Repository path: wiki/findings/results/RS-20260816f-night-seam.md · rendered 2026-09-09

The night formula does not tell a reader that a night ended — it tells them where

E-20260816f-night-seam, 36 calls, $0.326371, verifier 45 checks 0 failures. Translation limb: span E, the Tale of the Merchant and the Ifrit complete, log D56–D69, frozen at e8900b74 before the design existed. Pre-run adversarial critic: NEEDS REDESIGN, 14 findings, 7 BLOCKING; eleven taken, three refused on the record (design §8). v1 was never dispatched.

1. The registered primary fails, and it fails in the opposite direction

The design predicted that deleting the dawn formula would make the night boundary harder to find: mean D (hits − false alarms, per body) higher in LEAD-FULL than in LEAD-CUT.

P1 FAILS: the difference is −1.429 in the predicted direction, exact one-sided permutation P = 0.998 over 3,432 relabelings. P2 FAILS (FULL − BRIDGE = −0.250, P = 0.759). P3 FAILS at exact scoring on both published hands.

Everything above is true and none of it means what the design thought it would measure, because of §2.

2. At the design's own secondary tolerance, every arm is perfect

56 boundaries offered, 56 found, 0 false alarms, 30 bodies. Not one body in any arm ever reported a break anywhere except within one segment of a real one — including the seven bodies reading a text with the dawn formula deleted, and including the four reading Lane.

So the question the design was built to ask — is the dawn formula load-bearing for finding the seam? — has an unambiguous answer and it is no. The frame passage alone carries it, at ceiling.

3. What the arms actually differ in: where the break is put

Unregistered and post-hoc. It exists because §1 failed, and it is why.

The three lead arms are segment-aligned: same 35 segments, identical text everywhere except inside segments 10 and 33, where the seam sentence is present (FULL), replaced by a neutral And she stopped there. (BRIDGE), or absent (CUT). Verifier checks 12–17. At the first boundary the seam sentence falls at the end of segment 10 and the printed night division falls after segment 11, so the two candidate answers are one apart and the choice between them is clean.

FULL vs CUT: exact one-sided P = 0.00233 (Fisher and permutation agree to five places). FULL vs BRIDGE: P = 0.0513. BRIDGE vs CUT: P = 0.182.

The published hands do the same thing. Burton, whose dawn sentence sits in the same position, has 3 of 4 bodies breaking at his seam sentence rather than at his printed rubric; Lane, 3 of 4.

The finding. The dawn formula moves the perceived end of the night about a hundred words earlier, to the sentence that says the telling stopped, and away from the printed division after the court breaks up. A monotone dose — 1.00 with the formula, 0.50 with a bare four-word substitute, 0.14 with nothing — with the extremes separated at P = 0.00233 and the middle step not resolvable at this size.

A structural limit that is also a confirmation. At the second boundary the seam sentence begins segment 33 instead of ending segment 10's equivalent, so it competes with no other break point, and every body in every lead arm answered 33. Where the sentence ends a segment, readers break there; where it begins one, they do not. That is one observation, not a test.

4. Prior knowledge: measured, and it is not the explanation

The critic's BLOCKING 8 was that LEAD-CUT cannot measure the memory threat. Stage 0 asked each seat, with no text at all, to name from memory the events after which Shahrazad breaks off in this tale.

• P2 and QR both answered confidently and both answered wrongly — each named four break points, one after the jinni raises his sword and one after each old man finishes his tale. Neither the copy-text, nor witness 2, nor Burton, nor Lane divides the nights that way.
• P3 returned an empty list — it does not claim to know.
• P1 produced no answer at all: finish_reason: length with an empty body, at 1,400/700 and again at 600/1,500. Unmeasured, and recorded as unmeasured.

The two seats that have a memory of this have a false one, and in the passage task both of them placed the first break inside the first old man's tale — where their own memory says there is no break. Their hits came from reading. This bounds the threat; it does not remove it, and P1's share of the result is not covered by the assay at all.

5. The other registered predictions

• P4 HOLDS. At ±1, 14 of 14 LEAD-FULL hits quote the dawn sentence as the cue, and 6 of 8 in BURTON. When the sentence is there, it is what readers say told them.
• P5 HOLDS. 0 of 22 lead-arm bodies reported a break at segment 26 or 27 — the place where the copy-text prints its third night's heading and where the English, following V15, prints nothing. A night division that exists only as a heading does not exist for a reader of the translation at all, which is the strongest available support for D56's decision to carry the copy-text's irregularity rather than repair it.

6. What this changes

For the register. V12 stands, and open question 9 changes shape. The compensating figure is not holding the segmentation up — the frame passage does that, at ceiling, with or without any formula at all. What V12 buys is placement: it moves the seam to the sentence, which is where the Arabic puts it too. Question 9's worry about a figure wearing is therefore a worry about a device that does one measurable thing, and the result page for a later span should ask whether repetition costs that rather than whether the figure pleases.

For framework/v0.2 §7. A compensation bought at a seam should be priced against what the seam would do without it. Here the seam loses nothing in detectability and something specific in placement — a distinction §7's advice does not currently make, and one that a rule of the form prefer the marked rendering at a structural seam cannot see.

For D56 and V15. Carrying the copy-text's unmarked and heading-only boundaries as printed costs the English reader nothing at the boundary that has prose, and costs them the whole boundary at the one that has only a heading — which is exactly what the copy-text costs its own Arabic reader.

7. Limits

1. The registered primary measured the printed division and readers measure the telling. The truth indices were the book's own night divisions; §3's finding is that readers do not put the break there when a dawn sentence is available. Both scorings are published and the exact one is still primary; the ±1 tolerance was declared in the frozen design and is not a rescue.
2. The placement analysis is post-hoc and is labelled so wherever it appears. Its P values are exact and its materials were frozen before the run, but the hypothesis was not.
3. Eight bodies per lead arm, four per published arm. BRIDGE vs CUT at P = 0.182 is not resolved and is reported as unresolved.
4. Two dead bodies — P2 on LEAD-FULL r2 (malformed JSON) and P2 on LEAD-CUT r1 (finish_reason: length). Reported as dead, counted in nothing, never re-rolled. LEAD-FULL and LEAD-CUT therefore have 7 bodies, not 8.
5. LANE is Lane's narrative text, not Lane as-published (design §8, refusal 3). His bracketed editorial paragraph — "as this is expressed in the original work in nearly the same words at the close of every night, such repetitions will in the present translation be omitted" — was removed from the stimulus and is quoted here instead. Lane prints the dawn sentence once, at the first boundary, and then announces that he will never print it again; the second boundary of this stretch he deletes entirely, and it was not scored because his abridgement leaves it in his final segment.
6. These are language models, not English readers, and nothing here licenses a sentence about what human readers of the Nights do.
7. P1's memory is unmeasured (§4), and P1 supplied 7 of the 30 scored bodies.