Audit-Closed AI Scientist Protocol

This record provides the full source package for “Audit-Closed AI Scientist Protocol: Trustworthy Autonomous Scientific Discovery for Self-Driving Laboratories.”The work introduces an audit-closed governance protocol for autonomous scientific discovery in which every accept/reject/update decision is a deterministic function of public transparency logs and replayable incorporation certificates. The framework integrates: typed stochastic observation interfaces, tolerance-aware coherence auditing on finite acyclic audit graphs, always-valid sequential testing with e-processes/e-values, adaptive contextual sampling with logged propensities (including variance-adaptive betting), anti-griefing novelty accounting, and drift recovery with subgraph-local fault localization and exogenous-certified root-memory decay. It also specifies physical deployment safeguards, including attested measurement provenance, cross-sensor coherence witnesses, and hierarchical sentinel-based recalibration. The repository includes the manuscript source, complete appendix reproducibility code, and machine-readable output manifest. All reported synthetic benchmark values in Section 13 are generated from the included script via: python appendix_reproducibility_code.py with fixed seed 20260222. The output manifest includes SHA-256 digests and an automated manuscript-value consistency check.Benchmark claims are synthetic (not a physical lab deployment) and are conditional on explicitly stated Genesis assumptions and trust boundaries.