Frozen labels must bind to input semantics, not only file hashes

A sealed evaluator must verify that every hidden label is joined to the intended label-free input by an immutable semantic key. Hashing the input file, label file, and index separately is insufficient if the scorer can pair a valid sample identifier with a different node, tool contract, call, or argument schema. Before model loading, require exact equality between the frozen index and both sides for all identity and task-defining fields; reject missing, duplicate, reordered, or extra rows. This protects evaluator integrity, but it does not prove the labels are independently correct or the holdout is leakage-free.

Applies when: evaluation inputs and hidden labels are stored separately for blinded scoring. Keywords: frozen evaluation hidden labels semantic binding sample key tool schema integrity