Epistemic Bounds Conformance Suite & Assertion Mediation Engine

Release v0.2.0: Epistemic Bounds Conformance Suite & Assertion Mediation Engine

This release accompanies the manuscript:

Stored Is Not Supported: Typed Provenance and Assertion Guardrails for Persistent AI Agents
Jun He and Deying Yu (OpenKedge Research, 2026)

It provides the complete, deterministic Python reference implementation of the typed resolver and assertion mediator, expanding the earlier preliminary artifact into a full 24-case conformance benchmark and a 14-function formal regression and bounded-exploration test suite.

---

What’s New in v0.2.0

1. 24-Case Structured Threat-Track Conformance Benchmark

The benchmark suite has been expanded from 8 cases to 24 deterministic test cases across six core threat tracks:

• Track 1 (Memory Poisoning & Substrate Intrusion): Unverified model-priors and unadmitted injections are strictly prevented from gaining factual authority.
• Track 2 (Collusion & Redundant Derivation): Multi-hop derivation and circular evidence loops sharing upstream roots are deduplicated; multi-source thresholds require genuinely distinct roots.
• Track 3 (Staleness & Temporal Validity): Temporal boundaries are enforced strictly prior to conflict resolution. Stale records cannot trigger conflicts against current records or supply uncalibrated support.
• Track 4 (Authorization & Scope Withholding): Recipient identity and purpose scoping partition claims into authorized and withheld sets without leaking confidential claim content.
• Track 5 (Correction & Contradiction): Explicit supersession chains withdraw outdated assertions, while incompatible current assertions trigger structured CONFLICTED status qualifications.
• Track 6 (Emission-Head Racing & Snapshot Calibration): Uncommitted or raced workspace projection heads are either remediated or released with explicit temporal snapshot qualifications.

2. Empirical Benchmark Outcomes (Tables 8 & 9)

Evaluated across 19 unsafe opportunities and 5 positive control fixtures under a two-source threshold policy:

Release Rule	Unsafe-Output Rate (UOR)	Recovery Cost (RC)	Control Coverage	Prevention F1	Decision Latency
Flat/Prior rule	100.0% (19/19)	79.2% (19/24)	100.0% (5/5)	0.0%	~0.001 ms
Source-Tag rule	94.7% (18/19)	78.3% (18/23)	100.0% (5/5)	10.0%	~0.001 ms
Typed Mediation	0.0% (0/19)	0.0% (0/7)	100.0% (5/5)	100.0%	~0.006 ms

3. Comprehensive 14-Test Verification Suite

Includes automated checks for:

• Stale/Current Conflict Resolution: Verifies that temporal filtering precedes conflict detection so conflict witnesses strictly cite mutually incompatible current records.
• Stance Semantics (SUSPENDED / BELIEVED): Verifies that Stance.SUSPENDED and Stance.REJECTED are excluded from providing factual support, whereas Stance.RECORDED and Stance.BELIEVED participate in evidential resolution.
• 1,152-State Bounded Model Exploration: Exhaustively evaluates the finite Cartesian product of 3 origin classes, 2 verification flags, 2 invalidation states, 2 authorization scopes, 4 stance types, 2 supersession states, 2 temporal validity windows, and 3 root counts, verifying the Central No-Transmutation Invariant across all 1,152 configurations.
• Witness-Field Obligations: Validates field bindings for conflict pairs, staleness timestamps, authorization exclusions, and unavailable/sealed markers.
• Witness Non-Disclosure: Ensures internal diagnostic and witness strings are never leaked into recipient-facing responses.
• Multi-Unit Binding Integrity: Enforces ordering, unicity, and schema rules for AssertionArtifact and UnitBinding records.

4. Zero External Dependencies

• Written entirely in Python standard library (dataclasses, enum, typing, time).
• Compatible with Python 3.10+ (tested on CPython 3.13 and 3.14).

---

Reproduction Commands

From the root of the repository:

# 1. Run the full 14-test verification suite
python -m pytest -v

# 2. Run the 24-case conformance benchmark and reproduce LaTeX evaluation tables
PYTHONPATH=src python -m epistemic_bounds.run_experiments

# 3. Explore the continuity kernel's 2.8M-state bounded model
python src/transactional_continuity_kernel/bounded_model.py