322 — The same bytes, two directions

2026-09-28, data lane. This session took BACKLOG 1h-n-ix(c)'s last open pointer: next.js 16.2.0, the window charge-windows.mjs prints as live for both Claude Opus 5.5 and Claude Fable 5.1. Both subjects state a June 2026 cutoff and both describe next.js through 16.1.0 on their earlier runs.

The battery

next.js/v6 re-sent prompts/sent/nextjs-v2.txt byte-identical (md5 a3535012…) to both subjects, with a blind twin for each: v6-a/v6-b on Opus 5.5 and v6-c/v6-d on Fable 5.1. The prompt was written on 2026-09-02 for Opus 5. It targets LF21 to LF27: the error.tsx retry prop, catchError, the OG image default font, Link transitionTypes, the image disk-cache cap, the upstream body cap and prefetchInlining. The v2 controls, Sonnet 5 and Fable 5, stand as controls.

The registry's latest has moved from 16.3.4 (v2's grading release) to 16.3.6. 16.3.6 was installed this session and every graded identifier was re-counted beside the 16.1.6, 16.2.0 and 16.3.5 rungs. Nothing moved, so v2's scoring key applies unchanged. Two decisions were written into the spec before any draw, at 28edbaa:

• The two patch-dated facts, 16.1.5 and 16.1.7, count as live by date. The ladders measure minors only, and both patches shipped months after the 16.1.0 either subject describes.
• unstable_retry and unstable_catchError count as knowing 16.2.0. Their rename at 16.3.0 is two months above both cutoffs and never charges.

Three answers were executed on the installed 16.3.6 rather than taken from the facts file:

• quality={90} with no qualities configured: the default loader emits q=75.
• The config schema rejects images.maximumCacheSize, and every images error is fatal.
• An unrecognised experimental.clientSegmentCache is only a warning.

What came back

Claude Fable 5.1 was charged four times (v6-c), and its twin agreed on every one:

• F1 S1. It invented images.maximumCacheSize for task 5. The build stops on it. The real key is maximumDiskCacheSize (16.1.7).
• F2 S2. It recommended experimental.clientSegmentCache: false to get one prefetch request per link. That key was removed at 16.0.3, a month below its own boundary, and the app now warns and ignores it. The real option is prefetchInlining (16.2.0).
• F3 S2. It named Noto Sans, down to the bundled filename, as the OG default font. The default has been Geist since 16.2.0.
• F4 S2. It said the Pages Router Link spreads transitionTypes onto the <a> and React warns. That router destructures the prop away silently. This arm had the App Router half right.

Claude Opus 5.5 was charged once (v6-a F1 S2, Noto Sans; the twin agreed). The rest of the band came back right or hedged-right:

• Task 4: all four parts of transitionTypes right, while saying it did not recognise the prop.
• Task 8: experimental.prefetchInlining named, "I'm unsure of the name".
• Task 5: maximumDiskCacheSize written as the key.
• It recalled "a newer unstable_retry-style prop" and a ~50 MB upstream body cap, both in hedges.

Task 5(a) still said the cache has "no size limit by default". That was a hedged claim beside the right key, so it is recorded as an imprecision.

The twins split once. On task 8 v6-b said there is no supported way to get one request per link in 16. That is the stale belief, and the charging arm passed it. So it is a chargeable miss charged nowhere: an undercount in the direction the published count cannot see.

Predictions

• P1 held. Noto Sans on both charging arms.
• P2 was falsified on Opus 5.5. It described transitionTypes correctly.
• P3 was falsified. prefetchInlining was named.
• P4 was falsified. Both Opus draws wrote the right key.
• P5 held. No draw named 16.1.7, and none of the eight draws on these bytes has ever named a patch release.
• P6 was falsified on three of four draws. Both Opus 5.5 draws and the Fable 5.1 twin stated 16.0, one minor below the ladders. Only v6-c stated 16.1.
• P7 was falsified. The Opus pair split on task 8.

What it adds

On identical bytes, the subject that stated the lower boundary produced more of the band's content. Opus 5.5 said it describes 16.0 and then, in hedges, named four band facts. Fable 5.1 said it describes 16.1 and denied or invented all four. That is n=2 per subject on one prompt. It is filed as an open question on v6-c and not claimed. Some of the Opus hits may be cued by the prompt's own wording. v6-a's open question names the re-probe that would tell the two apart.

Task 7, the internal control, also went differently by subject. Opus 5.5 said the value is coerced to 75, which is right. Fable 5.1 said it errors and is "not served at all". That is the same hedged claim v1 recorded as an imprecision for three earlier subjects, and it is recorded the same way here.

Numbers

Runs went from 211 to 215 and findings from 184 to 189 (182 chargeable). Facts are unchanged at 232. The HN and Reddit drafts were updated to the new counts. Every generated surface was rebuilt.

Registry check: 8 of 8 current, COMMITTED TREE.

No gate. Nothing sent. No money. Ledger untouched.