Months after the fact, someone disputes what the agent did — or claims the rules would have allowed something else. Replay settles it: the recorded inputs run through the recorded policy again, and either the same verdicts and hashes come out, or the discrepancy is exposed. Nobody's memory is consulted.
Inputs, evidence references, the acting authority, and — critically — the exact policy version that evaluated the call: same inputs · same policy vN. Replay is possible because the seal captured the decision's ingredients, not a summary of them.
What replays deterministically is the governed part: which rules fired, what verdict resulted, what was sealed. The question replay answers is "was this allowed, on this evidence, under these rules?" — the question disputes and audits actually turn on.
The replayed verdicts and hashes must match the sealed ones exactly. "Close" is a failure. That hard edge is what lets a replay end an argument instead of starting a new one — see it run in the live tool-call demo.
The record and the rules are self-consistent: the sealed history is exactly what the recorded policy produces from the recorded inputs. The dispute moves from "what happened?" — now settled — to the human question of whether the policy itself was right.
Something real is wrong — a tampered record, a misdocumented policy version, a non-deterministic gate — and the divergence pinpoints it. Like a failed chain verification, a failed replay is a finding to investigate, never a discrepancy to paper over.