Repository path: workshop/experiments/E-20260829-radif-hands/design.md · rendered 2026-09-09

Can a published hand supply a counter-instance to the impossibilities §7.36.1 and §7.40 state?

This is v2. v1 was frozen at commit 31e6026b, put to two independent critic seats on two labs, and both returned NEEDS-REDESIGN — 25 findings, 9 BLOCKING; 20 accepted, 3 accepted in part, 2 remedies overruled in writing (critic-response.md). No call beyond the critic pass was dispatched under v1. The heading above is itself an amendment: v1 asked whether the predictor sorts carriages from failures, and both seats established that nine items cannot license a sorting claim. What nine items can license is a counter-instance to an impossibility, and that is what is registered.

ARM-radif-hands step 1 (T5). Tier D is NOT PASSED; nothing here is a quality judgment and no sentence of the result page will say that a reader hears, prefers or wants anything.

1. What is under test, and when it was written

framework/v0.2 §7.36.1 (published 2026-08-26) and §7.40 (published 2026-08-28) — both before this session — say when a Persian radif can be carried into English:

§7.36.1 — "A Persian radif can be any word, because Persian puts the verb last… An English clause ends on an object, a complement or an adverbial, so an English radif can only be something an English clause can end on." Its table calls a transitive preterite and an oblique pronoun unreachable.

§7.40 — "A radif standing in a genitive to its bearer cannot be carried, whatever its part of speech."

And one prediction is registered-but-unpublished: RS-20260828 §5.3 (S230) conjectures a third obstacle — that English cannot end a clause on a bare copula, on an imperative whose complement precedes it, or on a case particle — deliberately not written into the framework.

Both published cells rest on six Sa'di ghazals rendered by one hand, the lead's, which is also the hand that wrote the rule. That is the defect this experiment exists to remove.

2. Materials

• Source. Ganjoor's Hafez, Qazvini–Ghani text, all 495 ghazals, fetched 2026-08-29.
• Hand A, declared — Walter Leaf, Versions from Hafiz: An Essay in Persian Metre (1898), 28 odes. Printed declaration, Introduction pp. 8–9: "Several instances of this have been reproduced in the following translations."
• Hand B, undeclared — Gertrude Bell, Poems from the Divan of Hafiz (1897). No promise about form; her Introduction names "the beat of the refrain" as a beauty of the original.
• Hand C, complete — John Payne, The Poems of Shemseddin Mohammed Hafiz of Shiraz, vol. 1 (1901), odes I–CC, monorhymed English verse. Payne's ode numbering is Brockhaus's, which is the numbering Leaf cites, so the two hands are matchable to each other and to the source.
• Lead limb — T-hafez-ghazals-radif-R55-v1, two ghazals rendered whole, frozen at commit 7db1fedc before this design was written, contamination measured clean against all three hands.

2a. The nine items, named before a classification call was bought

Census and matching cost nothing and were run first, so the item list is frozen here rather than left to be discovered (critic C2-5). Leaf ode → Ganjoor slug → Brockhaus number → radif → bayts:

Five of Leaf's 28 are excluded and named under G1: his I and II, which he prints as (MS.) and which are not in the Qazvini–Ghani canon, and his XIV, XVI and XVII, whose opening couplets match nothing in it above ratio 0.75. Of the 23 that matched, 14 have no radif and are the P4 control set. Bell is not matched item by item and contributes only the undeclared-hand rate.

