The audit pack

Six months after go-live, someone with authority will pick cases and ask: what did the system do, on what evidence, under which approved version, who decided, and can you prove the record has not changed since? Application logs and traces will not answer that: they are sampled, short-lived, mutable and written for engineers. An audit log is a separate, complete, structured, append-only record of decision events. Each record holds references and fingerprints rather than raw personal data, chains to the previous record by hash, and is periodically sealed by a signed checkpoint stored where the system's operators cannot write, ideally on WORM storage. From those records you reconstruct, never replay, each decision into an explanation pack: what the AI was allowed to do, what evidence it used, what it proposed, what the person decided and what was sent. For LLM agents the explanation is the evidence and the controls, not saliency maps and not the model's own narrative of its reasoning.

• Audit evidence is a product with a customer: the auditor. Design its schema, retention and access deliberately.
• Log references and fingerprints, not copies. Content stays in the system of record under its retention rules.
• Tamper evidence needs an anchor outside the log. A hash chain the operators can recompute proves nothing on its own.
• Reconstruct from evidence; do not replay the model. Models are non-deterministic and retire.
• Explain with evidence and controls. Chain-of-thought text is not a faithful account of why a model produced an output.