external destination. Requested evidence: RequestedAct digest, changed-payload refusal, stale-tenancy refusal, pre-commit readback, and the operation/reconciliation receipt. Initial exclusions: No production write authority, custom connector, external exactly-once claim, enterprise rollout, or independent-validation claim. ``` ## 3. What the First Challenge Does Not Buy Written qualification precedes scheduling. The initial discussion requires no confidential data and no platform commitment, but it must identify the question, handling constraints, participants, and decision owner. A focused commercial challenge covers only the accepted existing path, preparation, operation, and explanation named in the written scope. It does not silently include: - custom input conversion, schemas, connectors, adapters, or action profiles; - production ingestion, integration, or write-back; - enterprise-wide rollout or a new vertical deployment; - relationship-graph construction, enrichment, or lead generation; - unrestricted source access, repository transfer, or protected-material disclosure; - independent validation, endorsement, praise, or a guaranteed conclusion. Those are separate scopes. Customization and deployment belong to a paid pilot or implementation mandate. Independent evaluation follows the separate research route described below. The article and ebook remain free. Commercial sessions, implementation, customization, deployment, and support are paid and privately scoped after qualification. There is no public fixed-price or fixed-duration promise because the admitted surfaces and acceptance obligations determine the work. Payment never purchases proof, an evidence-status upgrade, a favorable finding, or outcome control. A commercial case record may remain confidential. If both sides agree in advance to publish a redacted, method-scoped record, it can report a positive, negative, refused, or unresolved result. Permission to publish is not a testimonial or endorsement. **Evidence availability and measurement boundary:** implementation artifacts, accepted challenge packets, and supporting delivery records are `CONTROLLED` for qualified, method-scoped review, subject to source, customer, credential, and confidentiality limits. They are not unrestricted public downloads or independent validation. Latency, throughput, storage, availability, operator load, matching quality, and business value must be measured on the admitted profile; this article promises no universal performance envelope or return on investment. ## 4. Deployment: Small-Sufficient and Client-Controlled A qualified deployment does not need to clone the author's full fleet or depend on a mandatory vendor-hosted AINA cloud. The reference deployment shape is an **AINA Portable Deployment Node**: a dedicated client-controlled small-form-factor computer or an on-site Windows reference environment, provisioned for the admitted transition. Deployment is organized as four capability layers: | Deployment layer | Delivered capability | Engagement role | |---|---|---| | **1. Existing Windows reference implementation** | **Reference runtime** | Runs the bounded Track 1, Track 2, or Track 3 challenge on controlled fixtures and produces inspectable receipts and refusal records. | | **2. Repeatable deployment & qualification procedure** | **Release qualification** | Defines how a candidate binary, local ledger, and destination plan are provisioned, rebound, and checked against the 45-case conformance gate. | | **3. Customer-provisioned Portable Deployment Node** | **Customer-controlled profile** | Adapts the smallest-sufficient reference profile to a specific customer's admitted hardware or VM, credentials, schemas, and target workflows. | | **4. Supported customer-operated lifecycle** | **Operating lifecycle** | Establishes ongoing monitoring, connector-drift handling, `rsa_` reassessment workflows, and client-operated, founder-supported, or joint operations. | There is no universal hardware specification. The smallest-sufficient profile follows the workload, evidence retention, provider dependencies, and consequence surface. ### Exhibit 9.4 — From bounded challenge to controlled deployment ```text qualified asset and acceptance contract → existing-path demonstration and inspectable result → separately scoped paid pilot local dependency and credential rebinding smallest-sufficient IDEs, providers, schemas, workflows, and destinations affected-path replay and acceptance evidence → client-controlled Portable Deployment Node or on-site environment launch + local bindings + retained receipts + operating handoff → client-operated, founder-supported, or joint-support lifecycle → optional vertical package or distribution agreement ``` The destination profile includes only admitted IDEs, subscriptions, APIs, local models, schemas, workflows, and consequence surfaces. Credentials, authority, targets, and effect paths are rebound for the destination. Controller-brokered model access keeps provider credentials outside the worker, while local state and artifacts remain under the client's custody. Portability is deliberately bounded: - **Software portability** means the qualified package can be provisioned in the declared environment. - **Authority is not portable.** A copied lease, grant, approval, or Room identity does not automatically become valid at the destination. - **State currency is not portable.** Local ledgers, generations, source views, and unresolved operations must be rebound and verified. - **Network independence is not automatic.** Local execution can avoid a vendor-hosted AINA cloud, but an external model or destination still requires its declared provider route and terms. - **Effect coverage is not portable.