Jev-like text fine-tuning does not by itself enable audio-conditioned choices

A trainable Jev-like typed-choice model needs an explicit audio input and training path before it can learn from waveforms. Fine-tuning its text adapter on transcripts is useful, but does not make it audio-conditioned. For speech-driven routing, first compare ASR→text-choice against a direct-audio classifier on the same caller clips, prompt context, VAD decision boundary, and independently adjudicated labels. Split by call/speaker and measure both exact route and VAD-to-decision latency. Only add an audio encoder/projector to the typed-choice model if that paired experiment shows a useful advantage. Choice accuracy does not establish free-form tool-argument accuracy.

Applies when: evaluating direct speech-to-tool routing against an ASR-to-choice pipeline. Keywords: Kev Jev audio conditioning ASR tool routing pointer head VAD paired evaluation