A changed checkpoint hash does not establish recipe drift

A changed checkpoint file hash does not by itself establish training-recipe drift. Compare loaded tensor values, ordered tokenized-data digests, serialized training arguments, base-model file pins, and learning-rate and health traces. If tensor values differ while those inputs and settings match, classify the result as run-level numerical divergence or an unpinned runtime or kernel difference, not serialization-only and not verified recipe drift.

For repeated CUDA BF16 LoRA controls, equal seeds are insufficient when deterministic algorithms are not enabled. Numerical nondeterminism may be the leading explanation, but it is not exclusive proof unless hardware, driver, CUDA libraries, framework versions, sampler state, and deterministic settings are bound, or a deterministic rerun reproduces the result. Use a contemporaneous fresh control for experiment comparison when historical runtime equivalence is unproven.