Generator-Critic-Referee: Adversarial Self-Review Against a Scored Rubric

Run a three-role self-review loop on a single task. TASK: [TASK]. The only definition of quality is this rubric: [RUBRIC]. MAX ROUNDS: [MAX ROUNDS].

Play three distinct roles each round and keep them honest about disagreeing:
1. GENERATOR - produces the best attempt (round 1) or a targeted revision (later rounds).
2. CRITIC - adversarial by design. Vague praise is banned. Score the draft on EVERY rubric criterion 0-5 with a one-line justification each, then list each concrete, fixable defect with the exact change that would fix it. Judge ONLY against the rubric - not personal taste.
3. REFEREE - neutral. If every criterion meets its pass bar, ACCEPT and stop. Otherwise order a REVISE and hand the Generator only the flagged defects to fix. Record a one-line 'lesson' so a fixed defect is never reintroduced.

Think before round 1: restate what 'excellent' looks like for THIS task in one sentence, so the Critic has something to measure against. Stop at MAX ROUNDS even if imperfect, and state exactly what still falls short.

OUTPUT FORMAT: per round a compact block {"round":n,"scores":{},"defects":[],"lesson":"","referee":"revise|accept"}. Then 'FINAL:' with the accepted output in full, then 'Known weaknesses:' (one line). Do not reprint rejected drafts in full - only the review blocks and the final.