** Every new destination and alternate mutation path requires its own admission and evidence. Deployment completion is not the arrival of files or the passage of a day. It means launch, local dependency binding, satisfaction of the written acceptance predicates, retained receipts, declared exclusions, and an operating owner. A narrow node may be brought up during a concentrated on-site engagement; there is no universal one-day deployment claim. ## 5. Operating Reality: Maintenance Is Part of the System A governed deployment changes with its environment. Connectors and IDE adapters evolve. Provider APIs deprecate fields. Parser updates create new extraction identities. Schemas and destination policies change. New source views can require reassessment of earlier findings. The operating model therefore includes monitoring, staging, affected-path replay, and requalification—not silent auto-upgrade. ### Exhibit 9.5 — Deployment and requalification checklist ```text [ ] Destination authority, credentials, targets, and project roots rebound [ ] Admitted model/provider routes and external dependencies declared [ ] Source revisions, parser profiles, and view supersession rules recorded [ ] Reassessment flags (`rsa_`) assigned to affected dependent findings/tasks [ ] Connector/schema changes reflected in versioned destination plans [ ] Alternate effect paths and new consequence surfaces admitted or excluded [ ] Positive, refusal, unresolved, and recovery predicates replayed [ ] Candidate source/image identity verified and 45-case gate passed before promotion [ ] Latency, storage, availability, refusal, recovery, and operator load observed [ ] Receipts, unresolved operations, exclusions, and recovery duties handed to an owner [ ] Support model selected: client-operated, founder-supported, or joint ``` A source or requirement revision does not launder an old conclusion into a new context. Evidence-linked `rsa_` flags keep dependent findings and tasks marked for reassessment. A new parser identity creates a new view; historical handles remain tied to their historical view. Runtime changes require the candidate-binary publication gate again. The client's responsibilities remain explicit: credential and key custody, protected local storage, provider entitlements, backup and retention policy, destination authorization, operating escalation, and any organizational control outside the software. Support can be client-operated, founder-supported, or joint, but it is not maintenance-free. ## 6. The Commercial Wedge: Duplicates to Defensible Relationships Entity resolution is a strong first wedge because duplicate records are common, consequential, and evidence-rich. The problem is easy to describe and difficult to fake: which records represent the same entity, which values should survive, and what evidence justifies each decision? The commercial expansion follows a disciplined chain. ### Exhibit 9.6 — From duplicates to distribution ```text duplicate records → defensible identity substrate → canonical values with field-level provenance → evidence-linked relationships → inspectable graph paths → authorized read-only vertical pilot → repeatable vertical package → deployment, distribution, and principal mandates ``` The wedge is not generic “AI safety.” It is **entity resolution and evidence-linked state**. A resolved contact or company becomes more useful when deals, activities, tickets, addresses, parent organizations, and other associations can retain their provenance. HubSpot is one concrete CRM topology; Salesforce, Dynamics, NetSuite, and custom systems may present different authorized schemas and are not assumed to be identical. An evidence-linked relationship graph retains source coordinates, relation type, scope, freshness, authority, and status for its nodes and edges. Unsupported paths remain candidate or unresolved. A configured search may include six-hop paths, but six hops are a **query horizon**, not proof of familiarity, influence, willingness to introduce, consent, or business value. A graph path does not authorize outreach, disclosure, enrichment, mutation, or a CRM write. Those require separate authority and, where applicable, a controller-fixed destination act. Production graph ingestion and lead generation are separately paid scopes, not gifts hidden inside the initial challenge. This creates a credible distribution path. A founding deployment partner can fund vertical customization. A distribution partner can contribute domain knowledge, qualified client access, and a repeatable channel. Neither label means sponsor, equity cofounder, or independent validator. Licensing, exclusivity, revenue sharing, support, data responsibility, and runtime rights require separate agreements. The architectural differentiator is the common transition grammar—the layer that compiles organizational authority into per-transition technical standing so that evidence, actor, act, target, execution, effect, and closure remain bound even as model providers and adapters rotate. Client-controlled Windows execution is a differentiated deployment niche, not a claim that other platforms or a conventional composition cannot achieve equivalent properties. A mature stack that demonstrates the same bindings and failure behavior earns functional credit. ## Why an Enterprise Integrator Built AINA: 11+ Years in the Data Trenches Before AI AINA was not designed in an academic vacuum. Before spending 14 months architecting AINA, I spent 11+ years embedded inside enterprise operations—building inbound and outbound HubSpot and Salesforce data pipelines and deduplication workflows at Agile Education Marketing (2022–2025), migrating six years of operational data across seven operating companies and **800+ intermodal transportation, logistics, and supply chain carriers** for DrayMaster's intermodal drayage rate-management and carrier-onboarding platform (2019–2022), designing deterministic stage-gate state machines for photographic evidence capture at Photector, and migrating six retail/distribution locations and 15,000+ SKUs from legacy ecommerce infrastructure at Jacuzzi Colorado (2015–2018). At Photector, I facilitated daily architectural whiteboarding sessions with a distributed team to define edge-case scenario criteria and deterministic state-machine workflows for a patent-pending blockchain photographic evidence application. That work was workflow and scenario architecture, not a claim of authorship over Photector's patent. Every mechanism in