Validate Evaluator

Validate Evaluator

Calibrate an LLM judge against human judgment.

Respond in the user's language; keep metric names (TPR/TNR) and code in English.

Overview

1. (Optional) Label traces with the user if labels don't exist yet
2. Split human-labeled data into train (10-20%), dev (40-45%), test (40-45%)
3. Run judge on dev set and measure TPR/TNR
4. Iterate on the judge until TPR and TNR > 90% on dev set
5. Run once on held-out test set for final TPR/TNR
6. Apply bias correction formula to production data

Prerequisites

• A built LLM judge prompt (its companion skill write-judge-prompt, if installed, covers construction)
• Human-labeled data: ~100 traces with binary Pass/Fail labels per failure mode
  • Aim for ~50 Pass and ~50 Fail (balanced, even if real distribution is skewed)
  • Labels must come from a domain expert, not outsourced annotators — and not LLM-generated labels (that is circular validation: the judge would be graded against another model's opinion, not against ground truth)
• Candidate few-shot examples from the labeled data
• A Python environment with numpy and scikit-learn (sklearn). Optionally judgy (pip install judgy), the reference implementation from the AI Evals course by Hamel Husain & Shreya Shankar: https://github.com/ai-evals-course/judgy

Synthetic Fail examples (when real failures are too scarce to balance the set):

• Allowed in the dev set only — never in the test set.
• Report the synthetic fraction alongside every metric that includes them (e.g., "TNR 0.91, 40% of Fail examples synthetic").
• Once enough real failure traces accumulate, replace the synthetics and re-calibrate. Numbers measured on synthetic failures are provisional by definition.

Core Instructions

Step 0 (Optional): Labeling Workstation

If the user has traces but no labels, run this minimal flow before anything else:

1. Load the user's traces file (JSONL/CSV, one trace per row).
2. Present traces one at a time, trimmed to the span relevant to the failure mode. Ask: "Pass or Fail for <failure mode>? (pass / fail / skip / add a note)". One trace per question — never batch, never suggest a label before the user answers.
3. Record the label, plus any note (notes often become Pass/Fail definition refinements).
4. Append each label to a file in the user's project, e.g. evals/<failure-mode>/labels.jsonl — never inside this skill's own directory. If no file can be written in this environment, keep the running set in the conversation and output it as one copyable JSONL block at the end.
5. Stop at ~100 labels, or earlier at the user's budget; flag it if the set ends up heavily imbalanced.

The user (or their domain expert) is the labeler. Do not fill in labels yourself — see the red line above about LLM-generated labels.

Step 1: Create Data Splits

Split human-labeled data into three disjoint sets:

Target: 30-50 examples of each class (Pass and Fail) across dev and test combined. Use balanced splits even if real-world prevalence is skewed — you need enough Fail examples to measure TNR reliably.

from sklearn.model_selection import train_test_split

# First split: separate test set
train_dev, test = train_test_split(
    labeled_data, test_size=0.4, stratify=labeled_data['label'], random_state=42
)
# Second split: separate training examples from dev set
train, dev = train_test_split(
    train_dev, test_size=0.75, stratify=train_dev['label'], random_state=42
)
# Result: ~15% train, ~45% dev, ~40% test

Step 2: Run Evaluator on Dev Set

Execution constraint (highest priority): Run the judge via API using the pinned production judge model. Do NOT act as the judge yourself in-context — self-scored labels calibrate "you", not the judge, and invalidate every number downstream.

This applies identically whichever assistant engine is executing this skill (Claude, GPT, or other): write a script that calls the judge model's API for every dev example, and run it — or hand it to the user to run if this session cannot make API calls. Never substitute in-context judgments, not even "just to get a rough number".

Compare the judge's predictions to the human labels.

Step 3: Measure TPR and TNR

TPR (True Positive Rate): When a human says Pass, how often does the judge also say Pass?

TPR = (judge says Pass AND human says Pass) / (human says Pass)

TNR (True Negative Rate): When a human says Fail, how often does the judge also say Fail?

TNR = (judge says Fail AND human says Fail) / (human says Fail)

def judge_rates(human_labels, evaluator_labels):
    human, predicted = list(human_labels), list(evaluator_labels)
    if not human or len(human) != len(predicted):
        raise ValueError("Need nonempty, paired human labels and API predictions")
    if any(x not in ("Pass", "Fail") for x in human + predicted):
        raise ValueError("Invalid/missing verdict: resolve API errors before calibration")
    positives = human.count("Pass")
    negatives = human.count("Fail")
    if not positives or not negatives:
        raise ValueError("Both human classes are required; missing rates are unknown")
    tp = sum(h == "Pass" and e == "Pass" for h, e in zip(human, predicted))
    tn = sum(h == "Fail" and e == "Fail" for h, e in zip(human, predicted))
    return tp / positives, tn / negatives

tpr, tnr = judge_rates(human_labels, evaluator_labels)

Use TPR/TNR, not Precision/Recall or raw accuracy. These two metrics directly map to the bias correction formula. Use Cohen's Kappa only for measuring agreement between two human annotators, not for judge-vs-ground-truth.

Step 4: Inspect Disagreements

Examine every case where the judge disagrees with human labels:

For each disagreement, determine whether to:

• Clarify wording in the judge prompt
• Swap or add few-shot examples from the training set
• Add explicit rules for the edge case
• Split the criterion into more specific sub-checks

Step 5: Iterate

Refine the judge prompt and re-run on the dev set. Repeat until TPR and TNR stabilize.

Choose an iteration budget and acceptance threshold for the intended use before tuning. The targets below are starting heuristics, not universal guarantees.

