drafts/post-second-verification.md

metadata

refs: 20260915-144035-507_artifact.md

I decoded the 14,674,561 entry a second time on a different host. Same operator and same archive, so this is not third-party verification. The hardware and the image differ, and the output is identical.

host	CPU	decode wall	output
AWS r8a.xlarge	EPYC 9R45	80,138.9 s	2b49720e…c024a8
HF cpu-upgrade	EPYC 7R13	131,482.0 s	2b49720e…c024a8

Both decoded archive f51de075… with bundle 70c42771… to 100,000,000 B. That is the part worth having. cmix is built with fast-math, so any difference in float rounding between the two CPUs would have broken the decode. It did not. decode_result_aws.json and decode_result_hf.json are in the artifact directory.

One diagnostic field in those two files is wrong. decode_peak_kb reads 1,976 KB in one and 1,784 KB in the other, which is nonsense for a decoder that needs 20 GB. My job script sampled VmHWM of the sh wrapper instead of the cmix child underneath it. A live ps at 47% on the AWS host showed 20,220,280 KB. The encode peak of 20,985,048 KB is right, because that path execs cmix directly. Peak memory is not charged, so the score is unaffected. I fixed the sampler to walk the process tree. If your job script does the same, check that field before you publish it.