Tool-contract benchmark gains can represent caller-facing behavior changes

A tool-contract benchmark gain can be a real caller-facing functional gain when a tool-only runtime suppresses model-authored plaintext and substitutes deterministic recovery speech. This applies to interactive voice nodes whose frame pipeline enforces tool-only output. To classify the gain, trace four things in order: whether plaintext reaches TTS, whether missing-tool recovery retries the model or synthesizes a backend action, which counters change, and whether the node transitions. Limit: a retained-source trace and a queued TTS frame do not prove that this code is deployed or that audio was audible.

Evidence/example — 2026-09-10, retained application source snapshot and reference-contract audit: on a reference-collection node, a plaintext response such as a natural-language wait acknowledgment is dropped before TTS. At response end, the runtime synthesizes a repeat-or-clarify control action, queues a generic reference/lane clarification, and remains on the same node; it does not make a second model request or increment load/carrier business-failure counters, though generic interaction attempt bookkeeping advances. A correctly emitted wait control action instead selects the wait acknowledgment path. Provenance: local retained source snapshot plus its contract audit; deployment and playback were explicitly out of scope.