Suggest the weaker answer + error tags + a draft rationale for a task the evaluator is grading — VERIFY, don't author. Guardrails:

• Anti-peeking: gated behind the evaluator's independent-answer commit (like GET /tasks/{id}/answers) — the suggestion describes the A/B answers, so it must not exist pre-reveal.
• Never applied server-side: the verdict/tags/rationale stay untouched; the client renders the suggestion as a tap-to-accept hint only.
• Low-confidence suggestions (< ASCLEPIUS_ASSIST_MIN_CONF, default 0.6) are HIDDEN — returned as skipped so the UI never nudges on an uncertain call.
• Degrades to skipped=True with no LLM key — manual labeling always works.

The access gate is resolved IN THE BODY rather than through Depends, because the task id arrives in the payload and a path-parameter dependency cannot see it. Same rule as require_task_access, same two branches — an applicant on their own examination, or the full chain for everybody else.