1. The Specific Security Problem: Why Probabilistic Guardrails Cannot Govern Consequence
As autonomous AI systems gain direct access to operating systems, databases, code execution, and cyber-physical actuators, four structural failure gaps emerge at the boundary between a model's output and an external state mutation:
- Untrusted Proposals & Model-Monitor Failure Modes: Prompt instructions, RLHF, and secondary LLM-as-judge monitors operate over probabilistic token sequences. AINA deliberately does not rely on proving the general internal safety or alignment of an opaque neural network; instead, it restricts which externally observable operations can be authorized and checks the admissibility conditions for those operations inside a minimal, deterministic Trusted Computing Base (TCB).
- Multi-Agent Sycophancy & Echo-Chamber Quorum Laundering: Multiple agent instances can converge on a hallucinated or adversarial action without independent primary-source evidence.
- Approval-to-Execution (TOCTOU) Substitution: Between asynchronous human or policy approval and actual OS process execution, command-line arguments, environment blocks, target files, or inherited kernel handles can be substituted.
- Stale-Worker Overwrites & Unverifiable Rollbacks: Delayed workers holding expired clock-based TTL locks can overwrite newer state upon waking, and ordinary application logs rarely provide cryptographic proof that a refused attempt left the target resource untouched.
2. The Authority-to-Effect (A2E) Architecture & Three Enforcement Layers
AINA interposes a deterministic reference monitor at the commit boundary across four composable mechanisms and three explicit enforcement layers:
- Evidence Instruction Set Architecture (
EISA— Seven Closed Opcodes): Every proposed action must compile into a closed 10-key envelope (squad.eisa.act_envelope.v1.json) restricted to seven closed opcodes across four acyclic layers:BindEvidenceTag(Layer 0,Fracture),RequestHumanReview(Layer 0,Fracture),RejectUnlawfulAct(Layer 0,Fracture),AssertLayoutRelation(Layer 0,Forge),AssertDuplicatePair(Layer 1,Fracture),AssertTransitiveClosure(Layer 2,Fracture), andProposePartition(Layer 3,Fracture). - Three Explicit Enforcement Layers:
- Layer 1 (Pre-Parse & Deserializer Boundary): Enforces byte caps (
MAX_SUBMISSION_BYTES = 1,048,576,MAX_REQUESTED_ACT_BYTES = 65,536,MAX_EVIDENCE_REF_BYTES = 4,096),MAX_ACTS = 256, single-stage submission homogeneity, duplicate JSON key rejection (visit_map), and non-finite number rejection (visit_f64) before AST allocation. - Layer 2 (Recursive Draft-07 JSON Schema Validation): Enforces recursive
additionalProperties: falseclosure across all 12 schemas,allOfopcode-to-stage-to-operands binding, integer-onlyconfidencebasis points (0..10000, zero floats), 15 non-inferential evidence tags, 47 registered fault codes, RFC 6901 pointer depthMAX_POINTER_DEPTH = 8(regex^(?:/(?:[^/~\u0000-\u001f]|~[01])*){1,8}$) andMAX_POINTER_BYTES = 256,MAX_REASON_CHARS = 1,024,MAX_ID_CHARS = 64,MAX_REFS = 16,MAX_LIST = 1,000, andMAX_CHAIN = 4. - Layer 3 (Runtime Four-Gate Verdict Machine & OS Kernel Confinement): Enforces
MAX_CASE_BYTES = 4,194,304,MAX_CASE_RECORDS = 1,000,MAX_VALUE_CHARS = 1,024, andMIN_PHONE_DIGITS = 7over sealedeisacase_<32hex>substrates; acyclic Layer0 → 3DAG resolution;counterevidence_suppressedscanning; Disjoint-Set Union-Find partition verification; and Windows NT kernel process/handle confinement.
- Layer 1 (Pre-Parse & Deserializer Boundary): Enforces byte caps (
- Server-Cut Provenance (
srv_/ev_) & Sycophancy Collapse (consensusIdentity): Requires controller-computed SHA-256 digests and exact 64-bit integer byte-offset spans (squad.evidence_ref.v1.json) prior to admission, collapsing correlated agent claims lacking independent source spans. - Approval-to-Execution Act-Pinning (
act-pin) & Monotonic Lease Fencing: Binds a controller-fixedRequestedActdigest to the OS worker process instance across spawn and enforces monotonic integer generation fencing at commit time. - Proof-Carrying Non-Action (
PCNA): Emits canonical JSON refusal receipts binding the refusal fault code to pre-attempt and post-refusal SHA-256 digests of the measured target resource.
