| { |
| "schema_version": "1.0.0", |
| "taxonomy_version": "1.0.0", |
| "default_policy": "balanced", |
| "calibration_status": "uncalibrated", |
| "calibration_note": "Scores are raw softmax maxima for model spans and a fixed confidence table for rule spans. They are NOT calibrated probabilities: a span scored 0.90 does not mean it is correct 90% of the time. Per-type calibration is planned, not shipped. Choose a policy by measured precision/recall in models/evaluation.json, not by reading a score as a probability.", |
| "policies": { |
| "strict": { |
| "score_threshold": 0.5, |
| "intent": "Favour recall. Use when a human reviews every result." |
| }, |
| "balanced": { |
| "score_threshold": 0.75, |
| "intent": "Default. The published per-type results in evaluation.json are measured at this threshold." |
| }, |
| "precision": { |
| "score_threshold": 0.9, |
| "intent": "Favour precision. Use when a false positive is more costly than a miss." |
| } |
| }, |
| "per_type_overrides": {}, |
| "rule_confidence": { |
| "note": "Deterministic recognizers emit a fixed confidence by evidence strength, not a learned score.", |
| "checksum_validated": 0.99, |
| "format_and_context": 0.97, |
| "format_only": 0.93, |
| "weak_format_with_context": 0.8 |
| } |
| } |
|
|
