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
- (Optional) Label traces with the user if labels don't exist yet
- Split human-labeled data into train (10-20%), dev (40-45%), test (40-45%)
- Run judge on dev set and measure TPR/TNR
- Iterate on the judge until TPR and TNR > 90% on dev set
- Run once on held-out test set for final TPR/TNR
- 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
numpyandscikit-learn(sklearn). Optionallyjudgy(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:
- Load the user's traces file (JSONL/CSV, one trace per row).
- 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. - Record the label, plus any note (notes often become Pass/Fail definition refinements).
- 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. - 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:
| Split | Size | Purpose | Rules |
|---|---|---|---|
| Training | 10-20% (~10-20 examples) | Source of few-shot examples for the judge prompt | Only clear-cut Pass and Fail cases. Used directly in the prompt. |
| Dev | 40-45% (~40-45 examples) | Iterative evaluator refinement | Never include in the prompt. Evaluate against repeatedly. |
| Test | 40-45% (~40-45 examples) | Final unbiased accuracy measurement | Do NOT look at during development. Used once at the end. |
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)
from sklearn.metrics import confusion_matrix
tn, fp, fn, tp = confusion_matrix(human_labels, evaluator_labels,
labels=['Fail', 'Pass']).ravel()
tpr = tp / (tp + fn)
tnr = tn / (tn + fp)
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:
| Disagreement Type | Judge | Human | Fix |
|---|---|---|---|
| False Pass | Pass | Fail | Judge is too lenient. Strengthen Fail definitions or add edge-case examples. |
| False Fail | Fail | Pass | Judge is too strict. Clarify Pass definitions or adjust examples. |
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.
Stopping criteria:
- Target: TPR > 90% AND TNR > 90%
- Minimum acceptable: TPR > 80% AND TNR > 80%
If alignment stalls:
| Problem | Solution |
|---|---|
| TPR and TNR both low | Use a more capable LLM for the judge |
| One metric low, one acceptable | Inspect disagreements for the low metric specifically |
| Both plateau below target | Decompose the criterion into smaller, more atomic checks |
| Consistently wrong on certain input types | Add targeted few-shot examples from training set |
| Labels themselves seem inconsistent | Re-examine human labels; the rubric may need refinement |
Step 6: Final Measurement on Test Set
Run the judge exactly once on the held-out test set. Record final TPR and TNR.
Do not iterate after seeing test set results. Go back to step 4 with new dev data if needed.
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 PassTPR,TNR= from test set measurementtheta_hat= corrected estimate of true success rate
Clip to [0, 1]. Invalid when TPR + TNR - 1 is near 0 (judge is no better than random).
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
- True success rate is ~85%, not the raw 80%
Step 8: Confidence Interval
Compute a bootstrap confidence interval. A point estimate alone is not enough.
import numpy as np
def bootstrap_ci(human_labels, eval_labels, p_obs, n_bootstrap=2000):
"""Bootstrap 95% CI for corrected success rate."""
n = len(human_labels)
estimates = []
for _ in range(n_bootstrap):
idx = np.random.choice(n, size=n, replace=True)
h = np.array(human_labels)[idx]
e = np.array(eval_labels)[idx]
tp = ((h == 'Pass') & (e == 'Pass')).sum()
fn = ((h == 'Pass') & (e == 'Fail')).sum()
tn = ((h == 'Fail') & (e == 'Fail')).sum()
fp = ((h == 'Fail') & (e == 'Pass')).sum()
tpr_b = tp / (tp + fn) if (tp + fn) > 0 else 0
tnr_b = tn / (tn + fp) if (tn + fp) > 0 else 0
denom = tpr_b + tnr_b - 1
if abs(denom) < 1e-6:
continue
theta = (p_obs + tnr_b - 1) / denom
estimates.append(np.clip(theta, 0, 1))
return np.percentile(estimates, 2.5), np.percentile(estimates, 97.5)
lower, upper = bootstrap_ci(test_human, test_eval, p_obs=0.80)
print(f"95% CI: [{lower:.2f}, {upper:.2f}]")
Or use judgy (pip install judgy, source: https://github.com/ai-evals-course/judgy):
from judgy import estimate_success_rate
# judgy expects 0/1 integer labels (1 = Pass, 0 = Fail)
test_labels = [1 if l == 'Pass' else 0 for l in test_human_labels]
test_preds = [1 if l == 'Pass' else 0 for l in test_eval_labels]
unlabeled_preds = [1 if l == 'Pass' else 0 for l 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% CI: [{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.
- Improving TPR narrows the confidence interval more than improving TNR. The correction divides by
(TPR + TNR - 1), so a low TPR shrinks the denominator and amplifies estimation errors into wide CIs.
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.