3. Procedure, each step mechanical

1. Census. census.py enumerates every ghazal's radif by a frozen rule: rhyming positions are hemistich 1 plus every second hemistich; the radif is the longest common trailing word sequence over them, minus one word reserved for the qāfiya. Bound morphemes are not words and therefore never count as a radif.
2. Match. Leaf's 28 odes are matched to the census by consonant-skeleton similarity of the transliterated مطلع he prints at each ode's head, requiring ratio ≥ 0.75.
3. Code. carriage.py codes each hand's English: carried iff some trailing word sequence closes at least B − 1 of the poem's printed lines, where B is the bayt count of the Persian. The denominator is the source's bayt count and not the scan's line count, because OCR inflates lines and cannot inflate a matching tail. Every coded poem's line-ends are printed in the result page so the coding is checkable against the page.
4. Correspondence (critic A1). A cell the coder returns carried is kept only if the English tail renders the Persian radif. The Persian radif, a literal gloss and the English tail are printed side by side for every such cell in the result, and a tail that is not a rendering is recoded not carried. This is a judgment made in public on nine printed pairs, and is labelled as one.
5. Classify — bought, and bought blind. Each radif's grammar is classified by three independent seats shown only the Persian radif inside the two hemistichs of its own مطلع, on a closed checklist with one bit per predicate the published cells turn on (critic A3): part of speech · finite · preterite · transitive · bare copula · case particle · imperative with fronted complement · in ezāfe/genitive to the preceding word · UNSURE. The seats are shown no English, no rule and no prediction. This is the whole reason money is spent: it takes the classification out of the hand that wrote the rule.
6. Apply. §7.36.1, §7.40 and the S230 clause are applied to the seat-majority checklist by lookup.

4. Predictions, registered

• P1 — the published cells. §7.36.1 and §7.40 state their two cells as impossibilities. A single published counter-instance refutes an impossibility. Registered before coding was read: if any published hand carries a radif that the seat-majority classification puts in a cell §7.36.1 or §7.40 calls unreachable, that cell is downgraded from impossible to costly in framework/v0.2, and the cost is described from the hand's own page.
• P2 — declared against undeclared, DESCRIPTIVE ONLY (critic A8). Leaf's carriage rate and Bell's are reported with their own denominators and not compared as a test: the two hands translated different poems, so declaration is confounded with material. Registered only so that the confound is stated before the numbers are seen, not after.
• P3 — the S230 conjecture's case-particle clause. No hand carries a radif that is a case particle. Registered as a conjunction over the hands available on that item; one carriage refutes it.
• P4 — the false-positive channel. On ghazals the census finds to have no radif, a hand's English should not produce one. Registered as a control, not a prediction of interest: any hand-poem cell where a no-radif ghazal codes as carried is reported and explained rather than dropped.

5. Failure criteria and gates

• G1 matching. An ode whose best skeleton match is below ratio 0.75 is excluded and named. If more than 8 of Leaf's 28 are excluded, the matched corpus is not representative of his book and P2 is withheld.
• G2 inter-seat agreement (critic A2; lead-agreement is struck as a gate, because it made refutation conditional on agreeing with the party who wrote the rule). If fewer than 2 of 3 seats agree on a checklist bit, that bit is UNSURE for that item; if any item the analysis turns on has an UNSURE on the bit that decides its cell, that item is NO-TEST. Lead-versus-seat disagreements are printed as a diagnostic and withhold nothing.
• G3 denominator. If fewer than 6 radif-bearing ghazals survive matching, P1 is withheld for want of items.
• G4 coder control, on two named held-out sets (critic A10). Bell's 32 poems must code not carried; Leaf's ode XXI and Payne's ode 52 must code carried. Line-ends are printed for all three. A coder that fires on a hand who never attempts the device is measuring monorhyme.
• G6 coverage (critic A5). A cell with no matched item is reported NO-TEST, never as "not refuted". On the frozen list: case particle 1, transitive finite verb 4, oblique pronoun in genitive 1, copula 2, adverbial particle 1.
• G5 power, declared in advance rather than discovered. Nine items across three hands is not enough for a significance test of association, and none is claimed. P1 and P3 are existence and conjunction claims, which nine items can settle; the rate at which each hand carries is reported descriptively with its exact denominator and no P-value is computed for it. Step 2 of the arm exists to power the association, by matching Payne's two hundred.

6. What is bought

Declared ceiling: $1.20. Note (brr): P2 truncates a long critic prompt at 4000 and is not used for stage C. Note (brt): the cap is verified for the shape it is probed on; stage G's prompts are short and uniform and the first three returns are checked for truncation before the rest are dispatched.

7. What this design cannot do

It cannot say whether a reader notices a carried radif; nothing here is put to a reader and Tier D is not passed. It cannot separate register from capacity: if a hand who writes archaic inverted English carries what a hand who writes plainer English drops, this design sees the difference and cannot say the plainer hand could not have done it. And the lead's own limb is a craft report from a hand that knew the rule; it is evidence about what attempting the shape costs, not evidence about whether the rule is true.