Temporal counterfactual data diagnostics
train
- n_samples: 512
- active classes: 5 / 12
- dead classes (schema present, no data): 7
- stable fraction: 0.5000
- label_counts: {'stable': 256, 'pre_slip': 0, 'pre_drop': 0, 'collision_precursor': 0, 'fall_precursor': 0, 'navigation_dead_end': 0, 'wrong_click': 0, 'destructive_action': 73, 'workflow_divergence': 75, 'form_submission_failure': 61, 'stale_context_action': 47, 'invalid_tool_selection': 0}
- pair_count: 256
- pairs with byte-equal state: 256 / 256
- pairs with byte-equal context: 256 / 256
- pairs with byte-equal action_tokens: 256 / 256
- pairs with last-frame equal (1e-6): 256 / 256
- pairs with identical unordered frame bag (1e-6): 256 / 256
- per-step mean frame L1: [0.0, 0.02095, 0.02762, 0.02551, 0.02578, 0.02479, 0.02353, 0.0]
- per-step signal active steps mask: [0, 1, 1, 1, 1, 1, 1, 0]
- early(1-3)/late(4+) signal ratio: 0.9998
- per-family step signal:
- delayed_consequence (n=61): [0.0, 0.02647, 0.02175, 0.02137, 0.01972, 0.01939, 0.01901, 0.0]
- persistence_then_reversal (n=47): [0.0, 0.01436, 0.02765, 0.03327, 0.0365, 0.02231, 0.02873, 0.0]
- order_reversal (n=73): [0.0, 0.02241, 0.02954, 0.02045, 0.02492, 0.02497, 0.0277, 0.0]
- regression_after_progress (n=75): [0.0, 0.01919, 0.03052, 0.02894, 0.02481, 0.03059, 0.01988, 0.0]
- v2 event step pair counts: {'1,2': 10, '1,3': 15, '4,5': 16, '3,5': 19, '2,5': 21, '2,6': 18, '2,3': 23, '3,4': 15, '4,6': 19, '2,4': 22, '1,5': 17, '1,4': 18, '1,6': 11, '5,6': 14, '3,6': 18}
test
- n_samples: 256
- active classes: 5 / 12
- dead classes (schema present, no data): 7
- stable fraction: 0.5000
- label_counts: {'stable': 128, 'pre_slip': 0, 'pre_drop': 0, 'collision_precursor': 0, 'fall_precursor': 0, 'navigation_dead_end': 0, 'wrong_click': 0, 'destructive_action': 38, 'workflow_divergence': 36, 'form_submission_failure': 24, 'stale_context_action': 30, 'invalid_tool_selection': 0}
- pair_count: 128
- pairs with byte-equal state: 128 / 128
- pairs with byte-equal context: 128 / 128
- pairs with byte-equal action_tokens: 128 / 128
- pairs with last-frame equal (1e-6): 128 / 128
- pairs with identical unordered frame bag (1e-6): 128 / 128
- per-step mean frame L1: [0.0, 0.02847, 0.02344, 0.01944, 0.02572, 0.02866, 0.02546, 0.0]
- per-step signal active steps mask: [0, 1, 1, 1, 1, 1, 1, 0]
- early(1-3)/late(4+) signal ratio: 0.8936
- per-family step signal:
- delayed_consequence (n=24): [0.0, 0.02919, 0.01479, 0.02267, 0.03251, 0.01807, 0.02111, 0.0]
- order_reversal (n=38): [0.0, 0.0296, 0.03048, 0.0185, 0.02235, 0.03503, 0.02551, 0.0]
- persistence_then_reversal (n=30): [0.0, 0.03328, 0.02052, 0.02295, 0.02945, 0.02863, 0.0155, 0.0]
- regression_after_progress (n=36): [0.0, 0.02278, 0.02421, 0.01534, 0.02165, 0.02901, 0.0366, 0.0]
- v2 event step pair counts: {'1,4': 9, '1,2': 7, '2,5': 9, '2,4': 9, '1,3': 10, '1,5': 12, '5,6': 12, '1,6': 9, '3,5': 7, '4,6': 9, '4,5': 10, '3,4': 7, '2,3': 5, '2,6': 9, '3,6': 4}