Ai eval regression tester
Skill sisodiabhumca/agent-skills/skills/ai-eval-regression-tester
Production-Ready Agent Skills : product analytics, growth experiments, CRM, research synthesis, postmortems, data contracts, SaaS spend, compliance, architecture maps, and LLM eval and many more.
npx -y skills add sisodiabhumca/agent-skills --skill ai-eval-regression-testerAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 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
Use to run a regression eval suite over an LLM application — fixed test cases with deterministic graders (exact match, JSON schema, regex, embedding similarity, LLM-as-judge). Compares the candidate model/prompt against a baseline and gates a release on pass rate + per-tag thresholds.
SKILL.md
1.9 KB, as published. Nobody here has run it
AI Eval Regression Tester
When to invoke
- "Run the eval suite for the new prompt version."
- "Compare gpt-X vs the current baseline on our customer-support eval."
- "Block release if eval pass rate drops below 95%."
Inputs needed
- Eval YAML / JSONL — list of cases with
input, expected outputs, graders, tags. - Candidate runner — Python callable / HTTP endpoint that takes input and returns output.
- Baseline run — JSONL of prior outputs (optional, for diffing).
- Pass thresholds — overall and per-tag.
Workflow
- Load cases.
- Run candidate over each case (parallelized).
- Grade with configured graders.
- Aggregate — pass rate overall + per tag + diff vs baseline.
- Gate — exit 1 if any threshold fails (CI-friendly).
- Report — Markdown + JSONL of every case for diffing.
Eval case schema
- id: refund_basic
tags: [refund, policy]
input: "How do I request a refund after 60 days?"
graders:
- type: contains
values: ["return policy", "support team"]
- type: not_contains
values: ["sure thing", "lol"]
- type: regex
pattern: "policy"
- type: json_schema
schema:
type: object
required: [answer, citation]
Guardrails
- Always require deterministic graders before LLM-as-judge.
- LLM-judge results must include the judge's reasoning verbatim.
- Per-case results must be logged to JSONL for forensics.
- No silent retries on grader failures.
Reference code
run_eval.py runs cases in parallel, grades, diffs vs baseline, and exits with the right code.