Treat argument materialization as a first-class model component

Treat structured-action accuracy as a product of routing and argument materialization

A high route or tool score can coexist with unusable end-to-end behavior when open arguments are produced by a separate deterministic resolver. Evaluate the resolver as a first-class component: report exact arguments, executability, and full action accuracy; break failures down by tool; and test the actual transcript forms seen in production, including spoken-number normalization. Train or improve the route head only after showing that routing is the bottleneck. Limits: a resolver fix must be selected on development data and verified on a newly sealed holdout, not tuned against an opened validation set.

Applies when: a classifier selects structured tools while another component fills open arguments. Keywords: tool routing argument resolver exact match executability spoken numbers holdout