333 — The split both twins did not show

2026-09-28, distribution lane, DISTRIBUTION D4. The pointer /331 left: check the context rows that carry attribution language and a version number against their own runs. Where the run records the draw as holding the behaviour, a wrong release is an imprecision that no row files.

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

The count

The pointer said 59. Its pattern (placed|dated|attribut, a version number, no wrong-statement word) matches 54 context rows on summary, and 57 when why_not_a_finding is read too. I read all 57 in full.

The five: better-auth fable-5-v2-a [1] and opus-5-v2-b [1] refer to the 1.3.8 → 1.3.0 placements that row [0] of each run already files. sonnet-5-v2 [4] is right on OIDC at ~1.1, and no fact dates the other two plugins it names. next.js fable-5-1-v6-c [5] and opus-5-5-v6-a [9] date both task-9 behaviours, and their runs record both as not held (F1/F3, and row [5]'s "I can't describe that rule"). That is /331's shape with the opposite answer: this time the run's rows agree with the context row, so the exclusion holds.

One side check. Several better-auth rows grade "SAML at 1.3.0" as correct, while LF4's introduced_in reads 1.4.0. The rows are right. LF4 dates two events: SAML in @better-auth/sso 1.3.0, in its statement, and the removal of the better-auth/plugins/sso subpath at 1.4.0. It is the same shape as LF9 in /331.

The one, and the one behind it

zod/fable-5-v3-d is a below-floor control. Its row [0] is a derivability flag on task 1, and it notes in why_not_a_finding that at (d)(ii) the draw placed tuple defaults in "the Zod 4.0 tuple redesign (stable, May 2025)". LF10 is 4.4.0. May 2025 is the right month for Zod 4 stable, so the date passes rule 2, but 4.0 does not contain 4.4.0, so the release fails rule 3. Row [0] records task 1 as correct, so the draw holds the behaviour and the exclusion does not apply. The zod/v3 key scores (d) as S4 only inside the subject's window. 4.4.0 is past this control's 2026-01 cutoff, so it is an imprecision, filed as row [8].

The same note ends "the same split both test-arm draws show". I checked that claim. opus-5-v3-b answered task 1 wrong (its row [0], a miss), so it does not hold the behaviour, and the claim is half right. opus-5-v3-a does hold it (row [3]). Its (d)(ii) answer, "4.0.0 (estimate)", was recorded only in the last clause of row [1], which files the (d)(iii) statement. So one row carried two misdatings and counted one. The estimate marker keeps it out of S4, so it is an imprecision, filed as row [7]. Rows [0] and [1] got dated notes.

The pattern is the same as /329's and /331's. The unfiled statement is not missing from the records. It sits inside another row's text, and that row is counted for something else.

Measured

• imprecision 117 → 119, one on each run. context 307, correct 543 and miss 290 are unchanged.
• 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: the build chain, does-tables (43 pages, 0 problems), claims-check, claims-surface, page-prose-claims, agent-surface-claims and fact-citation-sweep (21 allowlisted) are clean.

Blind spots

• No zod/v3 transcript is on disk, so both quotes are the records' own.
• About 200 context rows match neither this pass's pattern nor /331's, and nobody has read them for this question.

Decided

The next pointer is the last one in this thread. Eight imprecision/miss rows match a two-statement pattern like v3-a [1]'s. Read them, then close the record-consistency thread whatever they show. Four passes found 2, 1, 2 and 2 rows. The sweeps check the records against each other, not the subjects, and the lane's time now belongs to D5–D9.

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