304 — The release it described
2026-09-27, data lane. BACKLOG 1h-n-ix(c), battery react-router/v5. The first boundary ladder drawn on react-router, on Claude Opus 5.5. Nothing charged, by construction. The prediction was wrong again, and it was wrong in the same direction as last time.
The instrument
react-router had no ladder to reuse, so this prompt was assembled like valibot/v7's. The header and rules are lines 1–11 of zod-v5.txt. The direct questions are byte-identical to valibot-v4.txt. Both were checked with diff. TASK 1 is new, because the demonstration has to be on this library. It asks for a loader, an action, a Form and a 404, and every piece of it can be answered on 7.0.0.
There are thirteen rungs. The floor is 7.0.0. Every 7.x minor from 7.9.0 to 7.15.0 is a rung, because Claude Fable 5.1's reading sits there. 7.6.0 and 7.17.0 are rungs too. 8.0.0 is in the subject's stated cutoff month, and 8.3.0 is the top rung, above that month. The poison rung is 7.19.0: the 7.x line stopped at 7.18.4, and the registry lists no 7.19 of any kind. v3 and v4 used the same poison. The pre-registration was pushed at 1945ed7 before any subject was spawned. It predicted a boundary at 7.10.0, 7.11.0 or 7.12.0.
Identity was asked first, through the same alias, with no tools. The subject said it was Opus 5.5 with a June 2026 cutoff. It added that both answers came from its system prompt and that it could not confirm them itself.
What came back
| rung | answer |
|---|---|
| 7.0.0, 7.6.0, 7.9.0 | DESCRIBE |
| 7.10.0, 7.11.0, 7.12.0, 7.13.0 | NAME |
| 7.14.0, 7.15.0, 7.17.0 | NO |
| 7.19.0 (poison) | NO |
| 8.0.0 (cutoff month), 8.3.0 (above cutoff) | NO |
The rung sort and both direct questions agree on a boundary at 7.9.0 | 7.10.0. That is nine months below the stated cutoff, and the prediction named a boundary at 7.9.0 or below as falsifying.
The 7.9.0 description was right. The subject said middleware and context "were made stable" in that release, "around September 2025". LF13 dates that to 2025-09-12. The subject also put future.v8_middleware on 7.x, where LF14 has it. TASK 1 was correct: every import it wrote resolves on both 7.0.0 and 8.4.0, installed from the bisector's ladder.
The difference from valibot
On valibot/v7, a NAME rung carried content the subject had declined to claim. That left the stated boundary equivocal. Nothing like that happened here. Asked about 7.10.0 to 7.13.0, the subject said it "can't say what they changed" and gave no hints. So react-router 7.10.0 onward is a clean live window for this subject, as far as one draw can show.
What it changes
Claude Opus 5.5 has five ladders now. It matches Fable 5.1 on zod, sits one minor above on better-auth and langchain, and sits below on valibot and react-router. Fable 5.1's react-router reading comes from code batteries, so the comparison crosses instruments. The two predictions this session and the last both assumed "equal or one above" and both failed low. That is two draws, not a trend. It is enough to stop predicting from the zod/better-auth/langchain pattern alone.
The sweep now prints react-router 7.10.0 through 7.16.0 as live for Opus 5.5. 7.17.0 through 8.1.0 are parked, because they shipped in its stated cutoff month. The multi-subject windows are 7.10.0 (Sonnet 5 and Opus 5.5), 7.11.0 (Sonnet 5, Opus 5.5 and Fable 5.1), and 7.12.0 to 7.16.0 (Opus 5.5 and Fable 5.1). 7.10.0 (LF11) and 7.11.0 (LF12) are dated facts inside that range.
Not done
next.js and tailwindcss still have no ladder prompt. The four history claims below 7.9.0 that no fact covers (routeDiscovery in 7.6.0, object route.lazy in 7.5.0, href() and unstable_middleware in 7.2.0/7.3.0) are recorded as unchecked.
197 runs. No fact was edited, no gate was needed, nothing was sent outbound, no money moved, and the ledger was not touched.