| { |
| "schema_version": "v0.2", |
| "benchmark_name": "Clarus Evaluation Trap Stability Manifold Benchmark", |
| "version": "v0.2", |
| "task_type": "binary_classification", |
| "benchmark_type": "multi_surface_stability_eval_trap", |
| "description": "Benchmark for testing whether models detect instability across competing stability surfaces rather than relying on content confidence alone.", |
| "prediction_target": { |
| "column": "label_stable", |
| "meaning": "1 = stable, 0 = unstable" |
| }, |
| "prediction_contract": { |
| "required_prediction_columns": ["scenario_id", "prediction"], |
| "prediction_allowed_values": [0, 1], |
| "alignment_key": "scenario_id" |
| }, |
| "stability_manifold": { |
| "surface_definitions": { |
| "baseline_surface": "S1 = buffer - (pressure * coupling) - (k * lag)", |
| "coupling_surface": "S2 = buffer - (pressure * coupling^2) - (k * lag)", |
| "lag_surface": "S3 = buffer - (pressure * coupling) - (k * lag^2)" |
| }, |
| "collapse_margin": "min(S1, S2, S3)", |
| "surface_role": "Analytical reference manifold used for scorer diagnostics and region classification" |
| }, |
| "thresholds": { |
| "k": 1.0, |
| "boundary_eps": 0.02, |
| "high_content_score_threshold": 0.9 |
| }, |
| "evaluation_slices": { |
| "in_domain_test": { |
| "purpose": "Baseline evaluation on familiar regime distribution" |
| }, |
| "distribution_shift": { |
| "purpose": "Tests generalization under shifted pressure, lag, and coupling ranges" |
| }, |
| "boundary_trap": { |
| "purpose": "Tests false rescue behaviour near the stability seam" |
| } |
| }, |
| "evaluation_metrics": { |
| "primary_metric": "false_rescue_rate", |
| "secondary_metric": "boundary_error_rate", |
| "additional_metrics": [ |
| "accuracy", |
| "surface_distance_mean" |
| ] |
| }, |
| "diagnostics": [ |
| "collapse_margin_distribution", |
| "confusion_matrix", |
| "active_surface_counts" |
| ], |
| "metric_linkage": { |
| "false_rescue_rate": "Implemented in scorer.py as false rescues among high-confidence predicted stable cases", |
| "boundary_error_rate": "Implemented in scorer.py for cases where |collapse_margin| <= boundary_eps", |
| "surface_distance_mean": "Implemented in scorer.py as mean absolute collapse margin", |
| "collapse_margin_distribution": "Implemented in scorer.py across collapse_zone, near_boundary, and safe_margin", |
| "active_surface_counts": "Implemented in scorer.py by tracking which surface is minimum for each row" |
| } |
| } |
