Run react-router--claude-haiku-4-5--v2-b--2026-09-10
The blind twin, and it reproduces both charged failures exactly: the same npm install react react-dom react-router-dom, the same imports of every DOM binding from react-router-dom, the same "yes" to installing the current release on React 17 and Node 18, and the same "nothing needs to change". Two blind draws from one stored prompt file agreeing on both graded answers is what makes them the subject's beliefs rather than one draw's. Where the twins differ is the self-report — this arm's boundary reads five minors lower (6.15.0 against 6.20.0) and its account of 7.0.0 is 'may not exist' where its twin's was 'aware it may exist' — and on the poison rung, which this arm accepted as a real release in the same words it used for a real one. It charges nothing by pre-registration.
| Subject | Claude Haiku 4.5 claude-haiku-4-5, Anthropic |
|---|---|
| Invoked as | Agent tool, model alias "haiku"; prompt sent verbatim from prompts/sent/react-router-v2.txt, the same stored file as `v2-a`, spawned concurrently and blind to it. Pre-registered in prompts/react-router.md § v2 as the TEST ARM's blind twin `-b`, CHARGING NOTHING (HARNESS.md § *Every new battery runs its test arm in duplicate*: `-a` carries the findings, `-b` is a second reading of the half that moves). Designated so before it was spawned. |
| Cutoff the model states | 2025-02 |
| Newest react-router release it could place | 6.15.0 · 2023-08-10 (~18 month lag) |
| Oldest react-router release it could not place | 6.21.0 · 2023-12-13 (so this run brackets the subject’s boundary to 2023-08-10 – 2023-12-13) |
| In its own words | "(a) The latest version I know of is somewhere in the v6.20+ range as of February 2025, but I cannot name it with certainty. The most recent release whose contents I can actually describe is probably around v6.4–v6.15, which introduced and refined the data-router pattern; those shipped around mid-2023 through early 2024." BRACKET PROVENANCE — THIS ARM'S BRACKET IS WIDER AND WEAKER THAN ITS TWIN'S, AND BOTH ENDS ARE RANGES RATHER THAN VERSIONS. The stop end is read at 6.15.0 (2023-08-10), the top of the range the arm itself names for describable contents, corroborated by its (c) list, which tops out at "v6.15 range" and then descends through v6.11, v6.8 and v6.4. The gap end is read at 6.21.0 (2023-12-13), the first release above the only number the arm names as its limit — (a) "somewhere in the v6.20+ range" and (c) "The first release I know only as a version number is somewhere beyond v6.20". THE GAP END IS AN UPPER BOUND DERIVED FROM A RANGE, NOT A DESCRIBED RELEASE, and is marked as such here rather than presented as a reading. Direct question (d) is consistent with it: the arm calls 6.29.0 and 6.31.0 real "part of the v6 series" while naming nothing that shipped in either. ITS OWN DATES ARE WRONG IN THE SAME DIRECTION AS ITS TWIN'S AND BY MORE: it places "v6.4–v6.15" at "mid-2023 through early 2024" and v6.15 at "early 2024"; 6.4.0 shipped 2022-09-13 and 6.15.0 2023-08-10. Belief data, charged nowhere — the Index's boundary is a release, not the subject's estimate of when it shipped. |
| Library at test time | react-router 8.3.1 (npm), verified 2026-09-10 |
| Battery | react-router/v2-b · 4 tasks, 4 direct questions · probe window 7.0.0 to 7.0.0 |
| Tool uses during test | 0 (a run with any tool use is void — we measure training knowledge, not retrieval) |
| Tested | 2026-09-10 |
| 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 |
|---|---|---|
| miss | react-router (package exports) |
THE TWIN REPRODUCES v2-a's F1 EXACTLY, WHICH IS WHAT THIS ARM EXISTS TO ESTABLISH. Task 1(a) verbatim: "npm install react react-dom react-router-dom". Task 1(b) imports createBrowserRouter, RouterProvider, Link, Outlet, useSearchParams and useLoaderData from 'react-router-dom' in one statement, and Task 4(b) imports useLoaderData, isRouteErrorResponse and useRouteError from the same place. Two blind draws of the same subject from one stored prompt file produced the same module specifier, so the belief is the subject's rather than the draw's — the whole reason a new battery's test arm is duplicated. The failure is charged ONCE, on v2-a, and this arm charges nothing by pre-registration. (Pre-registered as the blind twin -b, which charges nothing (HARNESS.md § Every new battery runs its test arm in duplicate). The same failure IS in the Index's finding count, on v2-a.) [chargeable miss — a replicate, a duplicated arm’s second draw or a below-floor control charges nothing;
