Critical decisions should be reconstructable from their inputs, versions, provenance, risk state, authority, execution state, and evidence.
An incident review cannot rely on a model’s current behaviour, a dashboard’s latest state, or a person’s memory of what the system probably saw. AI-native financial infrastructure needs a disciplined way to reconstruct the decision path as it existed when consequence was possible.
Replayability is that capability. It is not the claim that the past can be recreated perfectly. Markets, providers, networks, and models may not reproduce the same external conditions. The objective is narrower and more rigorous: reconstruct the governed record well enough to explain how evidence became a proposal, how controls and authority treated it, what execution state followed, and what outcome was observed.
Replay begins with captured context
A useful replay needs stable references to the inputs and their origin, the model and configuration that produced an inference, the rules and limits that were effective, the authority context, the instruction presented to the execution boundary, and the acknowledgements and reconciled state that followed.
Timing is part of that context. A data point available after authorization cannot be treated as if it informed the original decision. A rule updated later cannot be substituted silently for the rule that actually governed progression. Replayability therefore depends on versioned components, explicit state transitions, and provenance captured when events occur.
Different replays answer different questions
An evidentiary replay asks what the system knew and recorded. A control replay evaluates the captured proposal against the rule set effective at the time. A comparative replay may apply a newer model or rule set to historical evidence to study how behaviour would differ. These exercises are valuable only when their purpose is labelled clearly.
Confusing them creates false certainty. Re-running today’s model is not proof of what yesterday’s model produced. Applying a revised limit is not evidence that the old decision complied. A high-assurance replay preserves the historical account and treats later evaluations as distinct analytical results.
Why AI increases the need
Probabilistic systems can change through model revisions, provider updates, calibration shifts, prompts, tools, and configuration. Even when an input appears identical, output may differ. The system should therefore preserve not only the final recommendation but the relevant model identity, configuration, source references, uncertainty representation, and evidence needed to interpret it.
Deterministic layers benefit as well. Replaying qualification, risk, and authorization transitions can reveal whether policy was implemented as intended, whether evidence freshness was evaluated correctly, and whether an execution attempt remained inside its mandate.
Replay supports controlled learning
Incident reconstruction is the immediate use, but replayability also supports model evaluation, risk analysis, reproducibility, and controlled improvement. Teams can compare historical and revised components without mutating the authoritative record. They can identify where uncertainty entered, which boundary contained it, and whether an observed outcome reflects intelligence, control, authority, execution, or external-provider behaviour.
This is why replayability depends on both provenance and attestation. Provenance connects the decision path; attestation closes the operational loop. Replay turns that evidence into a reconstructable account without dissolving the distinction between what happened and what a later system would have done.