The first regression reported three after mismatches for float64-input gamma gradients. Parameters were default float32, so MLX correctly returned float32 parameter gradients; the reference used float64 values with a float64 tolerance. The corrected harness quantizes the independent parameter-gradient reference to the returned parameter-gradient dtype. No production code was changed to resolve this harness issue.
A proposed boundary [-192,192] did not isolate unbiased-only overflow: the original training output was also zero. Source inspection and direct primitive measurements show sum([36864,36864]) overflows FP16 before division by two. Thus even the representable biased variance 36864 is lost. The revised boundary script records this mechanism; the original script is preserved.
The first artifact validation had a checker-only error: it compared sorted changed method names with an incorrectly ordered literal list. Changed to set equality for the expected two methods. The patch and numerical results were unchanged.