3. Existing Baseline vs. Proposed Research & Engineering Capabilities
| Capability Domain | Completed Baseline (Built & Inspectable Today) | Proposed Funded R&D (Foreground Work — Not Claimed as Built Today) |
|---|---|---|
| OS Reference Monitor & Act-Pinning | Windows NT reference runtime (squad.exe): 56.2 MB Rust supervisor enforcing CreateProcessW(CREATE_SUSPENDED) + PssCaptureSnapshot pre-resume inspection, Win32 share-mode write exclusion, and SQLite row-version checks; passes 45/45 author-controlled isolated conformance cases and 20/20 server-cut evidence cases. | Linux Pre-Execution Authorization Barrier & Automated OS Hardening (AINA-LinuxHardener — PRJ-NAVY-01): Pairing clone3(CLONE_PIDFD) with a pre-execution synchronization barrier that verifies /proc/<pid>/fd, /proc/<pid>/maps, cgroup v2 membership, and an F_SEAL_WRITE-sealed memfd executable digest prior to applying Landlock LSM / seccomp-bpf and invoking fexecve(), demonstrated on Rocky Linux 9 / RHEL 9 + containerized ROS 2 Humble. |
Refusal & Schema Verification (PCNA) | Offline Refusal-Bundle & Schema Verifier (proof/verify_refusal_proof.py): Zero-dependency Python 3 script verifying (1) 39/39 files against SHA256SUMS.txt, (2) 3 static fixture SHA-256 invariance chains (pre_sha256 == post_sha256) + 1 absent-preimage stale-generation refusal, and (3) recursive additionalProperties: false closure + 18/18 positive and 26/26 negative Draft-07 and pre-parse bound tests across all 12 schemas. | Authenticated PCNA v2 & External Security Assessment (PRJ-LC-01): Adding Ed25519 / hardware-backed TPM 2.0 asymmetric receipt signing, monotonic supervisor counter binding, and commissioning an independent third-party security audit of the runtime and verification package. |
| Formal Methods & Crash Recovery | Closed EISA JSON Schema & Runtime Recognizer: 6 structural finiteness bounds, 7 closed opcodes, and 47 fault codes enforced across pre-parse, Draft-07 schema, and Rust recognizer layers. | Mechanized Lean 4 Semantics (eisa-lean4) & Crash-Recovery Benchmark (PRJ-TAIF-01): Machine-checked Lean 4 termination and verdict-uniqueness proofs across all 7 opcodes and 47 fault classes + a 10,000-trial mid-commit crash-injection benchmark for result_read_only successor grants (satisfies_original_return: false). |
| Empirical AI-Control Benchmarking | 45-case conformance & 20-case evidence battery: Author-controlled release qualification in isolated Windows NT fixtures. | 250-Case Multi-Model AI-Control Comparison Study (PRJ-CG-02): Controlled empirical comparison of (1) standard OS permissions, (2) frontier LLM-as-judge monitors, and (3) deterministic A2E enforcement across multi-agent collusion, tool manipulation, and TOCTOU races. |
| Hardware Actuation Interlock | Tier 1/2 Software/OS Reference Implementation + USPTO Provisional #3 architectural specification. | Tier 4 FPGA Proxy Substrate & Execution Latch Benchtop Demonstrator: Physical RTL / FPGA hardware bus-gating prototype for edge actuators. |
4. Publicly Reproducible Evidence & Exact Claim Ceiling
Anyone can verify the public evidence bundle offline in under 10 seconds with Python 3 and zero external dependencies:
git clone https://github.com/SquadForge/blayne.aina.git
cd blayne.aina
python3 proof/verify_refusal_proof.py
What verify_refusal_proof.py proves today (Three-Part Offline Verification): (1) Manifest Integrity: All 39 published files match their SHA-256 digests in SHA256SUMS.txt; (2) Fixture & Conformance Consistency: The 3 fixture-backed refusal records (evidence-selector.json, unleased-write.json, act-pinning.json) match the exact SHA-256 digests and byte lengths of the preserved static fixtures in proof/refusal/fixtures/, stale-generation.json records an absent preimage, and the 45/45 conformance and 20/20 evidence manifests match candidate attestation; and (3) Recursive Schema & Pre-Parse Bound Validation: All 12 public JSON schemas pass recursive additionalProperties: false closure plus 18/18 positive and 26/26 negative Draft-07 and pre-parse bound tests across all 7 EISA opcodes. What it does not claim by itself: It does not by itself prove independent third-party execution of the private squad.exe binary or verify an asymmetric hardware signature authenticating receipt origin—capabilities explicitly scoped under Authenticated PCNA v2 and our independent external security audit milestone (PRJ-LC-01).
