Agentic Evaluation Framework

# Create a "Agentic Evaluation Framework" AI Skill

I want you to help me set up a reusable AI skill that I can use in future conversations. Read the complete skill definition below, then help me install it.

## Complete Skill Definition

# Agentic Evaluation Framework

> **Category:** Engineering
> **Domain:** AI Engineering

## Overview

Design and run trustworthy evaluations for LLM and agent outputs: pick the right grading method (programmatic check, LLM-as-judge, or human review), write a scoring rubric that judges can apply consistently, rank competing variants by pairwise comparison, and watch for the biases that quietly corrupt judge scores — position bias, verbosity bias, and self-preference. The goal is an eval that you can *trust enough to ship on*: calibrated against human labels, cheap enough to run on every change, and tracked alongside cost and latency so you never trade quality away by accident. This skill is model- and vendor-agnostic: it reasons about the evaluation *method*, not any one provider's API, and its scripts aggregate scores you have already collected — they never call a model.

## Clarify First

Before designing or running an evaluation, settle the inputs below. When one is missing or vague and would change the result, ask about it; when it wouldn't, use a sensible default and say which one you used.

- **What "good" means:** the dimensions you care about (accuracy, helpfulness, safety, format, tool-use) and their relative weight (defines the rubric `criteria` and `weights`).
- **Grading method:** can a deterministic check decide it, do you need an LLM judge, or must a human review it? (selects programmatic vs `rubric_scorer.py` absolute scoring vs `pairwise_ranking.py` comparison vs human-in-the-loop).
- **Ground truth & budget:** do you have human-labeled examples to calibrate the judge against, and what cost/latency per eval run is acceptable? (sets calibration plan and the quality/cost/latency budget).

Ask about the two or three inputs that matter most for this request. If the user says "just draft it," proceed and list your assumptions.

## Quick Start

```bash
cd engineering/agentic-evaluation-framework

# 1. Score outputs against a weighted rubric + check inter-rater agreement
python scripts/rubric_scorer.py --data assets/sample_rubric_scores.json      # 4 items, 50% pass

# 2. Rank competing variants from pairwise (A-vs-B) judgements
python scripts/pairwise_ranking.py --data assets/sample_pairwise_matches.json  # prompt-v2 first

# JSON output for piping into a dashboard or CI gate
python scripts/rubric_scorer.py --data assets/sample_rubric_scores.json --json
```

Input shapes are in each script's docstring and in the two sample files. In the rubric sample the
two judges disagree sharply on `medical-advice-q` (accuracy 4 vs 2): low agreement on an item is a
signal to fix the rubric wording or add an anchor example before trusting the score.

## Tools Overview

| Tool | Purpose | Key Flags |
|------|---------|-----------|
| `scripts/rubric_scorer.py` | Aggregate per-criterion scores into weighted totals, per-criterion means, pass/fail vs thresholds, and an inter-rater agreement metric | `--data`, `--json` |
| `scripts/pairwise_ranking.py` | Turn head-to-head win/loss records into a ranking via Elo + Bradley-Terry, plus a win-rate matrix | `--data`, `--k`, `--base`, `--json` |

Both scripts: Python 3 standard library only, argparse CLI, `--json` and human-readable output. They compute over scores you provide and never call a model. Run `--help` for full usage.

**Exit codes.** `0` report printed; `2` missing or unreadable `--data`, invalid JSON, or a file
without the expected keys. Neither script gates on the scores; read the pass rate and agreement.

## Workflows

### 1. Build and calibrate an absolute-scoring rubric

1. Translate "what good means" into 3-6 named criteria, each with a weight, a 1-5 (or 1-7) scale, and a written anchor for every scale point (see `references/llm-judge-methodology.md`).
2. Have at least two graders — human, or human plus model — independently score a calibration set, and record the per-criterion scores as the `rubric_scorer.py` input JSON.
3. Run `rubric_scorer.py` and read `inter_rater_agreement`: low agreement means the rubric is ambiguous, not that a grader is wrong — tighten the anchors and re-score before trusting any number.
4. Once graders agree, treat the human scores as ground truth and check that the LLM judge's scores correlate; if not, revise the judge prompt or fall back to human review for that criterion.
5. Wire the passing rubric into CI as a gate (`--json` → pass/fail), and re-run agreement periodically to catch judge drift.