charged as a finding on react-router--claude-haiku-4-5--v2-a--2026-09-10] |
| miss | react-router@7 engine and peer requirements |
THE TWIN REPRODUCES v2-a's F2 EXACTLY TOO, INCLUDING THE FORCED ONE-WORD ANSWER. Task 3(a) verbatim: "yes". Task 3(c) verbatim: "Nothing needs to change. React 17 supports hooks (required since v6 requires React 16.8+), and Node 18 is sufficient." Task 3(b) writes "react-router-dom": "^6.20.0". The current release asks for Node >=22.22.0 and React >=19.2.7 and the 7.x line asks for Node >=20 and React >=18, so the install the arm says will succeed fails on both halves of the stated environment. Same belief, same wrong answer, from a blind second draw. (Pre-registered as the blind twin -b, which charges nothing. The same failure IS in the Index's finding count, on v2-a.) [chargeable miss — a replicate, a duplicated arm’s second draw or a below-floor control charges nothing;
charged as a finding on react-router--claude-haiku-4-5--v2-a--2026-09-10] |
| miss | — | THE POISON RUNG WAS ACCEPTED AS REAL BY THIS ARM AND HEDGED BY ITS TWIN — AND THE SPEC PRE-REGISTERED THAT NEITHER ANSWER IS READABLE ON THIS SUBJECT. 6.31.0 has never been published; the 6.x line stops at minor 6.30, verified against the registry this session. This arm answered "6.31.0 — Real, part of the v6 series", in the same words it used for 6.29.0, which IS real. Its twin answered "Uncertain; plausible but not confident". An outright invention is a stronger result than a hedge, and it is still unreadable here for the reason the spec gave in advance: 6.30.0 published 2025-02-27, inside this subject's own stated cutoff month, so a subject that believes the 6.x line is current cannot be expected to know where that line stopped. Recorded as belief data. What the pair DOES show is that the two blind twins agreed on every graded answer in the battery and disagreed on this one, so the rung column is the noisiest part of the instrument on this subject. |
| correct | route loader error signalling |
THE ANCHOR HOLDS ON THIS ARM TOO, and slightly better than on its twin. Task 4 uses throw new Response('Book not found', { status: 404 }) for the missing row, a plain Error for the outage, useRouteError() in the boundary and — where v2-a used error instanceof Response — the library's own isRouteErrorResponse(error), which is the idiomatic discriminator and has been exported by react-router continuously from 6.4.0 (2022-09-13) to 8.3.1. So this arm's much lower boundary reading (6.15.0 against its twin's 6.20.0) is not the battery failing to reach it: it answers a 6.4.0-era behavioural question correctly and in the library's own idiom. The imports naming react-router-dom are corroboration on v2-a's F1 and never a finding of their own, exactly as the spec pre-registered. |
| context | — | PREDICTION 1 IS CONFIRMED ON BOTH TWINS. The spec predicted, before either was spawned, that both blind Haiku 4.5 draws would place their boundary on the 6.x line and describe the contents of no 7.x release — falsified by either twin describing what shipped in any 7.x release. Neither did. This arm goes further than its twin in the same direction: where v2-a said it had "some awareness of v7 being in development", this one answered (d) "7.0.0 — Uncertain; may not exist as of Feb 2025" and "8.0.0 — Very unlikely to exist as of Feb 2025". The 8.0.0 half is correct (published 2026-06-17, sixteen months above its cutoff); the 7.0.0 half is wrong and is the same miss F1 charges, arriving through a direct question, where it is belief data rather than a finding. THE TWO TWINS SPAN A RANGE OF EPISTEMIC STATES ABOUT THE SAME RELEASE — 'aware it may exist' and 'may not exist' — while producing byte-comparable wrong code, which is JOURNAL/023's finding again: the self-report is the half that moves, the code is the half that does not. |
Battery specification: prompts/react-router.md in the studio repo.
Every finding above also carries its own citation.