Session: {{ session_id }}
Iteration: {{ iteration }}
Task: {{ task }}
Transcript so far:
{{ transcript }}
Latest assistant response:
{{ latest_response }}
Return JSON with:
- `verdict`: `"pass"` if the response advances the task acceptably, `"revise"` if it should be regenerated with feedback.
- `reasoning`: one short sentence justifying the verdict (not chain-of-thought).
- `critique`: when `revise`, a concrete actionable instruction for what the generator should do differently. Empty string when `pass`.
- `confidence`: float in [0.0, 1.0] reflecting how certain you are. Low confidence with `revise` will be treated as a soft warning; high confidence triggers the revision.