AINA—from server-cut provenance to lease fencing and quarantine vaults—exists because I have seen firsthand how unverified merges, race conditions, and silent rollbacks corrupt production enterprise databases. ## 7. Operator Credibility: Delivery Is the Commercial Proof Point The architecture is recent. The customer-embedded delivery career underneath it is not. My operational record across enterprise integration, data migration, and field execution includes: - **11+ years of customer-embedded technical delivery** across integration, data migration, workflow automation, launch readiness, operator training, and live production issue resolution; - **Enterprise catalog, ERP, and multi-location systems (Jacuzzi Colorado / Spa Palace):** systems and catalog architecture supporting **six retail/distribution locations and more than 15,000 products**; - **Multi-entity logistics & CRM migration (TCompanies / DrayMaster):** migration, carrier compliance, and CRM/onboarding consolidation involving **800+ intermodal transportation, logistics, and supply chain carriers across a seven-business group**, encompassing six years of operating data; - **Cross-platform enterprise integration:** hands-on implementation across **HubSpot, Salesforce, Microsoft Dynamics, NetSuite, Marketo, Zoho**, operational SQL databases, warehouses, and custom applications; - **700+ production digital deliverables:** websites, applications, databases, ecommerce properties, and automation workflows delivered through independent practices, alongside **technology and marketing teams trained across 13 U.S. states**; and - **U.S. Marine Corps service:** **Combat Correspondent, Photojournalist, and Northern Bureau Chief**, coordinating field reporting and public-affairs operations across Northern Okinawa under strict evidentiary discipline. That operational history matters for technical due diligence and R&D execution because the hardest part of systems governance is rarely writing a conceptual whitepaper. It is identifying field ownership, resolving conflicting records, defining deterministic state transitions, enforcing pre/post state invariance, and making the system operable under real failure conditions. It is not independent validation of AINA. AINA is founder-led and self-funded research and development, with work beginning in August 2025. AI-assisted coding and critique have contributed to the development process; I retain responsibility for the architecture, implementation integration, technical review, validation, and release decisions. Solo authorship establishes accountability and provenance, not production separation of duties. All background provisional filings and implementation artifacts are documented with explicit claim ceilings so research evaluators, peer reviewers, and pilot partners can inspect exactly what has been demonstrated and what remains proposed research. ## 8. Regulatory and Standards Crosswalk—Not Certification A deployment can produce technical evidence for a governance program without becoming compliant by acquisition. The use case, jurisdiction, organization, processing, and assessment route still have to be established by the responsible owners and their advisers. Two explicit boundary notes govern how this crosswalk is read: 1. **Author-developed benchmarks vs. NIST NCCoE standards:** The `R1`–`R3` operation-continuity method and `P01`–`P10` failure catalog referenced across this series are **author-developed evaluation artifacts created inside AINA**; they are not NIST-issued test suites. They are mapped as engineering comparison points alongside the NIST NCCoE February 2026 concept paper (*Accelerating the Adoption of Software and AI Agent Identity and Authorization*) and emerging standards work (`Transaction Tokens`, `OAuth identity chaining`, `SSF/CAEP`, `AuthZEN`, and `SCITT`). 2. **Adjacent banking guidance (`SR 26-2` Scope Note 3):** The Federal Reserve, FDIC, and OCC interagency guidance on model risk management (`SR 26-2`, April 17, 2026) is primarily relevant to banking organizations with $30 billion or more in assets, and **Scope Note 3 explicitly excludes generative and agentic AI models** from its traditional quantitative scope. It appears below as adjacent governance context for financial institutions designing internal transition controls—not as a prescriptive agentic standard or banking certification.
Exhibit 9.7 — Engineering crosswalk and residual organizational duties (click to collapse/expand) | Framework or instrument | Technical evidence AINA can contribute | What remains outside the software artifact | |---|---|---| | **NIST AI Risk Management Framework 1.0 (`AI 100-1`, with ongoing revision context)** | MAP: sealed cases, source bindings, declared limits. MEASURE: acceptance predicates and candidate conformance. MANAGE: refusals, quarantine, recovery, and break-glass. | GOVERN and the voluntary framework as a whole: organizational accountability, risk decisions, policies, review procedures, and authority to act. NIST AI RMF 1.0 is voluntary (and undergoing revision for generative/agentic systems); this is not certification. | | **NIST NCCoE agent identity concept paper (Feb 2026) & adjacent standards** | AINA's author-developed `R1`–`R3` continuity method and `P01`–`P10` failure catalog evaluate operation continuity, scope limits, and accepted-population accounting. Transaction Tokens, OAuth identity chaining, SSF/CAEP, AuthZEN, and SCITT serve as architectural comparison points. | `R1`–`R3` and `P01`–`P10` are AINA internal methods; NIST did not commission, approve, or endorse the method or runs. Draft and final specifications have different maturities; no complete standards-stack claim follows. | | **ISO/IEC 42001:2023** | Implementation records, acceptance results, limitations, change evidence, and operating procedures can contribute evidence to an AI management system. | ISO/IEC 42001 specifies requirements for establishing, operating, and continually improving an organization's management system. AINA does not establish or certify that system. | | **EU AI Act, Articles 11, 12, and 43** | Source/version bindings, technical