Replay console

Change a policy knob, replay the episodes in the store, see the delta before shipping.

Same action → observed outcome. If the replayed action matches what actually ran, we reuse the real result. No model is consulted.
Different action → modelled. The counterfactual is drawn from scoreRecovery under the replayed action — the same scorer production decides with — seeded per episode so a replay is reproducible.
The split is displayed permanently. A replay console that hides the observed/modelled ratio is a fiction generator.
Deltas are against what actually happened, over the same episodes, not against a second simulation.