inference/research
verifiedhigh · 0.90updated 2026-10-04

Method record: how source-record content was re-verified for silent revision on 2026-10-04 (spec sheets, MLPerf tables, pricing pages)

spec-sheet

Fields

id
sources/w3s-silent-revision-class-vendor-spec-and-pricing-pages
Publisher
NVIDIA
Kind
spec-sheet
Accessed
2026-10-04
notes
METHOD RECORD, so the audit is reproducible and so its NEGATIVES can be re-checked rather than inherited. url is a representative live NVIDIA LLM-readable spec endpoint used as the worked example; the sweep covered the whole of data/sources/, not this page. WHAT WAS RUN. (1) Every PDF source URL re-downloaded and text-extracted via the PDF text layer, not a rendered screenshot: 45 of 47 distinct PDFs parsed. (2) Every html source URL of kind spec-sheet/database/benchmark/release-notes re-fetched: 368 URLs, 309 returned 200. (3) For each record, the numeric figures in its notes field were searched for in the freshly-fetched text, comma-insensitively. A figure that no longer appears in its own source is a candidate silent revision. THE INSTRUMENT IS BLUNT AND ITS FAILURE MODES ARE THE POINT. It cannot tell a revision from a paraphrase, from a figure the record derived rather than transcribed, or from a figure that lives in a client-rendered table. Confirmed instrument failures, all kept as negatives rather than promoted to findings: pwr-mlperf-v4-1-h200-maxq-l70-server-summary flagged 9 missing figures because the record cites a github.com/BLOB/ html page whose numbers exist only in the raw file - fetching raw.githubusercontent.com instead cleared 6 of them; vb-cambricon-*, pd-cerebras-pricing-dom-rendered and every vast.ai / fireworks.ai / baseten.co pricing source flagged missing because those pages render their tables client-side and return under 4KB of shell to a static fetch, which is EXACTLY the repo's rent-cerebras-api-token-pricing-negative false negative and must not be repeated; sup-cloud-together-gpu-clusters-pricing and sup-cloud-vast-ai-pricing flagged missing for the same reason while the sibling sup-cloud-coreweave-pricing re-fetched with all 14 cited figures still present. DERIVED VALUES FLAG THEMSELVES, which is a property worth keeping: net2-rocm-mi350-microarch flagged '76.8' missing precisely because 76.8 GB/s is that record's own 38.4 Gbps x 16-bit derivation and correctly does NOT appear in the source. A missing derived figure is the instrument working. WHAT THE SWEEP FOUND: exactly one class of true content change, the NVIDIA workstation API rows in w3s-nv-rtx-pro-6000-and-4500-api-row-stale-in-6000-sheet. Every other 'missing' figure resolved to an instrument artefact on inspection. THE HONEST NEGATIVE: 309 of 368 html sources and 45 of 47 PDFs re-verified with content UNCHANGED is a statement about those fetches on 2026-10-04, not a guarantee. A pricing page that renders client-side was not verified at all; only a browser session verifies it, and the Arc Pro B60 Intel PDF was 403-gated and therefore unchecked.

Notes

METHOD RECORD, so the audit is reproducible and so its NEGATIVES can be re-checked rather than inherited. url is a representative live NVIDIA LLM-readable spec endpoint used as the worked example; the sweep covered the whole of data/sources/, not this page. WHAT WAS RUN. (1) Every PDF source URL re-downloaded and text-extracted via the PDF text layer, not a rendered screenshot: 45 of 47 distinct PDFs parsed. (2) Every html source URL of kind spec-sheet/database/benchmark/release-notes re-fetched: 368 URLs, 309 returned 200. (3) For each record, the numeric figures in its notes field were searched for in the freshly-fetched text, comma-insensitively. A figure that no longer appears in its own source is a candidate silent revision. THE INSTRUMENT IS BLUNT AND ITS FAILURE MODES ARE THE POINT. It cannot tell a revision from a paraphrase, from a figure the record derived rather than transcribed, or from a figure that lives in a client-rendered table. Confirmed instrument failures, all kept as negatives rather than promoted to findings: pwr-mlperf-v4-1-h200-maxq-l70-server-summary flagged 9 missing figures because the record cites a github.com/BLOB/ html page whose numbers exist only in the raw file - fetching raw.githubusercontent.com instead cleared 6 of them; vb-cambricon-*, pd-cerebras-pricing-dom-rendered and every vast.ai / fireworks.ai / baseten.co pricing source flagged missing because those pages render their tables client-side and return under 4KB of shell to a static fetch, which is EXACTLY the repo's rent-cerebras-api-token-pricing-negative false negative and must not be repeated; sup-cloud-together-gpu-clusters-pricing and sup-cloud-vast-ai-pricing flagged missing for the same reason while the sibling sup-cloud-coreweave-pricing re-fetched with all 14 cited figures still present. DERIVED VALUES FLAG THEMSELVES, which is a property worth keeping: net2-rocm-mi350-microarch flagged '76.8' missing precisely because 76.8 GB/s is that record's own 38.4 Gbps x 16-bit derivation and correctly does NOT appear in the source. A missing derived figure is the instrument working. WHAT THE SWEEP FOUND: exactly one class of true content change, the NVIDIA workstation API rows in w3s-nv-rtx-pro-6000-and-4500-api-row-stale-in-6000-sheet. Every other 'missing' figure resolved to an instrument artefact on inspection. THE HONEST NEGATIVE: 309 of 368 html sources and 45 of 47 PDFs re-verified with content UNCHANGED is a statement about those fetches on 2026-10-04, not a guarantee. A pricing page that renders client-side was not verified at all; only a browser session verifies it, and the Arc Pro B60 Intel PDF was 403-gated and therefore unchecked.

Cross-references

This record references no other records.