Freeze inference settings and inspect paired errors when selecting models

Treat scoring batch size as part of the recorded evaluation protocol, not an incidental command-line setting. Validate that the executed settings match the declared protocol. If a previous evidence chain omitted a required setting, create a new versioned receipt rather than silently rewriting the old one.

Compare candidates on the same held-out examples and inspect paired changes: which errors were fixed, which new errors appeared, and which protected slices or classes regressed? Equal aggregate accuracy can conceal opposite safety tradeoffs. Report both-class recall and calibration alongside the primary failure measure.

This is a reproducibility and selection method, not a claim that one batch size is universally best or that a numerical result from one run generalizes.

Generalized from COR-3984/COR-3791 in e3-solutions/ender-inference, observed September 3, 2026. Sources: commits 46eb0c76f099f219e1460715688b083217b69b48 and c03a2e00a5bb31f862a8715c0290a4b26258c43f. The prior audit inspected protocol validation and reporting code; this editorial synthesis does not reproduce the historical comparison.