Acceptance for spoken identifiers must couple model routing to stateful handler effects

For spoken-identifier agents, production acceptance should pair exact model routing with caller-visible handler and state assertions; a higher tool-call exact-match score alone cannot prove safe behavior. Use two hard-gated families: (1) unambiguous full restarts, where every pre-restart digit is discarded and the complete replacement is verified exactly, including leading zero/one and length boundaries; and (2) genuine partials, where no lookup or identity mutation occurs, exactly one backend clarification is presented, retry state changes once, and a later complete identifier is never concatenated with prior fragments. Include the configured retry-exhaustion boundary as its own scenario so terminal transition and duplicate-speech bugs cannot hide behind aggregate accuracy. Limits: authored acceptance cases validate the specified contract and integration path, not real-call prevalence, model generalization, or causal improvement in completion.

Evidence/example — 2026-09-09

A read-only synthetic runtime inspection at pinned source revision 5925a60 found that explicit repeat/clarify handling increments the same carrier-number retry counter as other recovery, stays in collection on attempts 1–2, and returns a terminal node on attempt 3. A successful later complete verification resets the counter and advances to carrier confirmation. This dated local evidence motivates the state assertions; it is not an attestation of current deployment behavior.