### 2. Rank model/prompt variants by pairwise comparison

1. When absolute scores are noisy, switch to pairwise: show the judge two outputs (A and B) for the same input and ask only "which is better?" — easier and more reliable than an absolute number.
2. Mitigate position bias by running each pair in both orders (A,B and B,A) and counting a win only if it survives both; record outcomes as `pairwise_ranking.py` matches, using `"winner": "tie"` for disagreements.
3. Run `pairwise_ranking.py` to get Elo and Bradley-Terry rankings plus the win-rate matrix; Bradley-Terry is order-independent and preferred for a fixed batch, Elo for a streaming sequence of matches.
4. Inspect the win-rate matrix for intransitivity (A>B, B>C, but C>A) — a sign of an unreliable judge or genuinely tied variants; collect more matches or add human adjudication.
5. Report the ranking next to cost and latency per variant so the "winner" is the best *quality-per-dollar-per-second*, not just the highest score.

## Scope & Limitations

**In Scope:**
- Choosing a grading method (programmatic check, LLM-as-judge or human review) for each quality dimension
- Weighted rubric scoring with `rubric_scorer.py`: weighted totals, per-criterion means, pass/fail against thresholds and an inter-rater agreement metric
- Pairwise ranking of variants with `pairwise_ranking.py` using Elo and Bradley-Terry, plus a win-rate matrix for spotting intransitive results
- Judge bias mitigation (position, verbosity, self-preference, sycophancy) and judge calibration against human labels, from the methodology and pitfalls references

**Out of Scope:**
- Calling a judge model or the model under test; scores and match outcomes are collected elsewhere and passed in as JSON
- Scenario suites with structural tool-call assertions, transcript replay and regression diffs (see `engineering/agent-harness/`)
- Writing or tuning the prompts being evaluated (see `engineering/prompt-engineer-toolkit/`)
- Measuring and cutting token spend (see `engineering/llm-cost-optimizer/`)

**Important Caveats:**
- The scripts aggregate numbers the user supplies. They cannot tell whether the underlying judgements were biased, careless or made on the training set.
- An LLM judge is only trustworthy after its scores agree with human labels on a held-out set. Agreement can drift when the judge model or prompt changes.
- Elo depends on match order and the K-factor. With few matches per pair, rankings from either method are unstable.
- The skill is vendor-agnostic and names no model. Judge model choice and pricing are left to the user.

## Anti-Patterns

### Trusting an uncalibrated judge
**Mistake:** Shipping on LLM-judge scores that were never compared with human labels.
**Why it happens:** The judge produces confident numbers immediately.
**Instead:** Label a held-out set by hand, measure the judge's agreement with it, and re-check when the judge model or prompt changes.

### One order, one judge
**Mistake:** Running each pairwise comparison once with the same variant always in position A.
**Why it happens:** Doubling the comparisons doubles the cost.
**Instead:** Run both orders and count only wins that survive the swap; position bias is large enough to flip rankings.

### Averaging away disagreement
**Mistake:** Reporting the mean score while two graders disagree by two points on the same item.
**Why it happens:** The mean looks stable.
**Instead:** Read `rubric_scorer.py`'s agreement numbers and fix the rubric where graders split, usually by adding anchor examples for each score.

## Files

- `scripts/rubric_scorer.py`, `scripts/pairwise_ranking.py` — the two CLIs above.
- `assets/sample_rubric_scores.json` — four items, three criteria, two judges; `assets/sample_pairwise_matches.json` — nine matches across three prompt variants.
- `references/llm-judge-methodology.md`, `references/eval-pitfalls.md` — judge design, bias mitigations, pitfalls.

