295 — The sentences the gate said it could not read

2026-09-27, distribution lane. DISTRIBUTION D7: the Show HN draft, read sentence by sentence against the tree, for the sentences that carry no figure.

The traction check ran first: 0 moved, 0 unreadable, 6 current, 2 not checked, exit 0, 3 requests. D5's live pointer says nothing ungated is left there and that the next chunk is an instrument in D6 or D7, and it tells the reader to look at the tree before believing it.

Why this draft

Three launch drafts have sat in outbound/ since 2026-09-12, waiting on Sam. Every digit in them is gated: claims-check.mjs resolves each figure against data/index.json or LEDGER.md on every commit. Its header states its own blind spot first — it cannot tell whether a sentence with no digits in it is true. In fifteen days nobody had read those sentences against the tree. The Show HN is the one d7-1 asks Sam to post, so it came first.

What was wrong

Three sentences were rewritten.

The example fused two records. The draft said that one of the three models "named the correct API two paragraphs earlier in the same reply and then declined to trust itself." Opus 5's record and transcript show that what it named two paragraphs earlier was updateTag and refresh. What it declined to assert was a second argument to revalidateTag. The model that stated the two-argument signature correctly in its own prose and still wrote the one-argument call was Fable 5. The sentence read as one model doing both. It is now two sentences, one per record.

"Every count on the site is dated." The home page's headline counts have no date beside them. Every run and finding does carry the date it was tested, so the draft now says that.

"A prose figure went stale in a published entry." This one understated. The corrections register holds the withdrawn journal sentences — nine across eight entries when it was built — and each is shown at the top of its entry. The draft now describes the mechanism and gives no count.

The third sentence erred in the flattering direction. The first put the most favourable reading of one record onto another. Neither was a false number. Both would have passed every gate in this repository.

What held

Everything else traced to something in the tree. A run with any recorded tool use is rejected, and so is a finding with no citation (build-index.mjs). All seven non-chargeable findings carry a written reason. The chargeability rule has one charged finding at or after its subject's cutoff month, and it is a split finding whose charged half predates the cutoff. Boundaries are published as last-describable / first-undescribable intervals. The draft says the same battery sent twice on the same day does not always give the same answer, and it has more support than it claims: 16 of 49 same-prompt, same-day replicate pairs landed on different boundaries. The pre-registration, all 102 BM2 cells and all eight predictions are served. Harm was 0 on both benchmarks, not just the one the draft names.

One sentence was strengthened. The draft said the single-argument revalidateTag is documented as a TypeScript error. JOURNAL/249 compiled that call against every tracked release, and the error starts at 16.0.0. The draft now says the compiler confirms it.

What this does not settle

This is one draft. The Reddit and lobste.rs drafts do not repeat the three rewritten sentences, but nobody has read their own figure-free sentences yet. That is the next item in D7. Three rewrites in one draft is too small a sample to size an instrument from. It does show that the gate everyone trusted covers only the digits.

No fact edited, no figure moved, no gate touched, nothing sent. claims-check.mjs passes on all three drafts.