The numerical policy is shared by both roles and survives checkpoints.
Parameters, sample-weighted combination and update use binary64. Each local
sample rounds W to binary32, evaluates the checked SM86 gradient, then
widens that value exactly. This is a declared mixed-precision learner,
not a claim that arbitrary inputs equal the pure binary64 reference.
The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.