dataset_schema_.json

{
  "version": "0.3",
  "benchmark_name": "CASSES",
  "description": "Clarus Adversarial State-Space Evaluation Standard v0.3",

  "splits": [
    "train",
    "in_domain_test",
    "boundary_trap",
    "distribution_shift",
    "counterfactual_intervention"
  ],

  "trap_types": [
    "false_stability",
    "boundary_masking",
    "trajectory_aliasing",
    "temporal_alias",
    "intervention_decoy"
  ],

  "difficulty_levels": [
    "easy",
    "medium",
    "hard"
  ],

  "pair_roles": [
    "safe_pair",
    "unstable_pair"
  ],

  "thresholds": {
    "low_boundary_distance": 0.10,
    "high_drift_gradient": 0.05
  },

  "required_prediction_fields": [
    "scenario_id",
    "prediction"
  ],

  "optional_prediction_fields": [
    "intervention_effect_direction"
  ],

  "required_truth_fields": [
    "scenario_id",
    "split_type",
    "pair_id",
    "pair_role",
    "difficulty_level",
    "pressure_obs_t0",
    "pressure_obs_t1",
    "pressure_obs_t2",
    "buffer_obs_t0",
    "buffer_obs_t1",
    "buffer_obs_t2",
    "true_label",
    "trap_type",
    "trap_active",
    "boundary_distance",
    "drift_gradient",
    "drift_acceleration",
    "recovery_feasibility",
    "regime_competition_ratio",
    "intervention_action",
    "intervention_magnitude",
    "boundary_distance_before",
    "boundary_distance_after",
    "intervention_effect_direction"
  ],

  "metrics": {
    "basic": [
      "accuracy",
      "precision",
      "recall",
      "f1",
      "rows_evaluated"
    ],
    "split_metrics": [
      "train_accuracy",
      "in_domain_test_accuracy",
      "boundary_trap_accuracy",
      "distribution_shift_accuracy",
      "counterfactual_intervention_accuracy"
    ],
    "trap_metrics": [
      "trap_accuracy",
      "false_stability_accuracy",
      "boundary_masking_accuracy",
      "trajectory_aliasing_accuracy",
      "temporal_alias_accuracy",
      "intervention_decoy_accuracy"
    ],
    "difficulty_metrics": [
      "easy_accuracy",
      "medium_accuracy",
      "hard_accuracy"
    ],
    "geometry_metrics": [
      "low_boundary_distance_miss_rate",
      "high_drift_gradient_miss_rate"
    ],
    "counterfactual_metrics": [
      "counterfactual_intervention_accuracy",
      "intervention_effect_direction_accuracy"
    ],
    "pair_metrics": [
      "trajectory_pair_discrimination_accuracy"
    ],
    "support_metrics": [
      "train_support",
      "in_domain_test_support",
      "boundary_trap_support",
      "distribution_shift_support",
      "counterfactual_intervention_support",
      "trap_support"
    ],
    "composite_metrics": [
      "manifold_generalization_gap",
      "casses_score"
    ]
  },

  "composite_score": {
    "name": "casses_score",
    "weights": {
      "in_domain_test_accuracy": 0.18,
      "boundary_trap_accuracy": 0.20,
      "distribution_shift_accuracy": 0.17,
      "trap_accuracy": 0.15,
      "counterfactual_intervention_accuracy": 0.15,
      "trajectory_pair_discrimination_accuracy": 0.15
    },
    "renormalize_if_missing": true
  },

  "manifold_generalization_gap": {
    "formula": "in_domain_test_accuracy - distribution_shift_accuracy"
  },

  "pair_discrimination_rule": {
    "description": "For each pair_id, one row must have pair_role=safe_pair and one row must have pair_role=unstable_pair. A pair is correct only if the unstable row is predicted 1 and the safe row is predicted 0."
  },

  "counterfactual_rule": {
    "description": "Rows in counterfactual_intervention evaluate whether the model correctly predicts the outcome label and, when provided, the intervention effect direction."
  },

  "validation_rules": {
    "prediction_binary_values": [0, 1],
    "truth_binary_values": {
      "true_label": [0, 1],
      "trap_active": [0, 1]
    },
    "intervention_effect_direction_values": [-1, 0, 1],
    "require_exact_scenario_id_match": true,
    "require_unique_scenario_id": true,
    "allow_null_trap_type": true,
    "allow_null_pair_id": true,
    "allow_null_pair_role": true
  },

  "interpretation": {
    "accuracy": "Overall classification correctness",
    "trap_accuracy": "Performance on trap-active rows only",
    "boundary_trap_accuracy": "Performance on boundary deception rows",
    "distribution_shift_accuracy": "Performance when observable manifold shifts",
    "counterfactual_intervention_accuracy": "Performance on intervention outcome rows",
    "trajectory_pair_discrimination_accuracy": "Ability to distinguish paired safe vs unstable trajectories",
    "low_boundary_distance_miss_rate": "Fraction of unstable near-boundary rows predicted stable",
    "high_drift_gradient_miss_rate": "Fraction of unstable high-drift rows predicted stable",
    "casses_score": "Composite state-space reasoning benchmark score"
  },

  "output_format": {
    "type": "json",
    "required_output_keys": [
      "accuracy",
      "precision",
      "recall",
      "f1",
      "rows_evaluated",
      "train_accuracy",
      "in_domain_test_accuracy",
      "boundary_trap_accuracy",
      "distribution_shift_accuracy",
      "trap_accuracy",
      "false_stability_accuracy",
      "boundary_masking_accuracy",
      "trajectory_aliasing_accuracy",
      "temporal_alias_accuracy",
      "intervention_decoy_accuracy",
      "easy_accuracy",
      "medium_accuracy",
      "hard_accuracy",
      "manifold_generalization_gap",
      "low_boundary_distance_miss_rate",
      "high_drift_gradient_miss_rate",
      "counterfactual_intervention_accuracy",
      "intervention_effect_direction_accuracy",
      "trajectory_pair_discrimination_accuracy",
      "train_support",
      "in_domain_test_support",
      "boundary_trap_support",
      "distribution_shift_support",
      "counterfactual_intervention_support",
      "trap_support",
      "casses_score"
    ]
  }
}