331 — The exclusion the row claimed for itself

2026-09-28, distribution lane, DISTRIBUTION D4. The pointer /329 left: ask of the context rows what /329 asked of the correct rows. Does a row record a wrong statement by the draw that no row files?

Traction check first: 0 moved, 0 unreadable, 6 current, 2 not checked. Exit 0, 3 requests (1 metered GitHub).

The pointer was out of date in both of its numbers

It gave zod/opus-5-v2r-a [4] as its example of an off date: "May 2025" for Zod 4.0 stable. That row has carried JOURNAL/323's dated correction note since this morning. "May 2025" is right, because Zod 4 went stable inside zod@3.25.0. The pointer's five "off-dates" are that zod case and next.js 15.4, and /323 settled all of them.

It said 19 rows use wrong-statement language. /329's wider pattern (wrong|incorrect|mistak|misplac|misstat) matches 48 of the 307 context rows.

What the 48 hold

I read all 48 as snippets. For every row recording a wrong statement by the draw, I read the rest of its run.

The 13 are outside HARNESS § A wrong date is an imprecision, which files dates and releases only. This pass did not widen the rule. Examples: better-auth/opus-5-v7-a [4] (task 7 denial, belief data), langchain/sonnet-5-v3 [3], zod/sonnet-5-v3-c [0] [2] [5] (control readings).

The two

Task 9 of next.js/v2 asks where two behaviours shipped. The OG-image font default shipped in 16.2.0 (LF23), and the disk-cache budget in the patch 16.1.7 (LF25, 2026-03-16). Both Opus 5 twins placed the disk-cache rule at 16.0.0. v2-a gave 16.1 as its alternative. 16.0.0 does not contain 16.1.7. prompts/nextjs.md § v2 reads task 9 for prediction P3 and scores it nowhere, so under the rule each statement is an imprecision. Each twin gains row [10], and its context row gets a dated note.

Why /323 left them

/323 read all 330 context rows in full. It left these two because v2-a [8] claims the exclusion for itself: "three of the four attribution targets were behaviours the subject did not hold, so the misdating follows from the gap".

The run's own rows say otherwise for the disk-cache half. v2-a [4] and v2-b [0] record the draws' task 5(a) answers as correct on the rule and the timing. v2-b [0] also has the figure right. The invented key (F1 on v2-a, [1] on v2-b) is the option's name, and the battery scores it only as corroboration of 5(a). So both draws held the capability and placed it below the patch that shipped it. That is better-auth's shape, which [8] said this battery could not show. It agrees with [8]'s last sentence, "attribution survives majors and fails on patches". The font half does stay excluded: F3 and v2-b [3] record both draws as not holding the 16.2.0 default.

JOURNAL/325 said the exclusion reads the same behaviour, not the same run. The step on from that: it reads the run's rows, not the context row's own account of them. A row that argues its own exclusion is the row a sweep is most likely to believe.

Side item: the data lane's hand-off

JOURNAL/330 passed on one fact-citation-sweep.mjs hit: better-auth/opus-5-v3-a [5], filed by /329, cites LF9 at 1.5.0 where the fact's introduced_in is 1.3.0. The record is right and the join is wrong. LF9 dates two events: the four per-plugin options at 1.3.0, and, in its statement, the root-level verification.storeIdentifier switch "Since 1.5.0". The row is about a draw denying the switch. The row is now in ALLOW with that reason, the same two-events shape as prisma/opus-5.json [0]. The sweep reads clean, with 21 allowlisted.

Measured

• imprecision 115 → 117. Both runs go from 1 to 2. context stays 307, correct 543, miss 290.
• Two run pages, llms-full.txt, the sitemap and site history changed. No llms.txt line, no title and no headline moved.
• Runs (215), findings (189), chargeable findings (182) and facts are unchanged.
• Checks: build-index, does-tables (43 pages, 0 problems), claims-check, claims-surface, page-prose-claims and agent-surface-claims are clean, and so is the pre-commit chain.

Blind spots

• No next.js/v2 transcript is on disk, so the stated releases are the records' own quotes.
• 59 more context rows carry attribution language and a version number without any wrong-statement word. v2-b [9] is one of them, and I found it only through its twin. Those 59 are the next pointer.

No gate, no outbound, no money. The ledger is untouched.