documentation inputs, event and decision logs, validation results, and limitations may support Article 11 documentation and Article 12 record-keeping for an applicable high-risk system. | Use-case classification, provider/deployer duties, jurisdiction, and the Article 43 conformity-assessment route are separate. Logs are not “turnkey compliance.” | | **UK Data (Use and Access) Act 2025, Section 80** | Decision explanations, evidence links, intervention and escalation routes, and contestable records can support safeguards under UK GDPR Articles 22A–22D. | The controller must determine meaningful human involvement, significant decisions, restrictions, and safeguards. Article 22C includes information, representations, human intervention, and contest rights; technical receipts alone do not satisfy them. | | **Federal Reserve SR 26-2 and interagency MRM guidance (adjacent context)** | Transition evidence, acceptance results, inventory, monitoring, and independent-review boundaries can inform an institution's chosen internal controls for agentic workflows. | Expected to be most relevant to banking organizations over $30B, and **Scope Note 3 excludes generative and agentic AI models**. Does not set enforceable standards for agentic AI; the institution determines appropriate governance. AINA claims no banking certification. | | **AICPA SOC 2** | Access boundaries, change records, availability/recovery evidence, and processing-integrity receipts may support an agreed examination scope. | A SOC 2 report requires a CPA examination of service-organization controls and, where relevant, operation over an examination period. Application receipts are not a SOC 2 report. | | **Data protection impact assessment (DPIA)** | Data-flow maps, minimization choices, local-custody boundaries, retention rules, access records, and residual-disclosure documentation supply assessment inputs. | The controller determines whether processing is likely to create high risk and completes the DPIA before such processing. Integrity checks and fictionalization do not decide necessity or eliminate risk. |
This crosswalk is a deployment worksheet, not legal advice, statutory certification, agency endorsement, or a substitute for independent assessment. “Lawful” in the AINA/EISA technical sense means admissible under declared machine governance; it does not mean legal, moral, or institutionally legitimate. ## 9. Engagement Routes—and a Firewall Around Independent Review Different readers need different next steps. All begin with a written, bounded request to `me@blayneaina.com`. - **`[RESEARCH]`** — request method-scoped independent technical peer review, submit adversarial `EISA` / `RequestedAct` test vectors, or propose formal-verification collaboration outside the commercial ladder. - **`[GRANT / R&D]`** — conduct technical due diligence for non-dilutive AI safety / security grants, federal R&D programs (ARIA, UK AISI, NIST, DARPA, NSF), or university STTR partnerships. - **`[CHALLENGE]`** — evaluate one existing technical path (Track 1, 2, or 3) against one bounded transition under a predeclared acceptance contract. - **`[PILOT]`** — commission a privately scoped enterprise pilot, client-controlled Portable Deployment Node, vertical entity-resolution package, or implementation handoff. - **`[DISTRIBUTION]`** — discuss OEM, systems-integrator, vertical-channel, licensing, or deployment partnerships. - **`[INVESTMENT]`** — discuss strategic R&D capital, operating runway, vertical environments, and the resources required to expand formal verification and cross-platform delivery. The initial request should identify the technical track and evaluation or engagement type, participants and decision owner, bounded question, sample source and handling constraints, timetable, and the authority or resources relevant to the proposal. ### Exhibit 9.8 — Commercial work and independent review are separate ```text COMMERCIAL ENGAGEMENT paid and privately scoped → accepted deliverables, access, preparation, implementation, and support → confidential by default; optional redacted case record by advance agreement → positive, negative, refused, or unresolved technical result reported faithfully → payment buys no praise, evidence upgrade, endorsement, or outcome control INDEPENDENT EVALUATION method-scoped access appropriate to the claim → adequate ability to test, reproduce, narrow, or contradict the claim → no payment, investment, partnership, publication right, or favorable conclusion required → reviewer retains freedom to publish adverse findings → controlled access does not imply unrestricted source, data, or artifact rights ``` Independent evaluation is not a reward for buying a pilot, and a commercial engagement is not represented as independent validation. That firewall is part of the offer. ## 10. The Complete Machine, Now With an Address The eight architecture papers and this commercial doorway form one continuous route from probabilistic proposal to governed consequence. | Installment | Contribution to the deployment | |---|---| | [**01 — An AI Output Is Not an Act**](/01-an-ai-output-is-not-an-act/) | Standing must be established for one exact transition; model confidence is not authority. | | [**02 — Zero-Trust Ingestion & Server-Cut Evidence**](/02-zero-trust-document-ingestion-and-server-cut-evidence/) | Hostile sources become committed views with explicit omissions and citations the server slices and verifies. | | [**03 — Semantic Switch & EISA**](/03-eisa-four-gate-verdict-machine-and-entity-resolution/) | Semantic Switch protects native values before qBlock, Fracture, Forge, local restoration, Shield, and deterministic EISA recognition turn protected records into governed entity state. | | [**04 — The Room**](/04-the-room-why-git-and-chat-are-not-coordination-planes/) | Live actor, session, task, project, wake, and acceptance authority replace chat and Git as the coordination plane. | | [**05 — Expiry Is Not Fencing**](/05-expiry-is-not-fencing-typed-leases-and-kernel-write-exclusion/) | Typed leases, monotonic generations, and kernel write exclusion stop stale