DETERMINISTIC REPLAY
Deterministic
replay
When something matters, you need the state as it was — not a summary, and not a fresh answer to an old question. Deterministic replay reconstructs recorded state from retained events, so a past moment can be revisited as recorded.
THE PROBLEM
Retellings drift
Ask a system to describe the past and you get a retelling: compressed, reordered, sometimes improved. Ask it to re-decide and you may get a different answer, because models, policies and context change.
Replay avoids both. It does not re-decide and it does not summarize. It returns the state that was recorded.
WHAT REPLAY IS
Reconstruction of recorded state
Replay walks the retained events in order and rebuilds the state they describe: what source was captured, what context was bound, what decision was recorded. The evidence playground demonstrates this with a sample history — evidence added, context bound, decision recorded — where each checkpoint shows the context recorded at that moment.
The result is the state as recorded, not a retelling of it.
WHAT REPLAY NEEDS
Conditions, not magic
Replay needs retained events: without the log, there is nothing to reconstruct from. It needs the recorded lineage — source, context, policy and receipt references — so the rebuilt state is traceable to what was recorded. It needs tenant scope respected at every boundary, so replay shows only the workspace it belongs to. And hash checks against a trusted reference are what let a reviewer trust that the rebuilt state matches the recorded one.
Where events are missing, replay can only go as far as the record goes. It does not invent what was never recorded.
WHAT REPLAY IS NOT
Not the same model answer again
Replay returns stored memory state. It does not guarantee identical model output: re-running a model on old inputs can produce a different answer, and that is not a failure of replay — replay never promised the model's answer, only the recorded state.
Replay is also not recovery. It does not restore lost data. If events were never retained, or storage failed, replay cannot bring them back — which is why backups and recovery procedures remain necessary.
A WORKED EXAMPLE
Walk the sample history
The evidence playground's replay section moves through a sample history in three checkpoints: evidence enters the record, the context is bound, the decision is recorded. Each checkpoint shows the context recorded at that moment, and the note is explicit: replay concerns stored memory state, not a guarantee of identical model output.
Samples are labeled. The demonstration shows the mechanism; it is not a production replay.
RELATED GUIDES
Continue with another guide
YOUR NEXT DECISION
Build with memory.
Stand behind the evidence.
Give consequential AI a history you can inspect.
