Field note /
Dry run first, with no provider calls, before any replay
Before replaying 120 saved requests, the replay tool checked the saved set, each request's identity and the fixed model versions. It made zero calls to the model provider.
How the finding changed
Belief
Once the saved replay pack existed, the next useful evidence would come from calling the provider.
Evidence
The executor could validate the full frozen pack, its exact checksum, the request and dispatch-order identities, model pinning, variant keys, matching requested and resolved models, and the selection, without running a semantic request.
Contradiction
Skipping that check would spend calls before confirming that the intended inputs and settings were the ones selected.
Correction
Stage provider replay like a release: frozen inputs first, a zero-call check second, a small four-variant test run next, and broader runs only after inspection.
Current claim boundary
The dry run made zero provider calls, produced no result or receipt files, and used no fresh evaluation tasks. It shows the executor is ready to run. No semantic response or mechanism effect exists yet.