Repository path: workshop/experiments/E-20260902-rhyme-slot/design.md · rendered 2026-09-09

E-20260902-rhyme-slot — where a rhyme-position sense lands in English, and whether the rhyme put it there

ARM-rhyme-family step 2 (T2). Frozen before any API call. Every number below that is not a count of stored material is a prediction.

1. The question

RS-20260830b-rhyme-family established, on Walter Leaf's twenty-eight odes, that a published English monorhymer's rhyme words almost never carry the senses his author put at the rhyme — pooled R = 0.0384, nineteen of twenty-eight odes at zero — while the same senses were found eight times as often somewhere in the last words of the same lines (R_set = 0.311). Its §8 named the successor question and this design takes it:

Is the pre-rhyme slot where the rhyme sense goes in English, in every hand, or only in these two?

Sharpened into something measurable and falsifiable: when a translator carries into English the sense a Persian ghazal puts at the end of a hemistich, how far from the end of the English line does it land — and is it further when that English line has to rhyme?

The subject-rule sentence (wiki/tracks.md §The subject rule): what this unit teaches about translating literature is where a fixed form's sense-bearing words end up in the receiving language, and whether the constraint the form imposes on the line's last syllable is what moves them. It is not a question about this project's instruments.

2. What makes the contrast possible, and why it is new

A ghazal rhymes some of its hemistichs and not others: both hemistichs of bayt 1 (the maṭlaʿ) and the second hemistich of every later bayt. The first hemistich of bayts 2…n ends on nothing in particular. Both kinds of hemistich end on a word with a sense; only one kind obliges the English line-end to chime.

So each poem carries its own control. COND is fixed by the Persian's verse form, before any English is read, and no property of any English line can move it. Step 1 had no such control: it compared poems, and had to build its null out of mismatched pairings.

Note (bsr) is discharged by construction. The predictor is a form label on the Persian; the outcome is a count of word tokens in an English line. Neither is in the other's denominator, and COND cannot move d arithmetically. This is written here because (bsr) was raised against the previous session's primary and the check it demands is a pre-flight, not a post-mortem.

3. Materials, all stored, all free

Two screens, both applied before any call, both declared here.

1. English-radif renderings are excluded — 14 of 57. Where a hand closes every rhyming line with a fixed word sequence, the line's last tokens are repeated text and the distance measure is not about the rhyme. This screen was derived today, as a free gate on the arm: re-locating the English rhyme word in each of Leaf's 28 by rime class shows it sitting outside the last two tokens in 6 odes (I, II, VII, XVI, XIX, XXV), which is where step 1's two-word en_bearer window could not see it. That gate also closed step 1's figure as sound: recomputing pooled R with those six odes dropped moves it from 0.0360 to 0.0415, so the floor is not a window artefact and no published figure is false. The screen is adopted here anyway, because this design measures position and step 1 measured presence.
2. OCR screen — a rendering is dropped if more than 6% of its tokens still carry a character impossible inside an English word after punctuation is stripped. It fires on 0 of the 43 kept.

4. Procedure

Stage G — gloss, seat P1 openai/gpt-5.6-terra. Input: one ghazal, hemistichs numbered in order, the line-final word of each marked. Output: a short English gloss of each marked word in its context, or UNGLOSSABLE. The words rhyme, qāfiya, English, verse and translation do not appear in the prompt. 41 calls.

Stage L — locate, seat P2 google/gemini-3.6-flash, a different lab. Input: one English rendering as numbered lines, and that poem's glosses in a seeded shuffle, with no indication of which hemistich any gloss came from. Output per gloss: the single English word that carries it and the line number it stands in, or NONE. 45 calls (43 published + the lead's two arms).

Screen — mechanical, in the analyser, never shown to a seat. The named word must occur in the named line. d = the number of word tokens standing after it in that line. AT-END ⇔ d == 0.

Seats are disjoint between the stages, which is what both of step 1's critics blocked on.

5. Registered predictions

6. Withholding gates and failure criteria

A gate means the figure may not be claimed in either direction. A failure criterion means the finding does not hold.

POSBIAS, the positive control. Six renderings drawn by seed. The analyser takes one line's actual last content word; P1 paraphrases it from the English line alone; P2 is then asked to locate that paraphrase in the whole rendering, in the stage-L format. A seat that can localise returns d = 0. 12 calls.

Not run, with the reason: no X-MIS control. RS-20260830b-rhyme-family §6 measured the in-field mismatch — one ghazal's Persian against another's English — at 0.1111 against the true items' 0.0384, inverted, and states that any future design pairing one ghazal's Persian with another's English "has this number against it". The prior is used rather than re-bought.

7. Declared limits, before the run

1. The gloss seat can see that the poem is a ghazal, and therefore which hemistichs rhyme in Persian. It cannot see any English. The outcome is a different seat's answer about a different text, and the route from "this word is the qāfiya" to "this English gloss is differently findable" is not one this design can rule out — only make implausible. It is stated, not solved.
2. The lead's pair is descriptive and cannot be otherwise. The lead conceived d before writing either arm (R59 §Known limitation). A hand that knows the measure can move it. The confirmatory weight sits entirely on Leaf and Payne, who could not have known it.
3. A Persian confound this design does not remove. The qāfiya of a Hafez ghazal is drawn from a stock lexicon and, in many ghazals, fuses with the copula; FREE hemistich-ends are ordinary words and sometimes function words. If CONSTRAINED senses are simply harder to place anywhere, that shows in the located rate, which is reported by condition — but a difference in placeability is not the same as a difference in position, and only the second is being claimed.
4. d is measured in word tokens from the line end, so a hand who ends lines on function words scores low d for reasons of English rhythm. Reported and not corrected.
5. Payne is OCR. 21 renderings survive a 6% damage screen; damage that survives it falls on both conditions and cannot create the contrast.
6. One seat on each side, four models, no sense Tier-D calibrated. These are descriptive lexical adjudications — the shape S015 found the panel usable on in the failing direction. provisional: true.

8. Budget

Ceiling declared $2.20 of the UTC-2026-09-02 day's $5.00, which has no prior row. Worst case built from the caps the requests actually permit, per note (abc): 41 + 45 + 12 = 98 calls, caps set per seat and per task shape by a probe first (note (bsf), fifth firing), plus 2 pre-run critic calls. Anything that does not fit is scaled down, not overspent.