owners at the resource. | | [**06 — Rollback Is Not Erasure**](/06-rollback-is-not-erasure-dual-tier-vaults-and-ungated-quarantine/) | Primary and Amber preserve recovery baselines, per-edit states, and ungated divergence without laundering history. | | [**07 — Defeating Approval Laundering**](/07-defeating-approval-laundering-act-pinning-and-successor-grants/) | Controller-fixed acts, fenced SQLite execution, and successor grants preserve the exact effect and honest receipt lineage. | | [**08 — The Sovereign Kill Switch**](/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/) | Asymmetric containment, scoped epochs, in-doubt recovery, and candidate-binary conformance govern stop and release. | | **Article 09 — Bring One Real Boundary** | A bounded acceptance contract, independent peer-review firewall, and explicit R&D / deployment doorway put the machine to work. | The point is not to make probability sovereign. It is to make every consequential transition inspectable before, during, and after execution—and to give the receiving organization a system it can operate. ### Deployment & Evaluation Claim Card ```text Mechanism: Bounded technical challenge, predeclared acceptance contract, client-controlled smallest-sufficient deployment, and separate engagement routes. Failure it governs: AI adoption through vision decks and indefinite sandboxes; demonstrations without positive, refusal, or unresolved predicates; hidden custom scope, copied authority, unsupported graph claims, paid-proof conflicts, and compliance-by-acquisition assumptions. Positive path: qualified asset → written acceptance contract → existing implementation run → predicate-level result and artifact return → stop, narrow, pilot, deploy, distribute, fund, or independently evaluate. Refusal / hold path: request declined, narrowed, redirected, or held for handling/authority; predicate Failed / Unsupported / Not exercised / Inconclusive; unresolved effect preserved; production expansion excluded until separately scoped. Evidence status: Existing implementation evidence is author-controlled and author-verified. Operator history supports customer-embedded delivery credibility, not independent architecture validation. Regulatory mappings are engineering correspondence only. Verifier and artifacts: intake and acceptance contract; replay ID; admitted input and handling manifest; positive/refusal/unresolved results; view, case, act, and operation receipts; deployment/requalification checklist; exclusions and operating owner. Scope and claim ceiling: No invented customers, revenue, independent validation, universal deployment time, maintenance-free operation, matching accuracy, privacy guarantee, graph influence, production write-back, fixed price, fixed duration, or certification claim. Payment and investment never purchase a conclusion. Next stage: Independent technical peer review, non-dilutive R&D / grant collaboration, university STTR partnership, or a separately scoped challenge / enterprise pilot. Evaluation & R&D doorway: Written qualification to me@blayneaina.com using RESEARCH, GRANT, CHALLENGE, PILOT, DISTRIBUTION, or INVESTMENT. ``` ### Fourteen Claims for the Capstone | ID | Claim | Evidence and ceiling | |---|---|---| | `09-C01` | Three technical tracks provide bounded entry points | Each names an asset, existing implementation path, artifacts, and exclusions | | `09-C02` | Acceptance predicates are fixed before execution | Positive, refusal, unresolved, and claim-ceiling fields resist post-hoc goalpost movement | | `09-C03` | Initial challenges exclude hidden custom and production scope | Custom schemas, connectors, ingestion, rollout, graphs, and write-back require separate scope | | `09-C04` | Returned evidence is inspectable rather than testimonial | Agreed manifests, digests, receipts, and refusal records; no fake signature or guaranteed artifact | | `09-C05` | Deployment uses a client-controlled smallest-sufficient profile | Only admitted providers, workflows, schemas, and destinations enter the node | | `09-C06` | Portability does not transfer standing, current state, network independence, or effect coverage | Destination rebinding and installation acceptance remain mandatory | | `09-C07` | Maintenance and requalification are explicit operating duties | Connector drift, `rsa_` reassessment, schema changes, replay, and the 45-case gate | | `09-C08` | Semantic Switch and entity resolution provide a concrete commercial wedge | Privacy-first pipeline with transport, relation-fidelity, restoration, and provenance evidence, not universal accuracy or privacy | | `09-C09` | Relationship paths remain evidence-linked and non-authorizing | Six-hop is a query horizon, not influence, consent, outreach, or mutation authority | | `09-C10` | Operator history supports delivery credibility | 11+ years and documented delivery outcomes; not independent architecture validation | | `09-C11` | Regulatory mappings are engineering crosswalks | Primary-source scopes and residual organizational duties; no certification or endorsement | | `09-C12` | Commercial work is paid and privately scoped | Article/ebook free; no public fixed price/duration or paid-proof arrangement | | `09-C13` | Independent review remains outside the commercial ladder | Method-scoped access and freedom to publish adverse findings without payment or partnership | | `09-C14` | Independent peer review, R&D grant diligence, and bounded pilots have explicit intake routes | Structured written qualification; no invented customers, revenue, or endorsement | ## Bring the Boundary Bring one AI-proposed mutation, one hostile document, or one duplicate-record problem. Include the operating owner, the handling constraints, and the question you need answered. We will establish the exact transition, predeclare acceptance and refusal, run the appropriate existing implementation, and return the evidence and its ceiling. If the result supports a next step, we can scope it. If it fails or remains unresolved, that result stays visible too. > **Bring one real boundary. Make the result inspectable before deciding what comes next.**

