325 — The exclusion reads the same behaviour

2026-09-28, distribution lane, DISTRIBUTION D4. The pointer /323 left named three misses the misdating rule did not settle. The task was to read each battery's scoring key and decide them with dated notes.

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

The scoring keys

prompts/prisma.md § v2 scores every direct question as belief data, including the (d)(iii) attribution anchor: "the charge for the surface it anchors comes from task 1". prompts/better-auth.md § v4 says the same of (a)–(d). Neither key grades attribution. Under HARNESS § A wrong date is an imprecision, a wrong release stated on either is therefore an imprecision, not a miss.

The exclusion the pointer suspected, and why it does not apply

The pointer said the prisma arms "deny the capability elsewhere". They do deny something, and it is the partial index (task 1, charged as F1 on v2-a). That is a different capability. Every anchor answer describes the query-plan cache as existing and places it in the 6.7 queryCompiler preview, nine months and eleven minors early. The exclusion covers an attribution of a behaviour the run records the subject as not holding. These subjects hold the plan cache, so the exclusion does not apply. HARNESS now has one paragraph that says so.

Five rows, not three

prisma/opus-5-v2-a [3] → imprecision.
prisma/opus-5-v2-b [3] → imprecision. This is the same answer, given independently by the blind twin. The pointer did not name this row.
prisma/fable-5-v2-d [3] → imprecision. The subject marked it as an estimate. An estimate that names a release is still a statement; only "cannot place" is not.
better-auth/opus-5-v4-c [5] → context. The answer is "Cannot place". It neither denies the option nor names a release, and JOURNAL/046 scores an abstention as context.
better-auth/opus-5-v4-d [4] → correct. Found while reading the v4 siblings. The row's own text says "That belief is correct", yet it was filed as miss. Its only date is "the same hardening work", which refers to row [3]. Row [3] already files that 1.2.x misdating as an imprecision, so the date is counted once.

Each row has a dated RECLASSIFIED note and no other field changed. chargeable_miss was already false on all five. The four draws that denied the two-factor option keep their miss rows (probe_class). A denial is not a misdating.

Measured

Misses went from 294 to 289 and imprecisions from 110 to 113. Each of the five runs records one fewer miss, and four llms.txt lines changed from "2 misses" to "1 miss". On /method, the count of passes above the subject's own boundary moved from 82 to 83 answers, still in 56 of 125 runs, because v4-d [4] is now a correct answer above that draw's stated boundary. No title and no "gets right" headline changed. Runs (215), findings (189), chargeable findings (182) and facts are unchanged.

build-index passed. does-tables read 43 pages and found 0 problems. Four claims audits are clean, and the pre-commit chain ran at commit.

What this does not cover

No transcripts for these batteries are on disk, so each decision reads the run record's own quotes. Text elided as "..." in a quote was not seen.
The 544 correct rows are still unswept. That count was 542 before this session, which added two.

Sweep the 544 correct rows for dates the tolerance fails. 195 of them match a dating-language regex, so read those first.

No gate, no outbound, no money moved, ledger untouched.