{
 "experiment": "PHASE1.4b_auroc_reconciliation",
 "band_L": 20,
 "N": 817,
 "EXPLORATORY": true,
 "cells_2x2": {
  "kinematic_MLP_ORIGINAL": 0.6683,
  "kinematic_logreg": 0.6956,
  "rawband_MLP": 0.7958,
  "rawband_logreg_CLEANROOM": 0.7826
 },
 "recorded_was": "kinematic_MLP (~0.731)",
 "cleanroom_was": "rawband_logreg (~0.787)",
 "feature_effect_raw_minus_kinematic": 0.1072,
 "classifier_effect_logreg_minus_mlp": 0.0071,
 "strongest_cell": "rawband_MLP",
 "strongest_auroc": 0.7958,
 "VERDICT": "2x2 on identical rows+folds: kin+MLP(original)=0.6683, kin+logreg=0.6956, raw+MLP=0.7958, raw+logreg(cleanroom)=0.7826. The 0.731-vs-0.787 gap is a METHOD choice (feature effect raw-minus-kinematic=+0.107, classifier effect logreg-minus-MLP=+0.007), NOT a data discrepancy -- same data, same splits. RECOMMENDED HEADLINE: the strongest+simplest detector 'rawband_MLP' = 0.7958 (raw commitment-band hidden state + logistic regression). Both numbers sit inside the cluster-bootstrap CI [0.702, 0.828]; adopt one headline, report the other as a method variant. Pending ratification."
}