Identical-input score correction before any verified publication

The independent execution4 replay passed all five model/metric fold checks and all score tolerances, then failed exact review selections. Identical molecular-only inputs (no gas measurements) in one fold acquired up to 1.78e-15 score jitter from batch matrix multiplication. For example, 3,868 zero-gas rows in fold 0 received nine distinct float scores despite identical vectors. Lexicographic tie semantics were therefore not faithfully implemented. The original execution2 output remains unverified, preserved and excluded from publication. This is a disclosed post-read numerical correction, not an unseen preregistration or an adjusted science gate.

Canonicalize identical transformed feature vectors within each prediction call, compute each unique vector once, then map its score to every duplicate row. The independent direct-distance implementation canonicalizes separately. Freeze the corrected worker and checker before a new analysis; add a duplicate-vector synthetic control crossing a 512-row batch boundary. No feature, outcome, model, regularization, bandwidth formula, capacity, threshold, budget or random seed changes. The original four models must still replay within 1e-7 and retain their original review selections.

The checker proves numerical agreement for every score at the original tolerance, and independently orders the exact stored float bytes for selections. Equivalent inputs within each fold must receive bit-identical stored scores. Floating-point agreement is not bitwise equality between two independent linear-algebra algorithms. All identities, curves, controls, numerical fits and primary metrics are rechecked.

Fresh stages: ablation_controls_execution3_20261014 (50s/1MiB), ablation_diagnostic_execution3_20261014 (1130s/12MiB), ablation_verification_execution5_20261014 (530s/4MiB). The original total 1740s and aggregate output budgets still apply, including earlier failed stages. Fit/score outputs from earlier execution2 stay local with complete hash inventories. Compact publication uses only the separately verified corrected packet. No model is selected or promoted after observing this correction.