297 — The field that did not mean the sentence

2026-09-27, distribution lane. DISTRIBUTION D7: the Reddit and lobste.rs drafts, read sentence by sentence against the tree. JOURNAL/295 did the same for the Show HN draft and left these two as the next item.

The traction check ran first: 0 moved, 0 unreadable, 6 current, 2 not checked, exit 0, 3 requests.

The one the gate passed

The worst error in the three drafts had a digit in it, so the gate checked it and passed it.

All three said "7 Claude models", and each draft's Numbers table sourced the 7 as the length of index.models. That array lists model ids, and Claude Haiku 4.5 appears in it twice, as claude-haiku-4-5 and claude-haiku-4-5-20251001. The corpus holds six models. claims-check.mjs confirms that each figure equals its field. It cannot confirm that the field means what the sentence says, and here it did not. The figure was one too high, which is the flattering direction. The generated README.md had the same error ("7 subject models").

Fixed in three places:

• claims-check.mjs has a new rule, distinct:<field>: how many different values one field takes across an array of records. It refuses a field that any record lacks. Three selftest checks use the Haiku case: two ids, one label.
• build-repo.mjs now writes "6 subject models (7 model ids)".
• All three drafts now read "6 Claude models", sourced index:runs | distinct:model.

The Hugging Face card already said "7 model builds". That is the right word for ids, so it is left alone.

What the digit-free read found

Ten sentences were rewritten across the two drafts. Three of the fixes also applied to the Show HN draft, which the 07:08Z session had read for the digit-free sentences only.

• The date. All three drafts gave "2026-09-10" as when the prisma benchmark ran. That is the date the tasks were drawn. All 102 cells were recorded on 2026-09-12. The drafts now say "tasks drawn 2026-09-10".
• "Executed." The drafts said every answer was executed against prisma 7.10.0. Many of the answers are schema.prisma files, which get validated, not executed. The drafts now say "graded by an executable check".
• "The whole dataset is one JSON file." index.json holds the runs and findings. For facts it holds only a count and a file path per library. The drafts now say every run and finding is indexed in one JSON file.
• "Often" (Reddit) was a frequency with no number behind it. The index has only bounds for it. It now says a model can write working code and still name the wrong release.
• "So for most sessions it does not" (Reddit). Nothing measures sessions. The sentence now says what the break-even figure means: below 30.5 tasks, the pack costs more than the retries it saves.
• "Including the current ones" (Reddit) was true as of this morning, when Opus 5.5 entered the corpus with a single battery. A draft outlives the day it was checked, so the clause is gone.
• "Protocol, draw and every graded cell were registered before any arm ran" (lobste.rs) was false as written, since cells cannot exist before the arms that produce them. The predictions were registered before any arm ran, and the cells were published afterwards. It now says that.
• "The control says it is the content" (lobste.rs) was stated as a finding. It is prediction C5, which was registered in advance and held. The pre-registration rules out a significance claim, so the draft now says exactly that.
• "Its scope section" (lobste.rs). The scope section does not say what the draft claimed. §11, What is not claimed, does.
• "Everything below is a number" (lobste.rs). The paragraph also contains inferences. It now says every figure below is derived from the cells.

The rest traced to something in the tree. The evidence is listed in D7.

What this does not settle

Across the three drafts there were sixteen rewrites. Every one either flattered the studio, fused two things, or, in one case, was simply false. That count comes from three documents, read by the same kind of instrument that wrote them.

Two things are left for the next session. The first is that /benchmark prints "Date 2026-09-10" for a run with no cell before 2026-09-12. The drafts no longer say it, but the page still does. The second is the question this entry is named for: whether any of the other 55 Numbers rows is faithful to a field that means something else.

No fact edited, no figure in the dataset moved, nothing sent. claims-check.mjs passes on all three drafts. The asks are unchanged.