Audit-Closed AI Scientist Protocol

AI scientist systems and self-driving laboratories require governance protocols that are statistically valid under continuous monitoring, auditable without privileged operators, and robust to physical-world failure modes. We present an audit-closed protocol in which every decision is a deterministic function of a public transparency log, with all semantic claims grounded in explicitly declared observable contexts. The formal core integrates typed stochastic observation interfaces, tolerance-aware predictive semantics, and non-vacuous coherence auditing on finite acyclic proof graphs. To resolve the deterministic-pipeline tension, we combine explicit deterministic-depth controls with a Lipschitz observation-translation theorem that converts latent Wasserstein closeness into observable totalvariation guarantees. For sequential inference we use always-valid e-processes with gate-wise alpha accounting, robust score sanitization plus integrity monitoring, and anti-griefing novelty dynamics with signal-qualified eligible-cost updates. For adaptive experimentation we provide two regimes: a conservative fixed-ratio IPS test and a variance-adaptive predictable betting process that preserves optionalstopping validity while avoiding high-ratio evidence paralysis. For physical deployment we introduce hierarchical primary/secondary sentinels, spoofing-aware coherence gates, and auditable calibration transitions that separate sensor aging from sentinel degradation. For long-horizon autonomy we use drift-mode state transitions, exogenous-only root-memory decay, and subgraph-local fault localization with familywise error control. We prove system-level guarantees covering adaptive optional stopping, variance-adaptive evidence growth, Byzantineresilient aggregation, legal-spam stability, drift-localization safety, and mixture re-anchoring coherence. A fully reproducible synthetic benchmark confirms improved availability under outliers, high-ratio contextual exploration, root-debt erosion attacks, and sentinel aging while maintaining conservative false-acceptance behavior; the released manifest includes cryptographic digests and an automatic manuscript-value consistency check.