A proof packet is the typed, signed, replayable record an AI agent must produce before a consequential action executes: the inputs read, the policy clauses consulted, the reasoning chain followed, the systems the action will touch, and the human or policy that authorised it — chained into a tamper-evident ledger.
The packet is typed so it can be validated mechanically, signed so it can be attributed, and replayable so that months later an auditor can re-run the determination exactly as it ran — same facts, same rules, same verdict.
Proof packets change the economics of oversight. Instead of reconstructing what a system did from scattered logs, a reviewer opens one artefact that was built for exactly that review.
This vocabulary is implemented, not aspirational — it describes how the MAIA substrate actually runs. See the platform or read the research.