Repository path: workshop/experiments/E-20260828-purchased-figure/design.md · rendered 2026-09-09

E-20260828 — where a rhyme gets paid for

v2, frozen after the pre-run critic and before any judging call. ARM-gulistan step 1's study limb. Translation limb: T-gulistan-bab2-R05-v1 span B, frozen and committed at b4daef0f with its log before v1 of this design was written and before any English of the chapter was opened (charter A4). The critic's findings and their disposition are §10; v1 is in the git history at b4daef0f's successor commit and is not reproduced here.

1. The wire

The translating produced the question. Rendering 39 of Sa'di's bayts under a rule that forbids buying a figure with a word the source has not got (V2), the translator refused six rhymes, and every one of the six was refusable for the same reason: the English rhyme was one small added word away, and the word would have stood at the end of the line (D24) — astray, at all, slain, door, gown, hall. Nothing was refused for want of a rhyme; the rhymes were there.

That is a claim about where the cost of a formal constraint falls, and it is testable on a hand that took the rhymes. Eastwick 1852 rhymes 38 of 42 bayts in this book (RS-20260822c, on باب اول), and the rhyme coding below finds 17 of his 20 sampled line ends chiming with another line end in the same stanza, against 0 of 20 for the unrhymed control.

PR-PRIMARY, in one sentence: in Eastwick's 1852 verse rendering of these ten tales, the word at the end of a line is more often a word with no counterpart in the Persian than a word from the middle of the same line is.

2. What this is not

• It is not a claim that Eastwick's rendering is bad. A verse translator who has undertaken to rhyme must put something in the rhyme position, and supplying it is the craft. The measurement is of where the constraint is paid for.
• It is not causal. The design does not randomise rhyme; it compares positions and hands. Both critic seats pressed this and both are right: a positive PR1 licenses an association between line-final position and supplied words in this hand, and nothing about rhyme causing it, except so far as PR2's control arm carries that weight.
• It is not a test of the lead against a published hand. The lead's arm is here because it is a rendering of the same 39 bayts under a policy that refuses the purchase — and its END−MID gap is partly entailed by that policy, which is why the critic struck it from the primaries (§10, A3).

3. Materials

Rhyme coding, mechanical, from tools/rhyme_pairs.py and its CMU dictionary, computed on the sampled lines before any judging call: a line end is coded by its best relation to any other line end in the same block.

UNR is verifiably unrhymed and EAS is verifiably rhymed; the arm contrast is what it says it is. LEAD's four are D22's and D23's — the rhymes that fell out of the plain sentence.

4. Arms — positions, not texts

Every item is one English word set against the Persian bayt(s) its block renders, and nothing else. The seat cannot see the arm, the hand, the line, or the hypothesis.

Selection is mechanical and seeded (build_items.py, seed 20260828). A line is eligible only if its final token and at least one interior token pass the same content filter, and an eligible line contributes one END and one MID item — so the two arms are drawn from an identical line set and cannot differ in lexical class by construction. 140 items, one shuffled order.

5. Procedure

Three seats, per config/models.md — P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, QR qwen/qwen3.7-max (P3 is the UNR translator and does not judge; P4 and P5 are out on notes (bps) and (bne)). One call per item per seat, no batching, temperature 0. Each seat is asked, for one Persian bayt and one English word:

Does the English word render something that is actually present in this Persian — a word, an image, an action, a thing named, a quality, a relation? … If you cannot read the Persian well enough to say, answer UNSURE — that is a real answer and it is better than a guess.

Item verdict = majority of the three seats; a three-way split or any tie is UNSURE.

Cap probe first, note (brt): the six longest prompts, three seats, 18 calls. Acceptance rule, registered (critic A7): all 18 bodies must parse and none may carry finish_reason == "length". If either fails, the cap is doubled and the probe re-run once; a second failure stops the run before the main block.

6. Registered primaries and predictions

Both two-sided, alpha 0.05, Holm across the two (critic A4). Significance by exact permutation over the arm labels within hand, 20,000 relabelings, seed 20260828, reported with the risk difference and its 95% interval. The test is associational, not causal (§2).

