Confirmed the large-mean InstanceNorm cancellation defect on an NVIDIA H100. The shifted two-pass repair passes all 18 admission checks; at mean offset 1000 its maximum absolute error is 8.31e-7 versus 0.410156 for the original kernel.
Regenerated 86 raw mutations into 85 retained records: 81 witnessed and 60 killed by official screening. Include 432 targeted matrix cells, raw journals, source fingerprints, and runtime details. The 66 original records remain archived.
Diagnosed a float32 oracle validity failure in an additional near-constant input and changed that added input noise from 0.01 to 0.1. Original suites, oracle, and tolerances remain unchanged. The measured kernel-time ratio is 1.509 on shape [8,64,512,512]. Keep the homepage erratum paragraph removed as requested.
Remove InstanceNorm erratum notice from dataset card
Use a shifted mean and two-pass population variance; add large positive/negative mean and near-constant regression suites. Include independent CPU numerical reproduction and successful NVRTC compilation. Archive the 66 baseline witness records and exclude InstanceNorm from active summary counts until actual CUDA re-admission and mutation regeneration complete. Add reproducible GPU validation and regeneration commands and document the historical scope of aggregate records.
