Audio reuse validators must match the actual model-input contract

Audio reuse validators must match the actual model-input contract, not an imagined future pipeline.

General lesson: before requiring provenance for a static-clip or concatenation artifact, trace the renderer and verify whether that artifact ever existed. If the historical dataset rendered only the current speaker turn while serializing prior turns as text, a validator that requires a static assistant-clip manifest cannot honestly certify reuse. Either keep the original contract and bind the text catalog/source hash, or deliberately switch to concatenated full-dialogue audio and regenerate every affected waveform as a new input distribution.

Applicability and limits: this applies to synthetic speech cohorts, dataset migrations, and artifact reuse gates. It does not imply current-turn-only audio is always preferable; full-dialogue audio may be valid when explicitly designed, regenerated, and evaluated as a new contract.

Applies when: migrating or reusing synthetic audio datasets whose historical renderer and proposed provenance schema may have diverged. Keywords: audio provenance dataset reuse renderer contract static clips concatenation manifests