5. The Independent Research & Assurance Programme (Non-Overlapping Work Packages)
| Work Package Code | Research / Engineering Focus | Primary Measurable Question & Deliverables |
|---|---|---|
PRJ-TW-01Assurance Initiative | Institutional Launch & Cross-Framework A2E Assurance Standard | Establish the independent public-benefit AINA Assurance Initiative, hire dedicated systems/kernel and AI-control engineers, commission two external lab audits, and publish cross-framework A2E enforcement adapters. |
PRJ-LC-01Open Verifier & Audit | eisa-core / pcna-verifier Release Packaging, Authenticated PCNA v2 & External Security Assessment | Package the standalone open-source admission validator and verifier, add Ed25519 / TPM 2.0 asymmetric receipt signing (Authenticated PCNA v2), and publish an unedited third-party security assessment. |
PRJ-CG-02Empirical AI Control | 250-Case Multi-Model Comparison: Deterministic A2E vs. LLM-as-Judge Monitors | Quantify unauthorized mutation rates, false-rejection rates, sycophantic false-consensus acceptance, and p50/p95/p99 latency across 4 frontier model families comparing OS permissions, LLM judges, and A2E. |
PRJ-TAIF-01Formal Semantics | Lean 4 EISA Mechanization (eisa-lean4) & Crash-Straddling Fault-Injection Benchmark | Prove bounded-step termination and verdict uniqueness in Lean 4 across all 7 opcodes and 47 fault classes, and evaluate result_read_only successor-grant crash recovery across 10,000 injected mid-commit faults. |
PRJ-NAVY-01Linux OS Hardening | Automated Linux Debloating, Hardening & Pre-Execution Barrier (AINA-LinuxHardener) | Automate RPM/ELF/module dependency analysis, OS debloating, Landlock/seccomp-bpf synthesis, Linux Pre-Execution Authorization Barrier, and deterministic ROS 2 Humble functional regression testing on Rocky Linux 9 / RHEL 9. |
6. Planned Open-Source Deliverables
eisa-core& Closed JSON Schemas: Standalone deterministic validator and closed 7-opcode schema definitions (squad.eisa.act_envelope.v1.json).pcna-verifier&Authenticated PCNA v2Tooling: Open-source offline bundle verifier and Ed25519 / TPM-attested receipt signature checker.eisa-lean4Formal Specification: Mechanized Lean 4 operational semantics and termination proofs for the 6-boundEISAgrammar.- 250-Case Multi-Agent AI-Control Benchmark Dataset: Reproducible adversarial test vectors and comparative evaluation harness for LLM-judge vs. deterministic reference-monitor benchmarking.
7. Background IP, Verbatim USPTO Provisional Schedule & Licensing Distinction
All background architecture developed prior to external grant funding is covered by three solo-authored U.S. Provisional Patent Applications filed in April–May 2026 by Blayne Manuia Aina (100% solo owned, zero third-party encumbrances):
- U.S. Provisional #1 (Filed April 12, 2026): Autonomous Intelligent Network Architecture Protocol; Systems, Methods, and Computing Architectures for Constitutionally Governed Machine Action, Machine Participation, Machine Encounter, Machine Execution, and Machine Finality Across Multi-Stage Verifiable State-Formation, State-Extraction, and Temporal Defense.
- U.S. Provisional #2 (Single Omnibus Application, Parts 2A & 2B — Filed May 2, 2026):
- Part 2A: AINA Semantic Switch and MX2MX; Systems, Methods, and Computing Architectures for Constitutionally Created Machine Domains, Typed Correspondence Governance, Protected Cross-Boundary Admissibility, Consequence-Bearing Re-Binding, and Monotonic Burden Accumulation Under an Effective Constitutional Envelope.
- Part 2B: AINA Topological Physics Pipeline; Systems, Methods, and Machine Architectures for Candidate-Universe Sealing, Asymmetrical Topology Adjudication, Claim-Graph Canonical Extraction, Reversible Patch-Plan Materialization, Baseline-State Defense, and Instruction-Bundle Constraints Under Stage-Local Constitutional Governance.
- U.S. Provisional #3 (Filed May 15, 2026): AINA Authority-to-Effect (A2E) Governance: Layer-Zero Constitutional Transduction of Machine-Computed Authority into Consequence-Bearing Actuation Under an Effective Constitutional Envelope.
Clean Background vs. Foreground Separation: Pre-existing background IP and the private Windows NT squad.exe runtime remain protected background assets (available for live demonstration and scoped evaluation under mutual NDA, or under applicable federal SBIR Data Rights for defense contracts). All public-interest foreground research deliverables—including the public JSON schemas, standalone verification packages (eisa-core, pcna-verifier), mechanized Lean 4 proofs (eisa-lean4), empirical benchmark datasets, and third-party security audit reports—are structured for open publication (MIT / Apache-2.0) to support independent scientific reproduction (me@blayneaina.com).