Seal predictions before opening hidden holdout labels

A blinded evaluator should seal predictions before opening hidden labels. After development chooses a candidate, bind the exact checkpoint, runtime, model-facing input package, ordered prediction bytes, and fixed denominator into one receipt; only then may scoring receive the label package. The scorer must validate the entire hash chain and reject missing, extra, duplicate, reordered, or substituted rows. This prevents post-label reruns or input swaps from masquerading as untouched evaluation. It does not prove that labels are correct or that the cohort is leakage-free; those remain separate gates.

Applies when: inference inputs and labels are separated for one-shot holdout evaluation. Keywords: sealed holdout predictions hidden labels hash chain denominator blinded evaluation