Repository path: workshop/experiments/E-20260902-rhyme-slot/amendment-v2.md · rendered 2026-09-09

E-20260902-rhyme-slot amendment v2 — the design after two NEEDS-REDESIGN verdicts

Two independent pre-run critics, disjoint labs, neither a seat the run uses: C1 openai/gpt-5.6-terra and C2 x-ai/grok-4.5, $0.104133000. Both returned NEEDS-REDESIGN; 19 findings, 5 BLOCKING. The finding-by-finding response is critic-response.md. This page is the amended design and it, not v1, is what runs. No data had been collected when the critics were dispatched and none has been collected now.

What the critics broke, in one paragraph

v1's predictor was COND, a property of the Persian: the qāfiya positions against the free hemistich-ends. Both critics, independently, said the same thing about it — a Hafez qāfiya is drawn from a stock lexicon and in this corpus fuses with the copula, so CONSTRAINED is a Persian lexical category and not an English rhyme obligation, and v1's predicted difference could arise with no English rhyme involved at all (C1-1, C2-2, both BLOCKING). Worse, C2-3: on a rhyming English line, AT-END is true exactly when the rhyme word carries the sense — which is step 1's R = 0.0384 measured again, so v1's primary was close to a re-measurement dressed as a new test. And C1-2 / C2-1: because stage L may find a gloss anywhere in an unaligned rendering, COND was not even attached to the line whose rhyme obligation was supposed to be doing the work.

The amended design

The predictor is now a property of the English line the sense actually landed in.

For each rendering, computed mechanically in the analyser and never shown to a seat:

• END-BLOCK — the maximal common word suffix of the rendering's rhyming lines (the English radif; empty for hands who do not use one).
• rhyme signature — the modal CMUdict rime over the content word immediately preceding the END-BLOCK on each rhyming line, under S233's non-rhotic licence.
• A line is RHYMING iff its own pre-END-BLOCK content word carries that rime. Every other line is OPEN: nothing about its ending is spoken for.
• d = word tokens after the located carrier, to the line's end. dq = tokens from the carrier to the qāfiya slot (the token immediately before the END-BLOCK); dq = d where there is no END-BLOCK.

SRC — whether the sense came from a CONSTRAINED or a FREE Persian hemistich — is retained, but it is now a stratifier, not the predictor.

The primary, which is a difference in differences

P1 (primary): P(AT-END | D) − P(AT-END | B) > 0.

Cells B and D hold only senses from ordinary Persian hemistich-final words — no qāfiya, no copula fusion, no stock rhyme lexicon. The Persian word class is held constant and the only thing that varies is whether the English line the sense landed in had to chime. That is the comparison C1-1 and C2-2 said v1 did not have, and it is available from exactly the same calls.

P1a: the same difference computed as a difference-in-differences, [P(A) − P(C)] − [P(B) − P(D)], reported with its interval. If the qāfiya's word class is doing the work rather than the rhyme, the CONSTRAINED row falls further than the FREE row and this is where it shows.

P2 — the pre-rhyme slot, which is the arm's actual question

Step 1 §3(c) claimed, on a single poem, that a displaced rhyme sense sits one word to the left of the rhyme. Registered here as a distribution, not a binary:

Among carriers in a RHYMING line with dq ≥ 1, the share at exactly dq = 1, against two comparators: (i) the same share among carriers in OPEN lines, and (ii) a mechanical null in which the carrier is placed uniformly over that line's non-qāfiya word positions. P2 holds if the observed share exceeds both.

This is not a re-measurement of anything: step 1 measured whether the rhyme word carries the sense, and this measures where the sense sits when it does not.

Inference: observational, with clustered uncertainty

C1-6 is accepted in full. SRC and line-rhyme status are not randomly assigned, so a label permutation is not a valid randomisation null and v1's was withdrawn. The run reports cluster bootstrap intervals, resampling whole renderings with replacement, 20,000 draws, and states in the result that every contrast is observational. No causal language is used for any published hand. C1-7's clustering objection is answered by the same bootstrap.

Registered gates, revised

Failure criteria. F1 the primary's bootstrap interval includes 0. F2 the sign differs between Leaf and Payne — the claim is then limited to the hand where it holds. F3 the sign reverses on the END-BLOCK (radif) stratum. AT-END governs; mean d is secondary, and if the two conflict the result says so and claims neither (C1-11).

Additional registered quantities, all from calls already being made

• ALLITEM (C1-4): the primary recomputed with NONE scored as not at end — a conservative bound immune to location selection. Reported beside the primary.
• MIGRATION (C2-6, C1-2): the share of located carriers standing outside the two English lines of their own bayt, for Payne, whose renderings are bayt-for-bayt by construction.
• RADIF stratum (C1-10): the 14 English-radif renderings are restored, not excluded. Excluding them removed a genuine realisation of the rhyme constraint on outcome-adjacent grounds. The sample is now 56 renderings — Leaf 28, Payne 28, 54 distinct ghazals, 538 CONSTRAINED and 426 FREE sense positions — and the OCR screen is the only screen that drops anything (it drops none of these).
• TERMPOS (C2-4, C1-3): the share of line-final tokens that are function words, by line type. A hand who ends OPEN lines on function words inflates P1 for reasons of English rhythm.
• X-POS, a new control (C2-8): 4 items pairing one poem's glosses with another poem's rendering, scored for the same position distribution. Step 1 §6's in-field-mismatch prior is about presence and does not cover position, so v1 was wrong to treat it as covering this; the control is bought.

What the critics asked for and this design does not do

C1's single most important change is a randomised paired translation experiment — independent translators rendering the same ghazals under randomly assigned rhyme / no-rhyme instructions. The project cannot reach independent translators, and the one hand it can commission is the lead, who conceived the outcome variable before writing either arm. T-hafez-boro-R59-MONO-v1 and -PLAIN-v1 are that experiment at n = 1 with a compromised hand, and they are reported as description, in the same posture step 1 §5 took, with Payne's ode 39 on the same ghazal beside them. Saying so is the honest response; calling the pair a test would not be.

Budget

Ceiling $2.20 stands. Calls: 54 gloss + 58 locate + 12 POSBIAS + 4 X-POS = 128, caps set by a per-seat, per-shape probe first (note (bsf)). Critics already spent $0.104133000.