Grade
Agent evals on autopilot: find quality bugs in your AI agent, ship a targeted fix, and prove it on a held-out set. Zero-dependency Agent Skill + CLI.
npx -y skills add avnath13/evalpilot --skill gradeAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 20 days oldThe repository was created 20 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Grade agent traces against rubrics using deterministic checks and binary LLM-as-judge (critique-then-label), no external eval framework required. Use when the user wants to score, judge, or evaluate agent outputs against criteria. Do NOT use to author a judge or validate it against human labels (use validate-evaluator/calibrate), and do NOT use to fix failures (use analyze-failures then optimize). Prefer a deterministic check over a judge whenever the criterion is objective (fix-before-eval).
SKILL.md
3.5 KB, as published. Nobody here has run it
grade: traces × rubrics → verdicts (the zero-dep heart)
Output: runs/<runId>/verdicts.jsonl, each line conforming to evalpilot/schemas/verdict.schema.json.
You, the host agent, ARE the autorater. There is nothing to install.
Every rubric answers one of two questions: correctness (did it do the right thing?) or quality/safety (was the way it did it safe and usable?). A serious suite covers both.
Procedure
For each (trace, rubric) pair, where rubrics resolve to evalpilot/rubrics/<name>.rubric.md:
- Read the rubric frontmatter.
kind: deterministic | judge,severity,threshold. - Deterministic rubrics → write and run a small check (regex / JSON field compare /
numeric tolerance / "was tool X called"). No model call. Example:
refund-hallucinationcross-checks any dollar figure inoutputagainst theresultoflookup_refundinsteps[]. Fast, exact, reproducible. - Judge rubrics → the rubric body IS the judge prompt. Read the trace + the rubric
criteria and emit critique-then-label:
{critique, result: Pass|Fail}plusreason,evidence,cluster_hint. Binary only, no Likert. - Always fill
reason,evidence, andcluster_hint. These are what make stage 4 clustering and stage 5 fixes targeted rather than guesswork. Quote the exact offending span inevidence. - Write one verdict object per line. Print per-rubric aggregate:
correctness 0.71 (65/92) | tool-use 0.88 | refund-halluc 0.14 (13 fails).
Calibration (do not skip for judge rubrics)
An unvalidated judge is a ruler of unknown length. Validate before trusting its scores
(python3 -m evalpilot validate):
- Label 30-50 cases spanning good / bad / borderline, ideally with two domain labelers. Where they disagree, discuss and resolve, if you can't resolve it, the rubric is underspecified: fix the rubric and relabel.
- Measure TPR/TNR vs human consensus (not raw accuracy); apply Rogan-Gladen bias correction. Gate: >0.90 trust · 0.80-0.90 iterate the prompt · <0.80 not ready.
- The calibration set becomes the judge's own regression test, re-run it whenever you edit the judge prompt and keep agreement above the bar.
Judge failure modes (anti-patterns to control for)
- Position bias, judges prefer the first of two outputs. Randomize order.
- Length bias, longer looks more complete. Control for it in the rubric.
- Over-lenience, helpful-trained models call "Acceptable" outputs "Good"; calibration surfaces this.
- Phrasing sensitivity, small rubric wording changes shift scores. Version the judge prompt.
- Model drift, a judge-model upgrade shifts scores unchanged rubric. Recalibrate after any judge-model change.
Guardrails
- Be a harsh, specific grader. Vague "looks fine" verdicts are useless downstream.
- Same trace + same rubric should grade the same way, prefer deterministic checks wherever the criterion is objective (a healthy suite is ~60-70% code-based).