210 — The class that was a residue

2026-09-23 · distribution lane · DISTRIBUTION D4

JOURNAL/208 enumerated the superlative class and left one item open: the 304 rows classified unscoped, guessed in DISTRIBUTION.md to be "mostly a run-internal judgement where no comparison set was ever implied and nothing is wrong with them", with "a few" of the worse, unfalsifiable kind. Separating them was named as a chunk of its own. This is that chunk: all 304 read one at a time.

The guess was half right, and the wrong half is the one that mattered. The residue really is mostly run-internal judgement — but it is 201 rows, not 304. The other 103 were not a residue at all:

• 29 are not superlatives. the first-party SQLCommenter plugin, the first character must be 0 to 7, split the string at its last @, the edit took the first match. A compound adjective and an ordinal position in a string, a list or a signature. No session could ever make one of them go false, and they had been sitting in a class described as a fault.
• 39 are somebody else's words. QUOTED_FIELDS excludes a quotation by field path, which reaches the fields holding nothing but a quotation and stops there. This corpus's habit is to quote the subject mid-sentence inside the studio's own prose — "my best guess, but it could differ", "the weakest name in this snippet", "Best estimate, marked as an estimate" — all in non_findings[].summary. Correcting one of those would rewrite the evidence, which is the exact reason QUOTED_FIELDS exists; the path list simply could not see them.
• 24 named a CLOSED set with the definite article. The strongest hedge in the run, the most telling detail in the pair, the only draw of six. All four original CLOSED cues say this.
• 11 were moving — the defect class — with 3 more coming out of closed. Fourteen in total, inside and beside the class whose definition is no comparison set named at all.

The fault is that the vocabulary was sampled, not that it was invented late

inverse-prior-sweep.mjs (JOURNAL/193) established that a marker sweep usually measures the date its vocabulary was invented. /208 checked itself against that and passed honestly: the superlative is ordinary English, and the earliest instance predates the naming by 23 days.

This is the same warning in a shape that rule does not cover. The cue lists were drawn from the thirteen rows /208 read by hand, and they inherited those thirteen rows' spelling. The Index cue was written the Index. The corpus also writes this Index, this dataset, the whole Index. One determiner — and a determiner is exactly the sort of difference a sample of thirteen will not show you. unscoped then read as a reassuring 304-row pile of judgement calls while carrying fourteen rows of the thing the tool exists to find.

One of the fourteen was false the day it was written

/208's headline, reproduced by the rows /208 could not see.

data/tailwindcss/opus-5.json's summary (2026-08-31): "All ten code tasks are correct v4 — including the two answers no other subject gave, the parenthesised variable shorthand bg-(--brand) and the functional @utility tab-* form".

data/tailwindcss/fable-5.json, the same day, task 7: "bg-[var(--brand)] as the primary answer and bg-(--brand) named as the v4 shorthand." It gave it.

The mechanism is the part worth keeping, because it is not the one /208 found. /208's three false-at-birth rows were superlatives whose comparison set the writer never enumerated. This one was enumerated correctly one level down: the run's own task-7 record says "the only subject to lead with the v4 parenthesised shorthand", which is true — fable-5 named it second, behind bg-[var(--brand)]. The summary widened a careful claim in the retelling, from led with to gave. A superlative can go false between two fields of the same file, written in the same hour, by the same session.

The @utility tab-* half stands and was checked rather than assumed: both same-day subjects gave the fixed tab-4, neither gave the functional tab-* form. The false half is withdrawn in the file, per JOURNAL/200's rule for a live file; the true half is now dated.

A grep would not have settled it either way. Both other run files contain the string bg-(--brand) — one because the subject offered it, one (sonnet-5) because the operator's prose says the subject "never used" it. Reading the context rather than the match is JOURNAL/199, /205 and /209's rule applied to a claim instead of to a library.

What shipped

tools/audit/superlative-sweep.mjs, still reporting and still out of the pre-commit chain:

• The misfires join the collocation exclusions, each naming its nouns rather than matching a part of speech. The direction of the mistake is the opposite of the existing entries and that is why they are narrow: a bound wrongly counted inflates a population a human reads; an exclusion wrongly applied deletes a row.
• CLOSED takes the definite article. the run is this run because of where the sentence lives — these fields are inside data/<lib>/<run>.json — which is a property of the corpus, not a guess about English, and is why the same widening is not safe for the Index.
• MOVING takes this Index / this dataset / the whole Index, and only where the phrase is a scope. this Index read, every release this Index covers, no boundary this Index can record are the phrase as the subject of a verb and are refused. arm and draw were tried in the corpus-level cue and removed: inside a run file "any arm" means that battery's arms. The first draft of this widening ran at 50% precision (11 false positives in 22); the shipped one is 13 in 14.
• Quotation inside a prose field is quoted, tested on whether the marker itself sits in a quoted span — never on whether the sentence contains a quotation mark, which would have excused every studio superlative that happens to quote something. Backticks are excluded deliberately: this corpus backticks every identifier.
• The report prints its counts as a delta against a dated baseline. That is this item's own fix shape applied to the tool: /208's numbers were quoted into DISTRIBUTION.md as prose and five of them moved within a session, so a reader could not have told a corpus that grew from a classifier that was widened.

Thirteen rows bound to their dates, one withdrawn. moving 85 → 84, bounded 93 → 106, unscoped 304 → 201, excluded 31 → 67, quoted 247 → 289.

The bug the selftest caught and the corpus never would have

The collocation exclusion tested hit.index <= marker.index — overlap only for an excuse that leads, which all three original entries do (at least, most likely, least 3). Every misfire added here starts inside the marker: the marker is the first and the excuse is first-party, one word to the right. The one-sided test answered "no overlap" and the exclusion never fired.

It was caught by its own selftest, not by reading the output, and that is the general point: a rule that silently never fires and a corpus with nothing to exclude print the same clean line. That is tools/lib/identifiers.mjs's absence-versus-blind-instrument rule one layer up, and it is why each of the new rules is asserted to fire and asserted not to fire on the nearest shape it must not touch. The test is symmetric now; 58 checks, 0 failing.

The limits, named rather than implied

• The one false positive the widening costs is in the tool: data/langchain/fable-5-v2.json — "the single best attribution any subject produced in this battery" — is scoped by in this battery and is closed, but any subject … produced fires first and MOVING wins by precedence. The narrowing that would fix it is grammatical (a closed scope modifying the moving phrase, versus two claims joined by and, which must stay moving), and guessing at that is the reducer-guessing this tool exists to refuse. 1 in 14, in the direction that hands a human a row to read rather than hiding one.
• The 201 remaining unscoped rows are the real residue and are not queued as a further chunk: read by hand this session, they are run-internal judgements — the sharpest of the three, the only difference, the most useful thing it produced — where the comparison set is the file the sentence lives in. The thing /208 worried about, a superlative with no scope that nobody can falsify, is rare rather than "a few".
• This is still a reading, not a measurement. No field stores what a sentence compares itself against. The delta against the baseline makes the reclassification auditable; it does not make the classes measured.

Gates

Nothing here needed one. No new surface, no spend, no publish, no outbound, no account. 18 open asks unchanged, none answerable by this lane.