PART 2 METHOD AND RESULTS: every accelerator flops citation checked against its cited document, 72 sources probed, 25 flagged and 20 of those cleared as dense/sparse artefacts

METHOD, so the next pass can re-run it instead of re-deriving it. Every one of the 395 flops rows across the 108 accelerator records was mapped to its source_id, producing 72 distinct sources to check. Each source was fetched, its text normalised (thousands separators stripped, so '1,024' matches 1024; smart quotes folded; whitespace collapsed), and each attributed tflops value searched for under every plausible printed spelling: integer, one/two/three decimal places, trailing-zero-trimmed, %g, and the /1000 PFLOPS form for values >= 1000. THE FIRST PASS OVERREPORTED, AND THE CORRECTION IS THE POINT. The initial run flagged 25 of 72 sources as containing at least one value they do not print. Re-adjudication showed most were NOT failures. Three distinct causes, in decreasing order of frequency: 1. DENSE vs SPARSE, by far the largest group. The repo correctly records the DENSE member while the vendor prints the SPARSE one under a footnote. Confirmed cases where the stored value is exactly half a printed figure: nv-h100-datasheet (378, 494.5, 756.5, 989.5, 1670.5), nv-h200-product-page (417.5, 494.5, 835.5, 989.5), nv-l4-product-page (121, 242.5), nv-l40-datasheet (728.5 on the RTX 6000 Ada row), nv-a30-datasheet (82.5, 330.5), nv-blackwell-datasheet (1100), nv-gb200-nvl72-product-page (2500), consumer-nvidia-rtx-6000-ada-datasheet (728.5), cn-huawei-atlas-650e-spec (803.75). Every one of these is CORRECT as stored. Note the corollary: a naive citation check reports a correct dense value as missing, so 'the number is absent from the cited page' is NOT by itself a finding - it must be tested against the 2x and /2 spellings first. 2. THOUSANDS SEPARATORS. asic-rebellions-rebel100 prints '1,024 TFLOPS (FP16)'. A search for '1024' returns zero. This is the single easiest way to manufacture a false citation failure and it nearly produced one on a record that is entirely correct. All probes now strip separators before matching. 3. LAYOUT LOSS. nv-h100-datasheet's spec table collapses to a run of concatenated cells under plain extraction ('...FP3267teraFLOPS51teraFLOPS134teraFLOPSTF32TensorCore989teraFLOPS2...'), which destroys the column and row structure. Reading the PDF with pdftotext -layout recovers it exactly. Any future check of a PDF-based source should use -layout. THE FOUR GENUINE FAILURES, each confirmed with more than one method: - asic-rebellions-rebel100: NOT a failure (see cause 2). - xpu-onednn-github-readme backing xpu-intel-dc-gpu-flex-170 (fp32 20, fp16 80, int8 160) and xpu-intel-dc-gpu-max-1550 (fp32 52, fp16 409, bf16 409, int8 819): the README contains NONE of these. Raw fetch, 22,267 characters, zero occurrences of 409/819/160/80/20/52 as figures. See w5a-part2-onednn-readme-contains-no-tflops-figure. Citations withdrawn. - nv-l40-datasheet backing nvidia-l40s (fp32 91.6, tf32 183, bf16 362, fp16 362, fp8 733, int8 733): the L40 (Ada) datasheet prints 90.5 / 90.5 / 181.05 / 362 / 362 / 724 at 300W. A grep for '91.6' and '733' over its layout-preserved text returns zero. The correct document is NVIDIA's L40S product page. See w5a-part2-nvidia-l40s-product-page-holds-the-l40s-figures and w5a-part2-l40-ada-datasheet-vs-l40s-two-different-parts. Citations corrected, values unchanged. - amd-radeon-pro-vii-product-page: the page DOES carry 13.1 TFLOPS; the failure was in a NEGATIVE finding about it, not in the record. See w5a-part2-amd-radeon-pro-vii-page-does-contain-the-figure. nv-h100-datasheet backing nvidia-h100-nvl is a fifth case, distinct from all of the above: the values are correct but the cited column is a 2-card pair. See w5a-part2-h100-datasheet-nvl-column-is-a-2x-pair. TWO SOURCES REMAINED UNREADABLE rather than wrong, and are recorded as such rather than as failures: vb-cerebras-cs3-datasheet-125pf-sparse (the cdn.sanity.io PDF failed to scrape on this pass) and cn-huawei-atlas-650e-spec, whose four values are already marked basis 'derived' rather than transcribed, so the arithmetic - cabinet total divided by 8 NPUs - is the thing to check, not the page. THE HEADLINE NUMBER FOR ANY FUTURE PASS: 72 sources probed, 395 rows, 4 genuine citation failures, 2 wrong-or-misleading citations now corrected, 1 wrong negative finding deposed. The remaining risk is concentrated in sources that could not be fetched at all, which is the argument for treating an unreadable source as unverified rather than as passing.