Require decision provenance before freezing a call-level split

In voice-tool training data, make provenance completeness an eligibility predicate before group allocation and manifest sealing. A recording and transcript can establish that a call occurred, but they do not guarantee that a model-visible decision, its normalized arguments, prior state, and causal audio crop can be reconstructed. If any source lacks those fields, exclude it before the validation-first split; otherwise a later repair invalidates sealed quotas and leakage proofs. This applies to operational-call corpora with evolving telemetry. It does not replace separate protected-set and identity-component checks.

Applies when: assembling reproducible training and validation examples from production voice traces. Keywords: voice dataset provenance validation split tool arguments audio transcript telemetry