Metrics

Guard Eval Harness computes binary classification metrics for each dataset in a run. All metrics are computed from the confusion matrix of ground-truth labels vs. model predictions (thresholded at the configured threshold).

Available Metrics

geh list metrics

Binary Classification

Metric	Formula	Description
Accuracy	(TP + TN) / Total	Overall correctness
Precision	TP / (TP + FP)	Of predicted unsafe, how many are truly unsafe
Recall	TP / (TP + FN)	Of truly unsafe, how many were caught
F1	2 * (P * R) / (P + R)	Harmonic mean of precision and recall
AUROC	Area under ROC	Ranking quality across all thresholds
AUPRC	Area under PR curve	Precision-recall tradeoff quality
FPR	FP / (FP + TN)	False positive rate (over-blocking)
FNR	FN / (FN + TP)	False negative rate (missed unsafe content)

Confusion Matrix

Metric	Description
TP	True positives — correctly identified unsafe
TN	True negatives — correctly identified safe
FP	False positives — safe content flagged as unsafe
FN	False negatives — unsafe content missed

How Metrics Are Computed

1. Each sample gets a unsafe_score in [0.0, 1.0] from the model adapter
2. The score is thresholded: unsafe_label = unsafe_score >= threshold
3. Binary metrics are computed from the confusion matrix of label vs unsafe_label
4. AUROC and AUPRC use the raw unsafe_score (threshold-independent)

Output Format

Metrics are written to metrics.json per dataset:

{
  "accuracy": 0.92,
  "precision": 0.89,
  "recall": 0.95,
  "f1": 0.92,
  "auroc": 0.97,
  "auprc": 0.96,
  "fpr": 0.11,
  "fnr": 0.05,
  "tp": 190,
  "tn": 230,
  "fp": 28,
  "fn": 10
}

Aggregated metrics across all datasets appear in summary.json.

Interpreting Results

For Safety Guardrails

• High recall is critical — you don't want to miss unsafe content (low FNR)
• Acceptable FPR depends on your use case — some over-blocking may be tolerable
• F1 balances precision and recall — good for overall comparison
• AUROC measures ranking quality independent of threshold — useful for model comparison

Threshold Tuning

The threshold parameter directly affects precision/recall tradeoff:

Threshold	Effect
Lower (e.g., 0.3)	Higher recall, more false positives (conservative)
Default (0.5)	Balanced
Higher (e.g., 0.7)	Higher precision, more false negatives (permissive)

Tip

Use AUROC to compare models independently of threshold, then tune the threshold for your deployment requirements.

Partial Predictions

If a model adapter fails on some samples (with drop_failed_predictions: true), the harness validates the prediction set and flags runs with high drop rates:

Drop Rate	Status
< 5%	Normal — metrics computed on available predictions
5–20%	Warning logged
> 20%	Run flagged as "partial" in manifest