Conversational role is not producer identity in evaluator evidence

Preserve producer provenance through every evaluator projection; a conversational agent role does not establish that a model authored the message. Backend-generated utterances can appear on the agent side of a transcript. If an adapter drops source and ownership fields, a downstream judge may apply model-only instructions to backend behavior and produce a false attribution.

Carry captured ownership and causal tool links through the typed source model, normalized evidence and rendered judge request. Test the final request, not only deserialization. Do not infer missing ownership from wording or role, and do not silently resolve contradictory provenance. Backend ownership also does not prove that recovery was correct: apply the policy belonging to the established producer and keep independently supported tool errors separate.

Applies when: evaluators judge conversations containing messages generated by models, tools and backend workflows. Keywords: evaluator speech ownership provenance backend model attribution source projection