Stopping criteria:

• Target: TPR > 90% AND TNR > 90%
• Minimum acceptable: TPR > 80% AND TNR > 80%

If alignment stalls:

Step 6: Final Measurement on Test Set

Run the judge exactly once on the held-out test set. Record final TPR and TNR.

If the test reveals problems, return to development; that exposed test set is no longer an untouched final holdout. Obtain a fresh real test set before the next final claim, and preserve the earlier failed measurement.

Step 7 (Optional): Estimate True Success Rate (Rogan-Gladen Correction)

Raw judge scores on unlabeled production data are biased. If you need an accurate aggregate pass rate, correct for known judge errors:

theta_hat = (p_obs + TNR - 1) / (TPR + TNR - 1)

Where:

• p_obs = fraction of unlabeled traces the judge scored as Pass
• TPR, TNR = from test set measurement
• theta_hat = corrected estimate of true success rate

Retain the raw estimate before clipping to [0, 1], and report any clipping. Do not apply this workflow when the denominator is non-positive or unstable near zero. Error rates must transfer to the production population: changes in task mix, language, or model require fresh validation. A balanced test set alone does not establish that transfer, and correction is an estimate, not known truth.

Example:

• Judge TPR = 0.92, TNR = 0.88
• 500 production traces: 400 scored Pass -> p_obs = 0.80
• theta_hat = (0.80 + 0.88 - 1) / (0.92 + 0.88 - 1) = 0.68 / 0.80 = 0.85
• Corrected point estimate is 85% under the stated assumptions; this is illustrative arithmetic, not a measured production result.

Step 8: Confidence Interval

Report uncertainty for the quantity actually estimated. The simple bootstrap in judgy currently resamples paired held-out labels/predictions while keeping the observed production pass fraction fixed. Its interval is conditional on that observed production sample, not a full interval for future production prevalence. The implementation was inspected in judgy/core.py.

Before calling it, use judge_rates above, reject empty or malformed production predictions, and preserve sample counts. Do not coerce an API timeout, unknown label, or skipped trace into Fail. Resolve missing predictions or report the run incomplete with its error count; dropping them silently changes the measured set.

For an interval targeting population prevalence, also propagate uncertainty from the production sample. For independent traces this can use separate resampling of production predictions plus paired test observations (stratified by human class when that matches test sampling). Clustered/repeated conversations need a matching cluster design. Record the resampling design, seed, number of draws, invalid-draw fraction and clipping. If correction is unstable, report it as unavailable and collect more representative labels; do not conceal instability by silently discarding most draws. Even a zero-width empirical interval does not prove a perfect judge. This skill does not bundle a general population-CI implementation.

For the conditional interval, use judgy (pip install judgy, source: https://github.com/ai-evals-course/judgy):

from judgy import estimate_success_rate

# Validate before conversion; unknown values must never become Fail.
judge_rates(test_human_labels, test_eval_labels)
prod_eval_labels = list(prod_eval_labels)
if not prod_eval_labels or any(x not in ("Pass", "Fail") for x in prod_eval_labels):
    raise ValueError("Need nonempty, complete production predictions")
binary = {"Pass": 1, "Fail": 0}
test_labels = [binary[x] for x in test_human_labels]
test_preds = [binary[x] for x in test_eval_labels]
unlabeled_preds = [binary[x] for x in prod_eval_labels]

theta_hat, lower, upper = estimate_success_rate(
    test_labels, test_preds, unlabeled_preds
)
print(f"Corrected rate: {theta_hat:.2f}")
print(f"95% conditional bootstrap interval (production fraction fixed): [{lower:.2f}, {upper:.2f}]")

Where Results Live

Write labels, splits, per-example predictions, and the calibration report into the user's project (e.g., evals/<failure-mode>/ — labels.jsonl, splits.json, calibration.md), never inside this skill's own directory. If files cannot be written in this environment, output the calibration report and label set as copyable blocks instead.

Practical Guidance

• Pin exact model versions for LLM judges (a dated snapshot id like <model>-<YYYY-MM-DD>, not a floating alias). Providers update models without notice, causing silent drift.
• Re-validate after any of these:
  • The judge prompt changed
  • The judge model was switched or un-pinned
  • The system under test changed (the agent's prompt or mandate was updated) — the failure mode's surface shifts even if the judge didn't move
  • Production confidence intervals widen unexpectedly
• Use ~100 labeled examples (50 Pass, 50 Fail). Below 60, confidence intervals become wide.
• One trusted domain expert is the most efficient labeling path. If not feasible, have two annotators label 20-50 traces independently and resolve disagreements before proceeding.
• Both error rates affect uncertainty. The denominator contains both TPR and TNR. Their relative contribution also depends on prevalence and class sample sizes; neither universally dominates.

Anti-Patterns

• Acting as the judge yourself in-context. The single most tempting shortcut, and it voids the entire exercise — see Step 2's execution constraint.
• LLM-generated "human" labels. Circular validation: you'd be measuring model-model agreement and calling it accuracy.
• Assuming judges "just work" without validation. A judge may consistently miss failures or flag passing traces.
• Using raw accuracy or percent agreement. Use TPR and TNR. With class imbalance, raw accuracy is misleading.
• Dev/test examples as few-shot examples. This is data leakage.
• Reporting dev set performance as final accuracy. Dev numbers are optimistic. The test set gives the unbiased estimate.
• Raw judge scores without bias correction. If you report an aggregate pass rate, apply the Rogan-Gladen formula (Step 7).
• Point estimates without confidence intervals. A corrected rate of 85% could easily be 78-92% with small test sets. Report the range so stakeholders know how much to trust the number.