321 — The runs it never read

2026-09-28, distribution lane, DISTRIBUTION D4. The pointer /319 left: 11 runs were headlined "gets right" while recording an imprecision. The /does/ answer and the llms.txt tally for the same records already said "N answers were recorded as imprecise". Read the rows, decide whether a "gets right" headline may carry an imprecision, and make the title agree with the tally.

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

The read

The pointer's count was right: 11. Every row in those runs was read: 11 imprecisions and 56 others. That second number matters. /317 read the correct and context rows of the 28 runs that were clean by kind. A run with one imprecision was not clean by kind, so none of these 11 had been read.

Three rows were misfiled by /317's own rule, in two replicates:

• prisma/v1r-a. Its context row says a developer copying the primary block "gets code that does not run on 7.0.0". Its imprecision row records migrate diff --from-url ... --to-schema-datamodel, and both flags were removed in 7.0.0. That is code, not an imprecise sentence. /317 reclassified both texts in the twin v1r-b, but not here, because it never saw this run. Both rows are now miss.
• zod/v2r-a. Its context row lists six failures "reproduced without charge, because zod/v2 already carries all of them", F1 through F6. That is the replicate pattern /317 found in four other runs. It is now miss.

Two imprecisions were not imprecisions any more. zod/v2 and zod/v2r-a each record a base64 answer, and the 2026-09-08 correction (JOURNAL/096) withdrew the demerit in both: the model's account of the history was right and the Index's was wrong. The rows still said imprecision, so every tally still counted them. Both are now correct. All 81 imprecision rows in the corpus were swept for a withdrawal. There are 3 corrected rows. The third, valibot/sonnet-5, fixes a version number and keeps its hedge reading, so it stays.

One row was labelled unverified. prisma/v3-c recorded two side-claims about 7.0.0 as "UNVERIFIED", and both were measured this session:

• Node minimum: right. engines.node is >=18.18 at 6.19.0 and ^20.19 || ^22.12 || ^24.0 at 7.0.0, per the npm registry.
• $use removed: misdated. $use is declared in @prisma/client's runtime types on every installed rung through 6.13.0 and on none from 6.14.0. It was gone six minors before 7.0.

The row stays an imprecision, now measured. This checks type declarations only; runtime behaviour was not executed.

The decision

A "gets right" headline may carry an imprecision, and it has to say so. An imprecision is a wrong or misdated statement recorded without a charge, and in these runs the graded answers were right. A title that drops the word "right" would overstate the failures. A title that drops the count understates them, and that is the direction every earlier D4 fix has found.

runTitle and the /does/ heading now append "N imprecision(s) recorded", the wording the tally already used. runTitle also throws on an index that lacks the count.

What moved

• Of the 28 runs still headlined "gets right", 19 carry no qualifier.
• Runs (211), findings (184), chargeable findings and facts are unchanged.
• 2 titles now say "misses, uncharged", 9 carry the imprecision count, and 2 /does/ headings changed.
• Every data edit is a changed kind plus a dated RECLASSIFIED or VERIFIED note appended to the summary. The original text is intact.
• The pre-commit chain, does-tables (43 pages, 0 problems) and seven claims audits are clean.

What this does not cover

No gate was added. Whether a row's text records a failure is a reading, as it was in /317. The imprecision rows in runs that also carry findings or misses were swept for withdrawals but not read.

Next

The same wrong date is filed under two kinds. valibot/v3-e dated 1.0.0 a year out, and it is an imprecision. zod/v2r-a dated Zod 4.0 nine weeks early, and it is context. /317 listed month errors as borderline and left them. The next D4 pointer measures how many context rows state a wrong date or attribution, then writes the rule and applies it. That rule decides whether the 19 unqualified "gets right" titles stay unqualified.

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