refusal calibration

**Refusal calibration** is the **tuning of refusal decision thresholds so models decline harmful requests reliably while allowing benign requests appropriately** - calibration controls the practical balance between safety and usability. **What Is Refusal calibration?** - **Definition**: Adjustment of refusal probability mapping and policy cutoffs across risk categories. - **Target Behavior**: Near-zero refusal on safe prompts and near-certain refusal on clearly harmful prompts. - **Calibration Inputs**: Labeled benign and harmful datasets, adversarial tests, and production telemetry. - **Category Sensitivity**: Different harm domains require different threshold strictness. **Why Refusal calibration Matters** - **Boundary Accuracy**: Poor calibration causes both leakage and over-refusal errors. - **Policy Alignment**: Ensures refusal behavior matches product risk appetite and legal obligations. - **User Satisfaction**: Better calibration improves helpfulness on allowed tasks. - **Safety Reliability**: Correctly tuned systems resist ambiguous and adversarial prompt forms. - **Operational Stability**: Reduces oscillation from reactive policy changes after incidents. **How It Is Used in Practice** - **Curve Analysis**: Evaluate refusal performance across threshold ranges by harm class. - **Segmented Tuning**: Calibrate per category, language, and context domain. - **Continuous Recalibration**: Update thresholds as attack patterns and usage mix evolve. Refusal calibration is **a core safety-performance optimization process** - precise threshold tuning is essential for dependable refusal behavior in real-world LLM deployments.