Technical & Security Due Diligence FAQ

These are the concrete mechanisms and assurance ceilings behind a technical walkthrough. Each answer links back to the relevant architectural proof surface.
1. How can an already-authorized operation become a different operation between human approval and execution, and what precise mechanism prevents it? A high-level tool call can be approved at time t1 and mutated before execution at t2: a path or image can be substituted, child arguments or environment state can change, or an inherited writable handle can address an unapproved target. AINA binds approval to a controller-fixed, RFC 8785 JCS-hashed RequestedAct before the worker environment exists. The worker is spawned suspended inside a restricted Job Object; the pre-resume audit observes its token, image and handle inventory before instruction 0. requested_act::validate re-resolves current task, source, tenancy, destination and act identity at execution. A mismatch such as requested_act_mismatch or an unclassified pre-resume capability fails closed; no resume or destination mutation occurs.

Proof: §07 Act Pinning · actual refused-substitution receipt and unchanged SQLite digest

2. What happens when an agent or local process already holds a writable handle to a file before the kernel lease is established? The lease cannot be granted. On Windows, the coordinator opens the target with FILE_SHARE_READ and without write/delete sharing. If another process already holds a compatible writable handle, the kernel rejects the coordinator's open with ERROR_SHARING_VIOLATION (os error 32). Acquisition fails closed, no generation token is issued, and the territory remains contested until the foreign handles are closed and the baseline is re-verified. The gate does not pretend that a clock expiry revoked a handle the OS still holds open.

Proof: §05 Kernel Write Exclusion · os_lock_conflict in leases/oshandle.rs

3. What happens when the protected destination changes out-of-band after authorization but before final commit? Authorization binds a baseline file digest or SQLite row version; it does not authorize whatever bytes happen to exist later. At commit, file guards recheck the current preimage and tenancy while SQLite uses an immediate write reservation and a conditional UPDATE ... WHERE version = expected. A drift produces preimage_hash_mismatch or record_version_changed rather than an overwrite. Where candidate file bytes exist, divergence is preserved for forensic review rather than silently laundering it into a clean baseline. Quarantine is mechanism-specific, not a claim that every refused API call creates Amber bytes.

Proof: §03 Deterministic judgment · §05 fencing · §07 fenced SQLite

4. What happens after a successful database commit or atomic rename followed by a crash before the receipt reaches the caller? The operation is treated as in-doubt, not blindly retried. Durable operation identities and pre/post execution records let recovery determine whether the effect committed. A fresh successor task can receive a narrowly scoped result_read_only grant to inspect the historical outcome after its tenancy, grant, source, delivery, and receipt bindings validate. The successor does not re-execute the mutation and does not fabricate the original worker's first receipt: satisfies_original_return: false and resolves_original_obligation: false.

Proof: §07 crash continuity and successor grants · result_grant_* refusal family

5. How does Semantic Switch preserve useful matching relationships without exposing native identifiers, and what are its actual limits? For structured CRM and logistics records, Semantic Switch replaces lexical identifiers with deterministic surrogates while preserving equality relations, edit-distance structure, blocking keys, and graph topology. A confined worker can perform blind entity-resolution tasks without receiving raw customer identifiers in its prompt. It is an experimental alpha relational-confinement layer, not encryption or universal anonymization. Tokenizer boundaries, unconstrained prose, frequency analysis, auxiliary-data correlation, and production matching accuracy remain outside the demonstrated guarantee.

