ADR 320 · Accepted
A run shows the model the provider named
Context
Every provider call of a run carries the run's uuid as its correlation id (ADR-153), and the run timeline joins the telemetry rows of that id. A row names the provider and model the call was sent to and, after a fallback, the ones that served it.
Since ADR-174 the same row also stores response_model: the model the provider named on its response. That need not be the model the call asked for. An alias such as gpt-5 resolves to a dated snapshot, and which snapshot answered is the fact that separates a model change from a skill or input change when two runs of the same configuration answer differently. TelemetryRepository::findByCorrelation() did not read the column, so the timeline could not show it.
Decision
TelemetryRepository::findByCorrelation() reads response_model and TelemetryCall carries it as responseModel.
A provider-call row of the timeline shows it as reportedModel when it is not empty and differs from the model the call was sent to: the serving model after a fallback, the requested one otherwise. This is the rule the serving pair already follows; a repetition of the name already shown is left out.
Nothing is derived from it. The timeline shows what the provider said; it does not decide whether the answer came from an acceptable model.
Consequences
● An administrator can read on the run which snapshot of a model answered each call, and tell two runs apart whose configuration and requested model are equal.
◐ The value is metadata like the rest of the row: an identifier, never content of the call. The timeline stays metadata only at every privacy level.
◑ The column is '' wherever the call produced no usage block: a cache hit, a failed call, the specialized operations and a plain streamed chat (see ADR-174). Where a provider's answer names no model, the adapter reports the requested id instead. The timeline shows nothing in both cases, which cannot be told apart from a provider that repeated the requested name.
◑ The value is what the adapter reports. GeminiProvider reports the model id of the request and never reads one from the response, so a Gemini call never shows reportedModel. The same holds for any provider that names the alias it was sent, not the snapshot behind it.