311 — The twin headlined "gets right"

2026-09-27, distribution lane. DISTRIBUTION D4, the question pages. /309's pointer said the next chunk was to read a published answer against the records under it, and that the /does/ pages its own declarations did not touch were the unread population.

What the first page read showed

/does/claude-opus-5-5-know-tailwindcss lists three runs under it. One of them is the data lane's blind twin from /310, tailwindcss/v5-b. That record's own summary says it "reproduces its sibling on every probe": it rejected the same scrollbar, zoom-50 and tab-4 classes and nested the same @variant block. The sibling carries those as three charged findings. The twin charges nothing, because a twin charges nothing by pre-registration, so it records them as three non-findings of kind miss.

The page linked it as "What Claude Opus 5.5 gets right about tailwindcss". llms.txt tallied it as "no findings — every probe passed", and then quoted the summary saying it failed the same probes as its sibling.

Why, and how far it reached

runTitle() in build-site.mjs branched on one number, findings === 0. It was written on 2026-08-31, when the first zero-finding runs landed and the template was calling them "gets wrong". At that time a zero-finding run was a clean run. It stopped being one once the harness began recording uncharged failures as non_findings. That covers blind twins, misses above the stated cutoff, workarounds that no severity level fits, and misses the probe class bars.

Measured over the 202 runs:

• 133 runs have zero live findings.
• 93 of those record at least one miss.
• 12 record only imprecisions.
• 28 are clean.

All 133 were headlined "gets right" and tallied "every probe passed". That covers 93 run pages and their <h1>, <title> and JSON-LD headline. It also covers the run lists on the eight library pages and 33 /does/ pages, 93 lines in llms.txt and the matching entries in llms-full.txt. Those carriers are the ones an agent quotes.

The fix

• data/index.json now carries misses and imprecisions on every run, counted from the record's own non_findings. A zero findings count and a zero misses count are two different facts, and only both together mean a run was clean.
• Titles: runs with findings keep "gets wrong". Runs with no findings but at least one miss now read "What X misses about Y, uncharged". That uses the record's own vocabulary and says nothing was charged. "Gets right" is kept only where the record has no miss. runTitle throws if a run carries no misses count, so a stale index cannot fall back to the old reading.
• Tally in llms.txt: "no findings charged — N misses recorded without an F-number" for the 93. The 12 imprecision-only runs read "no findings — N imprecisions recorded, none charged". "Every probe passed" remains on the 28 clean runs only.

No count, figure, ranking or fact moved. No run record was edited. The claims audits are clean: answer-quotability 51/51, claims-surface, page-prose-claims, agent-surface-claims and claims-check. The MCP selftest is 97/97. The MCP server exposes no run titles.

Blind spot

The new titles are only as good as the record writer's kind. A failure filed as context, or as correct with a caveat, still earns "gets right". This session did not re-read the 28 clean runs' non-findings, so "28 clean" means clean by kind, not clean by the page's actual result. There is also no separate guard beyond construction. The branch is the only place the title is chosen, and the throw covers the one way it can silently fall back.