Separate recognition errors from transcript delivery before choosing a voice fine-tune

When diagnosing a voice agent, distinguish wrongly recognized words, correctly recognized speech lost before the language model, and correct text evaluated with stale context or tools. A corrected-transcript replay only isolates recognition if the original text actually delivered, pre-decision history, tools, state, model and settings are held fixed. Stop at the next decision; whole-conversation replays let histories diverge. Treat missing exact request snapshots as reconstructed tests, not exact historical counterfactuals.

Use a representative call sample for prevalence and a separate enriched failure set for diagnosis. An excerpt pool selected for exact ASR agreement and longer utterances cannot estimate recognition failures or short-reply performance. Report paired rescues, unchanged failures and harms, with unknown/delivery failures separate. These are diagnostic safeguards, not proof that a particular recognizer is better.

Evidence/example — 2026-09-23: constructed example, with no customer data: a recognizer correctly emits a one-word answer, but a playback filter drops it. Injecting that answer into a text replay may repair the downstream decision, yet this demonstrates restored input delivery, not improved acoustic recognition. Similarly, a correctly received answer can fail if the question's tools have not become active. This extends the linked audit guidance from symptom/coverage separation to controlled next-decision comparisons. No new model evaluation or prevalence result is claimed.