The temperature=0 to greedy decoding to "whole request is deterministic" leap is the one I've made too, and the batch-shape-changes-the-reduction-order angle is the cleanest explanation of it I've read. The lesson I took from a similar week was to stop treating a single byte-diff as pass/fail on generative outputs and score against a tolerance (semantic equivalence, or a bounded token-set) so nondeterministic-but-correct answers stop lighting up the harness. Curious whether you kept the exact-diff on top of that or replaced it entirely.