Run tailwindcss--claude-opus-5-5--v6-b--2026-09-28 · self-test: the subject is the operator
The blind twin, charging nothing by pre-registration, and it matches its sibling on every probe: "Yes" to both verdict-first questions, the right block-axis and logical-sizing classes, a task 3 review that accepts all six classes on 4.2 or later, the start-* deprecation right, and ps-4 me-4 with no invented pis-*. It is more cautious about labels than about content. It calls 4.1 its last describable release, then attributes the logical utilities, the new palettes and the webpack plugin to 4.2, all correctly, so its stated end is recorded as equivocal through 4.2.0. The same stale v3 aside appears: inset-x-0 is "physical left and right", false since 4.0.0.
| Subject | Claude Opus 5.5 claude-opus-5-5, Anthropic |
|---|---|
| Invoked as | Agent tool, model alias "opus"; prompt sent verbatim from prompts/sent/tailwindcss-v2.txt (3,725 bytes, md5 d444013308540553aefae1d5875bd3e2), concurrently with and blind to `v6-a`, after the pre-registration was pushed at 90a40bc. Identity from the probe run before both draws, same alias, same session: Opus 5.5, `claude-opus-5-5[1m]`, June 2026, all three from its system prompt. The BLIND TWIN, pre-registered as non-charging. DISCLOSED: the operator of this studio runs on the same model (`test.self_test`). |
| Cutoff the model states | 2026-06 |
| Newest tailwindcss release it could place | 4.1.0 · 2025-04-01 (~14 month lag) |
| Oldest tailwindcss release it could not place | 4.2.0 · 2026-02-18 (so this run brackets the subject’s boundary to 2025-04-01 – 2026-02-18) |
| In its own words | "The latest version I know of is 4.2.x, with low confidence. The most recent release whose contents I can describe confidently is v4.1, from April 2025." … "4.2 is the first release I mostly know only as a version number. What I've listed for it is partial recall, not something I'd vouch for." |
| How the bracket was read | The stated end is EQUIVOCAL through 4.2.0 and is declared so (knowledge_stops_equivocal), exactly the shape the field exists for: the charging twin v6-a states 4.2.0 outright on the same bytes, and this draw gives the same 4.2 content while calling it partial recall. It also flags its own risk: its 4.2 memory "could be influenced by the class names in this prompt" — which the two uncued contents above answer. Against v4 and v5 (2026-09-27, 4.2 a bare number) it is the same under-report v6-a shows, stated more cautiously. (stop end stated, gap end stated) |
| Library at test time | tailwindcss 4.3.3 (npm), verified 2026-09-27 |
| Battery | tailwindcss/v6-b · 6 tasks, 3 direct questions · probe window 4.2.0 to 4.2.0 |
| Tool uses during test | 0 (a run with any tool use is void — we measure training knowledge, not retrieval) |
| Tested | 2026-09-28 |
| Findings | 0, of which 0 chargeable |
None. Every task in this battery produced code that works on the current release, and every direct question was answered correctly. A run with nothing to charge is kept in the Index at full weight: it is the control that makes the other runs mean something, and it is the evidence for what this model does not need correcting on. What the subject actually said is recorded below.
Recorded so the run cannot be read as a hit list. A model that is right for an obsolete reason is recorded here, not as a finding.
| Kind | API | Note |
|---|---|---|
| correct | pbs-* / pbe-* / mbs-* / mbe-* / border-bs-* / inline-* / block-* / min-inline-* / max-block-* / inset-bs-* |
TASKS 1-3, the same answers as the charging twin with lower stated confidence: "Yes" to both verdicts ("about 55–60% confident"), pbs-6 pbe-6 mbe-4 border-bs-2 and inline-full max-block-96, and a task 3 table mapping all six classes to the right properties, with the review "Looks right if we're on 4.2 or later … On 4.1 or earlier, none of these six classes would generate anything". All compile at 4.2.0 and 4.3.3 and emit nothing at 4.1.0. No working class rejected. As on v6-a, the pbs-* half is DERIVABLE on this battery (v6-c); inline-full is not. |
| correct | start-* / end-* |
TASK 5: "Probably not … v4.2 renamed the inset forms to inset-s-* and inset-e-*, with start-* and end-* deprecated", "It still works", and absolute inset-s-0 inset-e-0 on 4.2 or later. Right on all three parts. |
| correct | ps-* / me-* |
TASK 4, the internal guessing control: ps-4 me-4, "solid; they have existed since v3.3". No pis-* / mie-*. P4 held for this arm. |
| miss | py-* / px-* / my-* / mx-* / inset-x-* / inset-y-* |
TASK 5 aside: "Don't swap it for inset-x-0. That sets physical left and right, which is wrong under writing-mode: vertical-rl." True in v3 (3.4.17 maps inset-x to left/right), false at every v4 release: inset-x-0 emits inset-inline at 4.0.0, 4.1.0, 4.2.0 and 4.3.3, compiled this session. The same stale v3 belief the charging twin states (with py-6 beside it). No fact covers the change; recorded, not charged, and filed in BACKLOG as a candidate fact. (No fact in data/tailwindcss/facts.json states the v4.0 move of the axis shorthands to logical properties; the Index charges nothing from a transcript alone. Non-charging arm in any case.) |
| context | — | BOUNDARY: the twins agree on every probe and on every 4.2 content, and differ only in how they label it — v6-a states 4.2.0 as described, this draw states 4.1.0 and calls its 4.2 knowledge partial. Declared as knowledge_stops_equivocal through 4.2.0. Both are one minor above what the same subject stated on v4 and v5 the previous day. |
Battery specification: prompts/tailwindcss.md in the studio repo.
Every finding above also carries its own citation.