agentleFS
Sign inStart free
confident-aiconfident-ai/eval-skills/skills/eval-audit/SKILL.md

eval-audit

Skill11 starsChanged 44 days agoApache-2.0View on GitHub

Audit existing AI evaluation evidence, graders, datasets, and run reliability before trusting a score or optimizing an application. Not a generic code or security review.

5 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use validate-evaluator and write-judge-prompt.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • tim-osterhusSkill

    validate-evaluator

    Guardrail for validating LLM judges against human labels using data splits, TPR/TNR, disagreement review, and bias correction.

    Also in 3 projects that use this

  • eggboySkill

    write-judge-prompt

    Design LLM-as-Judge evaluators for subjective criteria that code-based checks cannot handle. Use when a failure mode requires interpretation (tone, faithfulness, relevance, completeness). Do NOT use when the failure mode can be checked with code (regex, schema validation, execution tests). Do NOT use when you need to validate or calibrate the judge — use validate-evaluator instead.

    Also in 3 projects that use this

The file

We split this skill into the rules it gives an agent

2 rules. 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 5 edited versions we read keep these.

  • 2

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read the operating agreement.

    line 12
  2. Distinctive:

    Trace validity: Do output, score, model identity, usage, and transcript correspond to the same execution?

    line 24

“Common” is measured: kept by at least 6 in 10 of the 5 edited versions of this file we could read. “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

Audit an existing evaluation

Read the operating agreement. Inspect artifacts; do not produce findings from a checklist alone.

For bundled local tools and exact commands, read toolkit setup. Run the workspace installer for this skill; it preserves existing customized files.

Gather evidence

Find the runner command, actual app entrypoint, cases, reference origins, full traces, human labels, grader definition/version, and raw results. Reuse connected observability tools or local exports. For the managed track read hosted workflow guidance.

Checks

  1. Task validity: Does the metric measure the desired outcome? Are failure categories observed or merely generic names? For an acting agent, does the check inspect resulting state?
  2. Data validity: Are inputs current, representative for the claimed rate, and legally retainable? Identify synthetic cases and model-generated reference answers. Check duplicates and shared conversations across splits.
  3. Trace validity: Do output, score, model identity, usage, and transcript correspond to the same execution? Inspect tool calls and retrieval where relevant. Missing content may be a capture-policy limit, not an app failure.
  4. Grader validity: Examine expert disagreements, false passes and false failures. Test empty output, an answer to a different question, a known pass, and a known fail. Verify case references before blaming the model.
  5. Harness validity: Induce a safe local error. It must be reported as an execution error rather than a capability zero. Check retry policy, timeouts, incremental writes, and resume identity.
  6. Comparison validity: Recompute the headline from raw rows. Compare dataset, judge, mode, app/configuration, and environment versions. Inspect error rates and missing-data denominators. Check that a parameter actually reaches the app and any subagents.
  7. Resolution: Estimate uncertainty and headroom against the smallest useful improvement. Repetitions over one generated index or memory store cannot measure the variance of rebuilding it.

Classify findings as blocking (invalid claim), important (weak evidence), or improvement. Each needs an artifact reference, observed consequence, and concrete fix. Mark unavailable evidence as unknown, not passed.

Repair without erasing history

Correcting a grader requires retaining old grades and regrading all compared stored outputs. Correcting app execution may require rerunning the baseline. Version the change, show whether rankings move, and return to the appropriate stage. Do not tune the app to work around a mislabeled reference.

Deliver the recomputed result, confidence limits, failure/error breakdown, prioritized findings, and the next executable step. If trustworthy, reuse the pipeline.

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.

Check your setup, free