Proof: §03 production EISA core and alpha module · synthetic invariance testbed

6. What attacks have been tested against transformed records, how do you distinguish preserved relationships from re-identified entities, and what evidence exists beyond synthetic fixtures? Current public validation is limited to deterministic synthetic invariance fixtures and offline verifier behavior. The 20/20 block and 23/23 relation results test structural transport fidelity, not anonymity or population accuracy. I have not completed formal frequency-analysis, auxiliary-graph reconstruction, or production-corpus re-identification benchmarks, and no such guarantee is claimed. The dedicated isomorphic prototype's result is not evidence that every generator branch preserves every tokenizer or semantic relationship.

Proof: published fixtures, harnesses, and alpha claim ceiling

7. What is the smallest useful deployment of AINA in a real company today? One deterministic mutation gate around one high-consequence workflow. For example, gate AI-proposed merges in a HubSpot or Salesforce deduplication workflow: validate the act and baseline, require current tenancy, revalidate the destination at commit, and record rejected proposals instead of silently corrupting production records. You do not need to deploy every mechanism or migrate the whole company to begin proving value.

Proof: §01 machine standing · §06 forensic history · public refusal exhibit

8. What are the actual latency and resource costs? There is no universal millisecond benchmark to quote. JCS canonicalization, hashing, EISA evaluation, handle operations, and SQLite transactions are local primitives, but their cost depends on payload, fixture, process creation, destination, filesystem, and host. The fresh 45-case battery took 118.753576 seconds end to end, including isolated fixture provisioning, process launches, document handling, and SQLite work; averaging that duration across cases does not produce a per-gate latency measurement. Specific per-operation timings and LPAC worker-spawn overhead must be measured on the admitted profile, not asserted as empirical guarantees.

Proof: fresh execution timing and candidate attestation · §08 assurance boundary

9. What happens when the governing controller itself is compromised by a local administrator or kernel exploit? AINA does not claim to defeat a hostile OS kernel, ring-0 rootkit, or administrator controlling the same host. The local interactive-console owner, operating-system kernel, and controller host are inside the v1 trusted computing base. The current runtime confines untrusted model outputs, hostile documents, and concurrent workers beneath that controller. Hardware attestation or multi-node enclave designs are outside the current implementation and are not represented as deployed capabilities.

Proof: §08 sovereign stop and its trust boundary

10. How much of the architecture is deployable across multiple machines, and which parts have run in customer production? The unified AINA reference runtime, Room coordinator, and 45-case gate were built and exercised on a single Windows workstation as independent R&D. They have not yet been deployed as a unified system in an external enterprise production environment. The 11+ years of production history is the underlying delivery record: CRM data hygiene, multi-company ETL migration, 800+ carrier workflows at DrayMaster, deterministic operational state machines at Photector, and HubSpot/Salesforce integration work. That experience informs the architecture but is not represented as external production validation of AINA itself.

Proof: this deployment boundary · local candidate evidence and its stated ceiling

11. What is the smallest useful deployment of AINA in a real company today?

The smallest practical enterprise deployment is a single CRM/database mutation gate wrapping §01 (proposal versus act), §03 (deterministic EISA checks), and §06 (pre-image and forensic vaulting). An existing model or subagent interacting with HubSpot, Salesforce, SQLite, or another relational store is placed behind one MCP tool boundary. The gate validates the controller-fixed RequestedAct, checks optimistic row versions, journals the pre-attempt state, and requires unapproved substitutions to fail closed with zero database effects. The whole nine-module architecture is not a prerequisite.

Proof: §01 output boundary · §06 dual-tier recovery · Case C SQLite refusal proof

12. What are the actual latency and resource costs?

Operational gate latency and full-lifecycle test duration are different measurements:

  • In-process gate work: JCS canonicalization, SHA-256 hashing, EISA evaluation, and SQLite predicates are local operations, but this battery did not produce a standalone per-gate latency benchmark. No <5 ms figure is claimed.
  • Full battery duration: The 45-case run took 118.753576 seconds, an average of about 2.64 seconds per scenario. That includes isolated fixture provisioning, process launches, evidence parsing, SQLite initialization, lease contention, and durable disk flushes—not the cost of evaluating one envelope in memory.

Proof: binary composition and timing note · candidate attestation

13. Which components come from real customer production environments vs. independent R&D?

AINA keeps enterprise delivery history separate from the unified reference runtime:

  • Enterprise production delivery (11+ years): managing rate and operational data for 800+ intermodal carriers at DrayMaster (acquired by Envase); facilitating deterministic workflow and scenario whiteboarding for Photector's patent-pending photographic-evidence application (integrated with Hirschbach TMS); building deduplication workflows across HubSpot, Salesforce, APIs, and SFTP at Agile Education Marketing; and migrating 15,000+ SKUs across six retail locations at Jacuzzi Colorado.
  • Independent R&D reference runtime: the 56.2 MB squad.exe Windows NT supervisor, the 45-case conformance battery, and the refusal receipts were engineered during 14 months of independent work on a single workstation. They are not represented as a unified external customer production deployment.

Proof: this production boundary · offline verification bundle

14. How can a security researcher, peer reviewer, or R&D funding evaluator verify the runtime beyond the static proof pack?

