| { |
| "schema_version": "v0.2", |
| "dataset_name": "eval-trap-stability-manifold-benchmark-v0.2", |
| "pretty_name": "Eval Trap Stability Manifold Benchmark v0.2", |
| "task_type": "binary_classification", |
| "benchmark_type": "multi_surface_stability_eval_trap", |
| "description": "Synthetic benchmark for exposing evaluation traps in which high-confidence predictions occur in mathematically unstable system states across multiple competing stability surfaces.", |
| "stability_manifold": { |
| "surfaces": { |
| "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)" |
| }, |
| "stability_rule": "label_stable = 1 if min(S1, S2, S3) >= 0 else 0" |
| }, |
| "constants": { |
| "k": 1.0, |
| "boundary_eps": 0.02, |
| "high_content_score_threshold": 0.9 |
| }, |
| "columns": { |
| "scenario_id": { |
| "type": "string", |
| "description": "Unique scenario identifier" |
| }, |
| "pressure": { |
| "type": "float", |
| "description": "External system load" |
| }, |
| "buffer": { |
| "type": "float", |
| "description": "Remaining stabilizing capacity" |
| }, |
| "lag": { |
| "type": "float", |
| "description": "Control response delay" |
| }, |
| "coupling": { |
| "type": "float", |
| "description": "Subsystem coupling strength" |
| }, |
| "content_score": { |
| "type": "float", |
| "description": "Primary model confidence or affinity score" |
| }, |
| "regime_hint": { |
| "type": "string", |
| "allowed_values": ["baseline", "coupling", "lag"], |
| "description": "Human-readable hint for dominant instability regime" |
| }, |
| "label_stable": { |
| "type": "integer", |
| "allowed_values": [0, 1], |
| "description": "Ground truth stability label where 1 = stable and 0 = unstable" |
| } |
| }, |
| "prediction_contract": { |
| "required_columns": ["scenario_id", "prediction"], |
| "prediction_allowed_values": [0, 1], |
| "row_alignment": "scenario_id", |
| "scenario_id_must_be_unique": true |
| }, |
| "splits": { |
| "train": "Training examples spanning all three instability regimes", |
| "in_domain_test": "Evaluation examples drawn from same distribution family as train", |
| "distribution_shift": "Evaluation examples with shifted parameter ranges", |
| "boundary_trap": "Near-manifold seam examples designed to trigger false rescue behaviour" |
| }, |
| "diagnostic_regions": { |
| "collapse_zone": "min(S1, S2, S3) < -boundary_eps", |
| "near_boundary": "|min(S1, S2, S3)| <= boundary_eps", |
| "safe_margin": "min(S1, S2, S3) > boundary_eps" |
| } |
| } |