• PR1 — P(ADDED | END, EAS) − P(ADDED | MID, EAS) ≠ 0, paired within line. Predicted positive, on D24.
• PR2 — [P(ADDED|END,EAS) − P(ADDED|MID,EAS)] − [P(ADDED|END,UNR) − P(ADDED|MID,UNR)] ≠ 0. Predicted positive. UNR is English verse with line ends, line-final stress, and clause closure — everything EAS's line ends have except the rhyme — and it was told nothing about fidelity, so its gap is free to be anything. This is the primary that can fail informatively: if UNR shows the same gap, the effect is about line ends in English verse and not about rhyme, and D24's claim is much weaker than it reads.

Reported, not registered as a primary: the LEAD gap, and the overall ADDED rate by hand. The LEAD gap is declared partly entailed — V2 forbade the lead to supply an unanchored word anywhere, so a small gap there is policy, not evidence (critic A3). It is reported because it is the wire back to the translation limb, and it is interpreted only as a check that the policy was kept.

7. Gates — what withholds the primaries

• G1 (can the seats read the Persian at all). CTRL-ANCH items coded ANCHORED at < 0.75 (fewer than 9 of 12) → both primaries withheld. This is the real calibration: known answers, blinded among the study items.
• G2 (ceiling diagnostic). Pooled DECOY items coded ADDED at < 0.55 (fewer than 14 of 24… 13 of 24, the rule being strictly below 0.55) → both primaries withheld. The bar is low and the reason is written: decoys are drawn from the same book and the same ten tales, so a decoy word is "not in this Persian" only in the sense that no hand used it to render this block, and shared vocabulary will anchor some honestly. G2 is a ceiling on the instrument, not proof of it; G1 is the proof.
• G3 (three-way UNSURE). UNSURE above 0.25 of non-control item verdicts → withheld.
• G4 (saturation). Non-control ADDED rate pooled at ≥ 0.95 or ≤ 0.05 → withheld.
• G5 (floor). Any of the six hand × arm cells below 10 complete items, or fewer than 132 of 140 items complete → both primaries withheld.

A gate that fires is the result. No primary is reported past a fired gate, and no number from a withheld primary may be cited anywhere in this repository.

8. Budget

Ceiling raised from $1.60 to $2.40, and the reason is the critic: two BLOCKING findings required a third translation arm (UNR, $0.150823200 spent) and forty more judging items. Spent so far: critic $0.093254750 (C2 re-dispatched once at a larger cap after its first body truncated), UNR $0.150823200. Remaining plan: 18 probe calls, then 140 × 3 = 420 judging calls. Worst case built from the cap actually set and not from an assumed output length, note (abc). Headroom on the day at design time: $3.769373050.

9. Limitations, written before the run

1. Not causal, and one published hand. PR1 is a statement about Eastwick 1852's rendering of باب دوم حکایات ۱۱–۲۰ and nothing wider. Arnold's exclusion (§3) is what costs the generalisation; span C can pay it back from page images.
2. Alignment is at block level, not hemistich level. A word absent from its own hemistich but present elsewhere in the same block codes ANCHORED. This makes ADDED a stronger claim and depresses all arms equally.
3. The UNR control is a model, not a person. It is the right control for line-end convention without rhyme, and it is not evidence about what human unrhymed verse translators do.
4. The design is not disinterested. The lead produced two of the three arms and the hypothesis.
5. Single-witness copy-text, inherited and declared.
6. G1 is 12 items. A seat that reads Persian badly but is right about camel and stone will pass it. It is a floor, not a certificate.

10. The pre-run critic, and what was done with it

Two seats, both adversarial, both given the frozen v1 design, the item builder verbatim, eight items as the seats would see them, and D24. C1 openai/gpt-5.6-terra returned NEEDS-REDESIGN, 7 findings, 2 BLOCKING. C2 google/gemini-3.6-flash returned NEEDS-REDESIGN, 5 findings, 2 BLOCKING. Raw bodies: raw/critic.json. Eleven of the twelve findings are accepted; one is overruled.