Observable-Only Audit Gate for Non-Markovian Components in AI Agents under Partial Logging

This paper introduces an observable-only audit gate for AI agents that operate under partial logging and may contain non-Markovian components (e.g., hidden state, delayed outcomes, or latent internal drivers). The core idea is to separate online decisions from replay-time certification using a small set of auditable events and a finalized prefix of the log stream. Online actions are constrained to depend only on a finalized, decision-eligible information set, while delayed evidence and outcomes are handled as PENDING until they are sealed and certified. The design specifies a minimal event vocabulary—such as AUDIT_PKT_PRE / AUDIT_PKT_POST, FINALIZE_ADVANCE, SEQ_ASSIGN, MEAS_BEGIN / MEAS_COMMIT, and CERT_BEGIN / CERT_END—to support deterministic replay, deduplication, and explicit gap handling across distributed collectors. To reduce ambiguity and cross-language drift, the paper uses canonical JSON (JCS / RFC 8785) and a string-safe rational numeric encoding, and it distinguishes content hashes from chain hashes via domain separation. The resulting audit stream enables stable reconstruction of what was known at decision time, while allowing post hoc measurement windows and certification results to be attached without creating causal loops. The paper is intended as a practical, implementation-friendly specification for auditability of AI agents under incomplete observability, emphasizing replay stability, counter-safety, and explicit lifecycle states for late arrivals, gaps, and delayed outcomes.