AUDIT NEGATIVE: the oneDNN README cited by xpu-intel-dc-gpu-flex-170 and xpu-intel-dc-gpu-max-1550 contains NO TFLOPS figure at all - a citation that does not support its number

PART 2 AUDIT FINDING, 2026-10-05. THE WORST-CASE CITATION FAILURE IN THIS REPO SO FAR, because the citation LOOKS like a first-party Intel document and is not. WHAT THE RECORD CLAIMS. data/accelerators/xpu-intel-dc-gpu-flex-170 carries fp32 20, fp16 80, int8 160 with source_id 'xpu-onednn-github-readme', basis 'vendor-docs', vendor_claim TRUE, basis_detail 'oneDNN (oneAPI) GitHub README.' data/accelerators/xpu-intel-dc-gpu-max-1550 carries fp32 52, fp16 409, bf16 409, int8 819, also sourced to the README, basis already demoted to 'secondary-aggregator' with the note 'Not an Intel figure. ... basis was vendor-docs until 2026-10-05, which overstated the...' (the sibling agent caught the basis but not the citation). WHAT THE DOCUMENT ACTUALLY CONTAINS. I fetched the RAW README (raw.githubusercontent.com/oneapi-src/oneDNN/master/README.md, 22,267 characters extracted) rather than the rendered GitHub page, because the rendered page was what produced the false positive in the first place. A substring sweep over the full raw text returns: '409' -> 0 occurrences, '819' -> 0, '160' -> 0, '80' -> 0, '52' -> 0. The strings TFLOPS, teraFLOPS and TOPS do not occur as a specification of any product. WHAT IT DOES CONTAIN, verbatim: the GPU section is a bulleted list of SUPPORTED ARCHITECTURES with no numbers - 'The library is optimized for the following GPUs: * Intel discrete GPUs: * Intel Arc(TM) A-Series Graphics (formerly Alchemist) * Intel Data Center GPU Flex Series (formerly Arctic Sound) * Intel Data Center GPU Max Series (formerly Ponte Vecchio) * Intel Arc B-Series Graphics and Intel Arc Pro B-Series Graphics (formerly Battlemage) * future discrete GPUs based on Xe3p-XPC architecture (code name Crescent Island)'. The occurrences of '52' that a naive search finds are inside a Base64-encoded camo.githubusercontent.com image URL in the OpenSSF Scorecard badge ('...projects/8762/badge...'), i.e. a badge asset id, not a spec. WHY THIS MATTERS MORE THAN A NORMAL MISSING FIGURE. A README that names 'Intel Data Center GPU Flex Series' and 'Intel Data Center GPU Max Series' is the single most plausible-looking citation for an Intel GPU's compute figure: same vendor ecosystem, same product family names, machine-readable, present in the record's sources list. It looks sourced in a way that no aggregator citation does. That is exactly the failure mode AGENTS.md warns about - 'a citation that does not support the number is the worst failure mode in this repo, because it looks sourced' - and here it also carried vendor_claim: true. WHAT WOULD SETTLE IT: Intel ARK (ark.intel.com) for the Flex 170 and Max 1550. Prior passes already established that ARK publishes NO compute row for these parts at all (see w4h-intel-dc-gpu-flex-170-ark-spec-table and w4h-intel-dc-gpu-max-1550-ark-spec-table), and the Flex Series product brief's '16 TFLOPS (FP32) / 250 TOPS (INT8)' is a third, different pair. So there is currently NO Intel-published compute figure for either part, and the honest state for both records is contested with the unsupported citation named.