Judge Loop

Produce → independent judge scores 1–10 with evidence → revise the lowest criterion → repeat. Default 3 rounds, bar ≥ 8. The user overrides both by saying so ("5 rounds", "bar 9", "until it clears 8").

Core rule: the score is a claim that needs evidence, and the judge is never the producer. A number without a named strength or defect is not a score.

1. Frame (once, before round 1)

Write the rubric to <scratchpad>/judge/RUBRIC.md and freeze it. Never loosen a criterion later to make the number go up; if the rubric was wrong, say so in the report and restart the loop.

Pick criteria from the task, not from a template. Starting points:

Ask: what would make the user reject this on sight? That becomes a hard floor. Requirements stated in the task ("must handle German", "stable output") are hard floors or criteria, never ignored because the producer didn't address them.

2. Round loop

round n:
  produce / revise   (producer agent or yourself — fixes the lowest-scoring criterion first)
  judge              (fresh agent, never the producer, gets: RUBRIC.md + artifact + previous JUDGE-n.md)
  record             <scratchpad>/judge/JUDGE-n.md
  overall ≥ bar AND no criterion < 6  → PASS, continue the real work
  else if n < max rounds              → next round
  else                                → STOP, report the gap (see §4)

• Judge model: Opus by default (taste); Fable for high-stakes or subtle work. Sonnet only for purely mechanical rubrics. Never Haiku.
• Judge sees the previous report so its scale stays consistent across rounds and it must say whether each criterion moved up, down, or held. A revision that raises one criterion and silently drops another is not progress.
• Several candidates in one round (A/B variants, a batch of assets): one judge call, all candidates in one context, ranked. Same judge for the whole round.
• Verify before judging when the floor is checkable: run the gate, open the screenshot, load the render. Judge on the real output, not the producer's summary.

3. Judge prompt (fixed shape)

You are the judge, not the author. Score the deliverable against RUBRIC.md verbatim.
For each criterion: score 1–10, then ONE line of evidence naming the specific
strength or defect (file:line, element, sentence, region of the image).
Check every hard floor first; any failure caps overall at 3.
Overall: holistic 1–10, 8 = an expert would ship it as-is.
Then list the 3 fixes that would raise the score most, ranked, each tied to a criterion.
Do not fix anything. Do not praise effort. A 7 is a good draft; an 8 needs no excuses.
Output the table below and nothing else before it.

Output contract (the judge fills this; missing evidence = invalid report, rerun):

| Criterion | Score | Δ vs prev | Evidence |
| ...       | n/10  | +1/–/=    | specific, checkable |
| Hard floors | pass/FAIL | | which one |
| **Overall** | **n/10** | | |
Top fixes: 1. … (criterion) 2. … 3. …
Verdict: PASS / REVISE

4. Report to the user

One table, then the verdict:

| Round | Overall | Lowest criterion | What changed |
| 1 | 5.5 | correctness 3 (umlauts dropped) | — |
| 2 | 7.5 | tests 6 | added transliteration |
| 3 | 8.5 | — | property test for stability |
Verdict: PASS at round 3 — continuing with <next step>.

Failed after max rounds → say "did not clear the bar", the last score, the criterion that blocked it, and what a round 4 would attack. Never round 7.5 up. Never continue the real work on a REVISE verdict without the user saying so.

Red flags — the loop is being gamed

• Judge and producer are the same agent or the same context
• Rubric edited after round 1
• A score with no evidence line, or evidence that is a restatement of the score
• "8 because it improved a lot" — improvement is Δ, not the score
• Overall above the lowest criterion by more than 3 without a stated reason
• Threshold passed by averaging away a hard-floor failure
• Round 3 declared PASS to finish, with the same defect still named in the evidence