Evaluations

Use the real application examples. They require an explicitly selected connected account/model and store actual responses, timestamps, runtime metadata and reported usage in a private receipt. No response is prerecorded.

Assess transport completion and answer quality separately:

• Brand directions must retain the supplied product name, audience and constraints.
• Literature answers must distinguish paper findings from proposed experiments and cite only supplied sources. Citation presence alone does not prove support.
• Email drafts must preserve release status and pending work without claiming to have sent mail or completed adoption.

Use an inexpensive explicitly selected model, public or permitted inputs and one attempt per request. Never select an alternative account or billing mode after a failure. Model output is nondeterministic; record and review the actual answer.

The old synthetic evaluation runner has been removed. Earlier reports do not qualify current connections. See the recorded real Prometheus checks for exact models and limitations.