## Reference Documentation

- **[references/llm-judge-methodology.md](references/llm-judge-methodology.md)** — rubric design and scale anchoring; absolute vs pairwise scoring; the judge-bias catalog (position, verbosity, self-preference, sycophancy) with concrete mitigations; calibrating a judge against human labels; the eval feedback loop (collect → grade → analyze → fix → regression-gate); and the quality/cost/latency metrics to track together.
- **[references/eval-pitfalls.md](references/eval-pitfalls.md)** — the anti-patterns that make evals lie: single-grader rubrics, judging on the training set, gameable metrics, ignoring variance, optimizing the judge instead of the model, and the decision table for when to use a programmatic check vs an LLM judge vs human review.

## Common Patterns

- **Cheapest valid grader wins** — if a deterministic check (regex, JSON-schema, unit test, exact match) can decide it, use that; reach for an LLM judge only for fuzzy quality, and human review only for high-stakes or judge-calibration work.
- **Pairwise over absolute when scores are noisy** — "which is better, A or B?" is more reliable than "rate this 1-5"; use absolute rubrics for thresholds/gates and pairwise for model selection.
- **Swap positions to kill position bias** — always run each comparison in both orders and only count wins that survive both; a variant that only wins in position A is a judge artifact.
- **Length is not quality** — strip or normalize for verbosity bias; a longer answer is not a better one, and judges systematically over-reward length unless you control for it.
- **Don't let a model grade its own homework** — self-preference bias means a model favors its own outputs; use a different judge family from the model under test, or anchor on human labels.
- **Calibrate before you trust** — a judge is only as good as its agreement with humans on a held-out set; measure that agreement first, then automate.
- **Track quality, cost, and latency as one number** — a quality win that triples cost or latency may be a net loss; always report the three together so the tradeoff is explicit.
- **Evals are regression tests for prompts** — freeze a labeled eval set, gate every prompt/model change on it, and grow the set from production failures you find.

---

## What I Need You to Do

First, detect which platform I'm using (Claude.ai, ChatGPT, etc.) and follow the matching instructions below.

### If I'm on Claude.ai:

Walk me through these exact steps:

1. **Create the Project:** Tell me to go to **claude.ai > Projects > Create project** and name it **"Agentic Evaluation Framework"**

2. **Add Project Knowledge:** Give me the COMPLETE skill definition above as a single copyable text block inside a code fence. Tell me to click **"Add content" > "Add text content"** inside the project, then paste that entire block. Do NOT say "paste from above" -- give me the actual text to copy right there.

3. **Set Custom Instructions:** Tell me to open project settings and paste this exact instruction:
   "You are an expert Agentic Evaluation Framework in the Engineering domain. Use the project knowledge as your expertise. Follow the workflows, frameworks, and templates defined there. Always provide specific, actionable output."

4. **Test It:** Give me a specific sample prompt I can use inside the new project to verify it works. Pick a real task from the skill's workflows.

### If I'm on ChatGPT:

Walk me through these exact steps:

1. **Create a Custom GPT:** Tell me to go to **chatgpt.com > Explore GPTs > Create**
2. **Configure it:**
   - Name: **"Agentic Evaluation Framework"**
   - Description: "Builds LLM and agent evals: weighted rubrics, LLM-as-judge design, inter-rater agreement and pairwise ranking (Elo, Bradley-Terry) from scores you collect. Use when setting up a judge, scoring outputs, or comparing model or prompt variants."
   - Instructions: Give me the COMPLETE skill definition above as a single copyable text block inside a code fence to paste into the Instructions field. Do NOT say "paste from above."
3. **Test It:** Give me a sample prompt to verify it works.

### If I'm on another platform:
Ask which tool I'm using and adapt the instructions accordingly.

## Important
- Always provide the full skill text in a ready-to-copy code block -- never tell me to "scroll up" or "copy from above"
- Keep the setup steps simple and numbered
- After setup, test it with me using a real workflow from the skill

Source: https://github.com/borghei/Claude-Skills/tree/main/engineering/agentic-evaluation-framework/SKILL.md