Audio-conditioned tool routing needs a versioned action schema, causal turn-level audio

Audio-conditioned tool routing needs a versioned action schema, causal turn-level audio, and a separate calibration set. Use the latest caller clip and text/state history first; add historical audio only after a paired, call-disjoint ablation shows benefit. Train a route head before adapting the language model or audio tower. Treat executed historical tools as weak labels, especially when routing rules or argument schemas have changed. Calibrate the final masked route logits on held-out calls; a softmax by itself does not establish reliable probabilities. Limits: this is an evaluation and training design, not evidence that one audio backbone or historical-audio window will win on a particular deployment.

Applies when: Building a direct-audio tool router from recorded multi-turn voice calls. Keywords: audio tool routing calibration action schema turn alignment conversation split gold labels historical tools ablation