fix(site): a ranking of seeds was printed as a ranking of formats — and the PASSPORT

[gHashTag]

An audit of the public claims confirmed 25 false or unsupported statements. This lands the site half, plus the document they were audited for.

Withdrawn

• The neuron table's MHz/LUT and # columns. LUT count is bit-identical across all five synthesis seeds; the frequency it was divided by is not. The column ranked seeds, not formats — and the rank beside it counted a position among twenty off eleven printed rows. The table is now read by area.
• The same frequency axis in the posit prose, the hero card, and twitter:description — which had been contradicting its own og: twin.
• Two citations that named T25 for results T25 does not give; they now name the theorem that does.
• README.md: GF16 dynamic range 10⁵× → ~6.55×10⁴×; G = π³γ²/φ → 0.09% withdrawn (it evaluates to 1.068, and its only input γ = φ⁻³ is rejected nine lines above); GF16 is not a Galois field; "0% DSP, 19.6% LUT" → 0 DSP, 4,267 LUT (6.7% of XC7A100T); "first autoregressive ternary LM on FPGA" loses "first"; throughput stated as measured (~34 tok/s) with the projection marked as one; power → not measured, no instrument or method on record.

Added

The PASSPORT — the record a reported result must carry, proposed to the OCP neuromorphic working group, with the three measured cases of ours that pay for it: a false difference (4.44 SE between two readings of the same weights), a false agreement (43.70% of parameters differ behind a metric that does not move), and an incomparability (5.4× LUT).

One content module, two faces: /passport (the record) and /passport/research (the evidence), plus a fourteenth view in the Queen's rail on key b. The site and the working group read the same sentences, so they cannot drift.

The atlas is generated with scan provenance keyed to a commit — a cell added there by hand would be a fabricated provenance inside a document about fabricated provenance. The entry goes through the module contract instead, which qa/agents-spec-contract.mjs checks (14 views, 14 modules, keys 1234567890tprb).

Still owed

src/content/tnf.ts:273-274 carries two quantitative claims ("LUT is bit-identical across all five seeds", "up to 37 of 210 pairwise throughput verdicts change winner") that are public on the OCP list but have no in-tree provenance. The 463 LUT hero card is still tagged measured.

🤖 Generated with Claude Code

Work report

{
  "version": 1,
  "head_sha": "d848712dcee82c607b0dec9c0351db9e216e72ed",
  "summary": "An audit of the site's public claims found 25 false or unsupported statements; this lands the site half of the fix plus the PASSPORT document those claims were audited for.",
  "changes": [
    "Withdrew the neuron table's MHz/LUT and rank columns: LUT count is identical across all five synthesis seeds but the frequency divided into it is not, so the column ranked seeds rather than formats.",
    "Removed the same frequency axis from the posit prose, the hero card and twitter:description, which had been contradicting its own og:description twin.",
    "Repointed two citations that named T25 for results T25 does not contain, so each now names the theorem that actually gives the result.",
    "Corrected README claims: GF16 dynamic range, the G formula that evaluates to 1.068, the DSP and LUT utilisation figures, and throughput now stated as measured with the projection marked separately.",
    "Added the PASSPORT: one content module rendered as two faces, /passport for the record and /passport/research for the evidence behind it.",
    "Added a fourteenth view to the Queen's rail on key b, routed through the module contract rather than written into the generated atlas by hand."
  ],
  "tests": [
    {
      "command": "npm run -s check:passport",
      "result": "Passport figures contract passed all 195 checks against the committed content module.",
      "status": "passed",
      "evidence": "Printed 'Passport figures contract: PASS (195 checks)' and exited 0."
    },
    {
      "command": "node qa/agents-spec-contract.mjs",
      "result": "Module contract holds with 14 Queen views, 14 modules and the key set 1234567890tprb.",
      "status": "passed",
      "evidence": "Exit code 0; output reports 'Queen views 14, modules 14, keys 1234567890tprb'."
    },
    {
      "command": "npm run -s build",
      "result": "Vite build succeeded and emitted the passport as its own lazy chunk.",
      "status": "passed",
      "evidence": "Exit 0; emitted assets/index-Lacmn5nj.js at 325.45 kB and assets/Passport-cw2pj-_a.js at 81.00 kB."
    },
    {
      "command": "npm run -s check:silicon",
      "result": "Silicon coverage self-test passed and every spec naming a verified format is classified.",
      "status": "passed",
      "evidence": "Exit 0; scanned 1407 specs, 19 mapped, 7 family, 23 excluded."
    },
    {
      "command": "npm run -s check:queen-honesty",
      "result": "Queen honesty contract fails on one decision-line assertion about the gold block.",
      "status": "failed",
      "evidence": "Exit 1 with the identical single failure on origin/main, so it is inherited rather than introduced here."
    },
    {
      "command": "npm run -s typecheck",
      "result": "TypeScript reports 179 errors spread across 26 files, all of them outside this change.",
      "status": "failed",
      "evidence": "Exit 2; intersecting the 26 erroring files with the 30 files this branch changes yields an empty set."
    }
  ],
  "limitations": [
    "The Queen honesty contract and the TypeScript build were already failing before this branch and are not repaired here; the typecheck failure was confirmed to touch no file this branch changes.",
    "src/content/tnf.ts lines 273 to 274 carry two quantitative claims that are public on the OCP mailing list but have no provenance recorded inside this repository.",
    "The 463 LUT hero card is still tagged as measured, which overstates what a post-route utilisation report from an open flow can establish.",
    "The inherited conclusion that an accumulator can be sized before it is built reads more strongly than the data supports, since the predicted value was 0.3466 against a measured 0.3700.",
    "No die of any design described here exists, so every hardware figure is a post-route estimate for one Xilinx Artix-7 part rather than a silicon measurement."
  ],
  "tags": [
    "passport",
    "provenance",
    "website",
    "audit"
  ],
  "blog": {
    "title": "A ranking of synthesis seeds printed as a ranking of number formats",
    "summary": "An audit of the public claims on the site confirmed 25 false or unsupported statements, the sharpest of which divided an invariant area figure by a frequency that moves between placer seeds. This lands the corrections together with the record they argue for.",
    "outline": [
      "How a column labelled MHz per LUT came to rank synthesis seeds instead of number formats, and why the area half of that ratio was the only half that stayed still across all five seeds.",
      "The two descriptions of the same page disagreed with each other in the page source, which is how the withdrawn frequency ranking survived long enough to be quoted back by a reader.",
      "What the PASSPORT asks a reported result to carry, and the three cases from our own measurements that pay for it rather than an argument from first principles.",
      "Why the erratum travels with the document instead of being edited quietly out of it, which is the practice the whole record exists to argue against."
    ]
  }
}