Replace the calibration-embedded v2 packages with the original inference-only BU risk heads. The five outputs are direct sigmoid probabilities and no calibration transform is bundled or applied.
Keep the same fitted coefficients and model tensors while renaming the overlap field to state that 942 test exams come from patients shared with validation; the exams themselves are disjoint.
Embed validation-fitted calibration in BU risk heads
Replace the two inference-only BU risk heads with v2 releases. Each release embeds its matching five-horizon Platt calibration, fitted on known validation outcomes only; no test outcome was used for fitting.
Replace BU risk folds with new S1-to-S2 and S2-only heads
Replaces the five historical fold checkpoints and manifest with the two new BU pretraining-exam-holdout risk heads. The S1-to-S2 head is the default release. Both heads use the frozen Mammo-FM_ASU_Trained_CLIP.tar encoder and produce raw, uncalibrated sigmoid scores. Their evaluation split is exact-exam-disjoint but not patient-disjoint.
