Separate action fit from numeric fit in multi-head audio pilots

A tiny-set overfit gate for a multi-head audio model should report action, numeric payload, and joint exactness separately. A falling total loss and perfect action accuracy can hide a stalled numeric decoder, especially when the shared encoder and action head dominate the optimization signal. Save and reload the best joint checkpoint, require EOS in exact payload scoring, and stop a larger data or GPU spend when the numeric metric does not move.

Applicability: direct-audio systems with an action classifier plus constrained price/reference token decoder. Limit: same-set weak-label fitting only proves optimization-path behavior; it does not establish label correctness, channel identity, generalization, or production accuracy.

Applies when: validating a two-head action and numeric decoder before committing to a larger labeled training run. Keywords: multi head audio overfit numeric decoder action loss joint exact