Work

Decoupling Attribution from Inference: Strengthening Legal-Geometric Auditing (RFC-006) via Deterministic Gates, Cryptographic Seams, and Anti-Conflation Provenance (RFC-007)

Year: 2026
Type: preprint
Abstract: 📌 TECHNICAL DISCLOSURE & FOLLOW-UP PREPRINT ARCHIVE (RFC-007) This record archives Request for Comments (RFC-007): "Decoupling Attribution from Inference: Strengthening Legal-Geometric Auditing (RFC-006) via Deterministic Gates, Cryptographic Seams, and Anti-Conflation Provenance." It serves as a methodological audit and architectural upgrade to RFC-006 ("The Geometry of Institutional Enclosure", DOI: https://doi.org). Related Source Repositories & Artifacts: • Legacy Preprint (RFC-006): https://doi.org • deterministic-ai-gate (v1.0, MIT): https://github.com/Borealiscodes/deterministic-ai-gate • active-inference-commit-seam (v0.1.0, Dual-Boundary MIT / CAUSA NC 1.1): https://github.com/Borealiscodes/active-inference-commit-seam/tree/main • Upstream CAUSA Library (Stell, CAUSA Non-Commercial License 1.1): https://github.com/stell2026/Causa ⚠️ EXPLICIT ATTRIBUTION CLARIFICATION & EPISTEMIC STATEMENT To maintain strict scholarly and engineering provenance, this record issues three explicit clarifications regarding RFC-006 and upstream software dependencies: 1. Zero Normative Attribution to Stell: Stell—author of the Anima cognitive architecture and the stell2026/Causa causal agency library—has NEVER made any statements, claims, or legal arguments regarding transnational repression, administrative denial of remedy, or international human-rights treaties. CAUSA provides domain-agnostic mathematical metrics (directional ownership and expression congruence) over abstract numeric vectors. Any application of those metrics to legal or administrative case files rests solely with the downstream researcher (NDH-RESEARCH-PILOT). 2. Empirical Audit of LLM Citation & Tool Conflation: The blurred attribution layers present in the legacy RFC-006 draft are analyzed here as an instructive, real-world demonstration of how Large Language Models (LLMs) conflate co-occurring context windows—attaching normative legal statements to technical libraries and inventing unified annotation narratives. This meta-finding directly motivates replacing LLM-managed citations with fail-closed deterministic CPU verification gates. 3. Reference Prototype & Dual-License Boundary: The companion software baselines (deterministic-ai-gate and active-inference-commit-seam) are clean-room reference prototypes requiring domain-specific empirical validation prior to production deployment. Core verification modules are MIT-licensed with zero external dependencies, while the optional CAUSA telemetry adapter is isolated behind an explicit boundary respecting the CAUSA Non-Commercial License 1.1. 🛡️ SUMMARY OF CONTRIBUTIONS • Clean-Room Input/Output Verification (deterministic-ai-gate): Locks raw legal and administrative source documents into pre-LLM SHA-256 chunk manifests, constrains LLMs strictly to schema-bounded text-to-JSON parsing, and verifies quotes, 2048-bit bounded rational arithmetic (Q_B), and Kahn topological DAG timelines deterministically on CPU. • Proposal-to-Commit Architectural Seam (active-inference-commit-seam): Isolates exploratory triage (Salience Δ, Structural Resonance Φ, Active Inference queries, and optional CAUSA permutation baselines) on the proposal side of a strict commit boundary (Record ≠ Admit ≠ Commit), enforcing the Law of Non-Self-Warranting Memory (repetition ≠ evidence). • Formal Separation of External Warrant & Internal Agency: Defines admissibility into evidentiary ledgers as the strict conjunction of external SHA-256 provenance (W_ext) and statistically significant internal agency above shuffled permutation baselines (A_int), ensuring internal vector coherence can never substitute for primary source truth.
Language: English
Open Access status: green
OpenAlex ID: W7234411720
Cites:
Cited by:
Related to: 10