agentleFS
youcislayoucisla/Agent-Foundry/skills/core/judge-integrity/SKILL.md

judge-integrity

Skill1 starsChanged 23 days agoMITView on GitHub

Use when a model judges model output: separate judge failure from a low score, run deterministic checks first, gate escalation, and keep scoring separate from generation.

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 1 is distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Every judge attempt returns a status: ok, unparseable, call_failed, or timeout.

    line 19
  2. Very specific:

    Run the task verifier. Stop on pass or fail only if the task declares the verifier authoritative.

    line 30
  3. Distinctive:

    Call the judge with a timeout.

    line 31

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

judge-integrity

A judge is a measurement instrument. An instrument that returns zero when it is broken is worse than no instrument.

Contract

  • Every judge attempt returns a status: ok, unparseable, call_failed, or timeout.
  • scores is present only when status is ok. A parse failure is never a zero.
  • Transport errors retry with backoff. Refusals and parse failures do not retry into another score.
  • A deterministic verifier runs before the paid judge. A definitive pass may skip the judge; unknown falls through.
  • Judge and evaluation calls pin temperature 0 and preserve structural markers.
  • Escalate to a stronger judge only below an explicit score threshold.
  • Store the judge prompt hash and status beside the score.

Procedure

  1. Parse the response with balanced-brace extraction; fences and nested objects are normal.
  2. Run the task verifier. Stop on pass or fail only if the task declares the verifier authoritative.
  3. Call the judge with a timeout. Record status and duration even on failure.
  4. Store scores, status, model, and prompt hash. Never overwrite the generation record.
  5. Re-score stored output with a separate command; do not re-run generation to re-judge.

Worked example

Three stored judge records: scored, unparseable, and call_failed.

{"status": "scored", "score": 7.5, "model": "judge-v2", "prompt_hash": "sha256:..."}
{"status": "unparseable", "score": null, "model": "judge-v2"}
{"status": "call_failed", "score": null, "error": "timeout"}

Anti-patterns

  • Mapping unparseable output to 0.0 and calling it a score.
  • Treating a provider outage as a low-quality answer.
  • Re-running the full generation to apply a new judge.
  • Pooling judge scores across model revisions or rubric versions.

References

  • status

Verification Checklist

  • A raising provider yields call_failed, not a score. (step 3)
  • Prose output yields unparseable, not a score. (step 1)
  • Nested JSON parses correctly. (step 1)
  • Re-scoring a stored execution performs no completion call. (step 5)
  • Every score has a status and a model fingerprint. (step 4)

Composes With

  • structured-output-contract — parsing and constraint precedence
  • execution-observability — judge status and request metadata
  • feedback-loop — calibration against human feedback

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.