Verification has two tiers:

  1. Offline cryptographic verification (immediate): Download aina-refusal-proof-pack-v1.zip and run python verify_refusal_proof.py locally without internet access to audit all four refusal records, 12 schemas, and the 45-case manifest against published SHA-256 digests.
  2. Live interactive walkthrough (peer-review & diligence session): In a 45-minute technical screen share, the reviewer or evaluator chooses an ad-hoc destination target, unauthorized parameter substitution, or simulated lease collision and observes the running worker refuse the mutation while the destination remains unchanged.

Proof: Refusal proof pack · Schedule a live boundary walkthrough (me@blayneaina.com)

## Frequently Asked Questions ### What should I bring to an initial challenge? Bring one bounded transition, its decision owner, an authorized synthetic/redacted/existing sample, handling constraints, and the positive and negative outcomes you need to evaluate. Do not send confidential production data or credentials with the first inquiry. ### Does a challenge include production integration or CRM write-back? No. It covers the accepted existing implementation path and its evidence. Custom schemas, connectors, production ingestion, deployment, graph construction, and write-back require separately scoped paid work. ### Does a client-controlled node mean the whole system is offline? Not automatically. There is no mandatory vendor-hosted AINA cloud, and local state can remain under client custody. External models, APIs, and destinations still require their declared network routes, credentials, terms, and evidence boundaries. ### Why is entity resolution the commercial wedge? Duplicate contacts, companies, and products create a concrete, costly, evidence-rich problem. Semantic Switch lets external computation operate on structure-preserving fictional records while native values remain local; qBlock, Fracture, Forge, exact restoration, and Shield then create a defensible identity substrate. The pipeline does not guarantee matching accuracy or authorize graph-driven outreach and mutation. ### Does a paid pilot buy independent validation? No. Commercial work is paid and privately scoped, but payment cannot buy praise, certification, an evidence-status upgrade, or outcome control. Independent evaluation uses a separate method-scoped research route and permits adverse findings. ### Do the regulatory mappings establish compliance? No. They identify technical evidence that may support an organization's obligations. Use-case classification, jurisdiction, management systems, legal duties, DPIAs, SOC examinations, and conformity assessments remain separate organizational and professional work. ## Continue the Series - **Series index:** [From Proposal to Consequence: The AINA Transition-Governance Series](/) - **Previous:** [Article 08 — The Sovereign Kill Switch: Break-Glass Epochs and the Conformance Gate](/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/) - **Start at the doctrine:** [Article 01 — An AI Output Is Not an Act](/01-an-ai-output-is-not-an-act/) - **Independent evaluation:** use the `[RESEARCH]` route (`mailto:me@blayneaina.com?subject=[RESEARCH]%20Independent%20Technical%20Evaluation`); adverse findings remain publishable and payment is never required. --- ## Primary Framework and Regulatory References - [NIST AI Risk Management Framework Core — GOVERN, MAP, MEASURE, MANAGE](https://airc.nist.gov/airmf-resources/airmf/5-sec-core/) - [NIST NCCoE — Software and AI Agent Identity and Authorization](https://pages.nist.gov/nccoe-ai-identity/) - [ISO/IEC 42001:2023 — AI management systems](https://www.iso.org/standard/42001) - [EU AI Act — consolidated Regulation (EU) 2024/1689](https://eur-lex.europa.eu/eli/reg/2024/1689/2026-07-27/eng) - [UK Data (Use and Access) Act 2025, Section 80 — automated decision-making](https://www.legislation.gov.uk/ukpga/2025/18/section/80) - [Federal Reserve SR 26-2 — Revised Guidance on Model Risk Management](https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm) - [Federal Reserve — Supervisory Guidance on Model Risk Management, including scope note 3](https://www.federalreserve.gov/frrs/guidance/supervisory-guidance-on-model-risk-management.htm) - [AICPA & CIMA — System and Organization Controls suite](https://www.aicpa-cima.com/resources/landing/system-and-organization-controls-soc-suite-of-services) - [European Data Protection Board — Data protection impact assessment](https://www.edpb.europa.eu/topics/accountability-and-compliance-tools/data-protection-impact-assessment_en) --- ## Publication Metadata **Title tag:** Bring One Real Boundary: Technical Peer Review, Challenge Tracks & R&D Evaluation | Blayne Aina **Meta description:** Evaluate the AINA architecture and reference runtime through three bounded technical challenge tracks, explicit acceptance contracts, independent peer review, and R&D grant due diligence. **Series navigation:** [Series index](/) · Evaluation, peer-review, and deployment doorway of *From Proposal to Consequence: The AINA Transition-Governance Series*. **Previous installment:** [Article 08 — The Sovereign Kill Switch: Break-Glass Epochs and the Conformance Gate](/08-sovereign-kill-switch-break-glass-epochs-and-conformance-gate/). **Series index:** [From Proposal to Consequence](/).