Open-access systems papers and reproducible empirical benchmarks on cryptographic tool receipts,
runtime verification for Model Context Protocol (MCP) agents, provenance-gated memory, and
neuro-symbolic state grounding by Sean Maxwell
(ORCID: 0009-0004-0133-0463).
Technical Reports & Preprints (2026)
RS-TR-2026-05 (Preprint) • November 2026 • osaudit-verified artifact
Language-model agents using the Model Context Protocol (MCP) face three runtime verification gaps: (i) action fabrication, where post-tool prose narrates side effects that failed, timed out, or never ran; (ii) provenance collapse, where unverified model prose or web text enters durable memory alongside executed tool outputs; and (iii) syndicated echo amplification, where mirrored web pages satisfy multi-source corroboration from one origin. Existing MCP firewalls inspect wire calls before execution, while LLM-judge guardrails require multi-gigabyte VRAM without binding prose to receipts. We present epistemic-guard (v0.5.2), an open-source Python library and out-of-process stdio MCP receipt proxy (15 modules, 0 MiB GPU VRAM). epistemic-guard isolates Ed25519 signing keys out of process, writes RFC 6962 Merkle logs with AER-1 receipt translation, checks post-turn prose on CPU with SHA-256-verified OpenVINO INT8 encoders (4 models, 279.48 MB total; 82.67 MB on the supervision path), gates SQLite writes across 7 provenance domains, and collapses verbatim and lightly edited wire-syndicated RAG mirrors. On conformance suites, epistemic-guard rejects 100.0% of tampered receipts (200 Merkle leaves), passes 58/58 official AER-1 ingestion vectors and 50/50 reference-verified AER-1 exports, verifies 100.0% of 40 fault-injected wire calls and 100 concurrent appends with 0 key leaks, and blocks 100.0% of 100 unauthorized memory writes. Across N=250 frozen-benchmark traces, epistemic-guard reaches 93.2% accuracy at 17.9 ms warm median CPU latency (8.62 ms single-clause; 921.8 ms cold-start load), beating 12B-14B GPU judges by 6.4-6.8 points (p=0.0259, p=0.0095) at 4.26-4.5x lower latency; on N=60 live mcp-proxy traces, it catches 96.15% of hazards at 76.67% accuracy (38.24% FPR; +18.34 points over Stage-1-only, p=0.001) in 9.16 ms p50.
Contemporary agent architectures commonly decompose persistent state into five layers: Working Memory, Episodic Memory, Semantic Memory, Procedural Memory, and an offline Meta-Memory control plane. However, conventional implementations connect these layers via unconstrained asynchronous extraction and reflection jobs that read raw episodic transcripts and promote them directly into semantic facts and procedural skills. We present the unified five-layer memory and skill-induction architecture of Project Lucca and report empirical measurements across every layer boundary on local consumer silicon. Six HKDF-derived HMAC-SHA256 domain signers and SQLite triggers separate autobiographical recall from factual grounding, reducing multi-session self-poisoning from 100.0% in Naive RAG to 0.0% across four Gemma-4 parameter scales while preserving 100.0% autobiographical continuity. Within Semantic Memory, a zero-VRAM OpenVINO MiniLM-L6 cross-encoder evaluates semantic conflicts in 12.23 ms mean CPU latency, routing contradictions into Belnap's four-valued paraconsistent state paired with OpenNARS Beta evidence weights, lifting LongMemEval accuracy from 23.3% to 70.0% at 8.1 ms mean retrieval latency, while Z3 SMT quarantine verifies induced MicroScheme procedural rules prior to activation.
Autonomous desktop and cloud agents must not assert that state-modifying actions have occurred when no verifiable execution receipt exists. While deterministic speech-regex pattern matchers are widely deployed as low-latency safety guards, live production audits reveal substantial failure rates on paraphrased side effects. We establish an empirical taxonomy of four grammatical evasion modes: active progressive participles, passive and telegraphic confirmations, unlisted verb lemmas, and markdown formatting syntax. Across a pre-registered held-out benchmark of N=120 unseen prompts and a 720-turn blinded semantic re-audit (validated by a second independent rater at kappa = 1.0000), plus 1,000 completions across four Gemma-4 scales, we compare regex supervision, expanded grammar matching, and Structured-Output Action Contracts (SOAC) verified against cryptographic tool receipts in 6.82 microseconds mean lookup latency.
Autonomous agent architectures deploying persistent Retrieval-Augmented Generation commit interaction history into long-term storage, introducing recursive self-poisoning when transient model confabulations emitted in Session t are retrieved in Session t+k as ground truth. We evaluate an immutable Cryptographic Provenance Gate enforced through SQLite database triggers across a pre-registered benchmark of N=100 paired multi-session probes against Naive RAG, No-Assistant Memory, and In-Prompt Soft Tagging. Under sampled adversarial decoding (T=0.7), In-Prompt Soft Tagging degrades to 40/40 (100.0%) poisoning while the Provenance Gate remains at 0/40 (0.0%), and preserves 29/30 (96.7%) benign multi-step continuity compared with 0/30 (0.0%) under assistant-turn deletion.
Autonomous agents driven by Large Language Models operating in extended multi-session settings suffer from recursive hallucination compounding and fabricated action claims. We present the architecture and mechanical evaluation of Project Lucca, a local-first neuro-symbolic agent runtime enforcing a Six-Signer Cryptographic Provenance Gate using domain-separated HKDF HMAC-SHA256 signatures for persistent memory commits and a Deontic Truth-in-Receipt Output Supervisor that screens generated speech tokens against deterministic execution receipts before user presentation, evaluated across a 73-test unit battery and a 100-probe live action-hallucination testbed.