Extraction Provider Response Facts
What the provider said about its own response.
Every field is present only when the provider reported it. None of it is inferred from ExtractionRequestedModelConfig: a run that asked for gpt-x and got no model name back records a null responseModel rather than echoing what it asked for, because the whole point of the field is telling those two cases apart.
Response bodies, message content, and SDK objects are not stored — these are the identifiers and the classification, which is what an incident needs to correlate with the provider's own logs.
EXPERIMENTAL. The shape may still change while extraction runs (DICE #67) land.
Properties
Why the provider stopped generating, in its own vocabulary
The provider's id for the response, for correlating with its support logs
The model the provider says answered
The provider's fingerprint for the backend configuration that served the call