Repository path: workshop/experiments/E-20260830-leaf-contract/design.md · rendered 2026-09-09

E-20260830-leaf-contract — two English Hafizes who print a formal contract, audited against their own pages

ARM-leaf-contract step 1 (T4), 2026-08-30 (S233). Frozen before any coding call and before any English of the translated poem was opened. Tier D is NOT PASSED; nothing here is a judgment of quality.

1. The question, and why the shelf needs it

A-hariri-hands is the only anchor on the shelf whose central evidence is translators writing down, in their own prefaces, what formal feature of their author they will and will not reproduce, and why. It is Arabic. framework/v0.2 §7.36.1, §7.40, §7.41 and §7.42 are all recommendations about carrying Persian verse form into English, and every one of them now stands on two published hands — Leaf 1898 and Payne 1901 — that the project has read twice (S231, S232) and never given an anchor page. D-20260724-04 requires a release to declare what each recommendation stands on; these point at a result page, which is a record of one coding run, not at an evidence point.

So the deliverable that needs this: §7.41 and §7.42 are traceable to an anchor or they are not traceable. That is the T4 growth condition (wiki/base/anchors/README.md) met in the form the shelf demands.

The question the anchor is built to answer is not did these two men translate well. It is: each printed a contract about form before his first line; which clauses does his own page keep?

2. The two contracts, quoted from the page

Leaf 1898 prints his in an Introduction of twenty-one pages, itemised and hedged. Five clauses are stated plainly enough to be checked:

Payne 1901 prints his in one unqualified sentence, on the title page of all three volumes:

NOW FIRST COMPLETELY DONE INTO ENGLISH VERSE FROM THE PERSIAN, IN ACCORDANCE WITH THE ORIGINAL FORMS

P-FORMS is that clause. It is sweeping where Leaf's are narrow, and it is the only formal statement vol. 1 makes.

This asymmetry is the anchor's point and it is not a finding of this run: it is what the two books print. Whether the hedged contract or the sweeping one is better kept is what is measured.

3. Materials

Storage: both English texts are public domain and are stored whole in materials/. The Persian author died c. 1390.

4. The five measures, each mechanical, each with its rule fixed here

Rhyming positions of a ghazal, in both languages: hemistich 1 of bayt 1, hemistich 2 of bayt 1, and hemistich 2 of every later bayt. A poem of B bayts has B + 1 rhyming positions.

C1 — L-RHYME / P-FORMS: is the monorhyme carried? For each ode, take the final word of each rhyming position; the ode's rhyme is the final word of position 1. Score the fraction of the remaining positions whose final word rhymes with it. Two codings, both frozen here:

• STRICT — tools/rhyme_pairs.py's existing rule, unchanged and not tuned for this run.
• LEAF-LICENCE — STRICT, plus the three licences Leaf prints and defends: (i) pre-consonantal and final r is silent, so dawn : morn is a rhyme (his footnote, p. 12); (ii) the identical word may stand as its own rhyme; (iii) the rhyme may fall on an unaccented syllable.

C2 — L-METRE: is one measure held? For each ode, count syllables in every printed verse line by a frozen heuristic (vowel-group counting with a fixed silent-e and -le rule, stated in syllables.py and not adjusted after any count is seen). Report, per ode, the modal syllable count and the fraction of lines at it, and the fraction within ±1. Line reconstruction is the threat here — OCR wraps long printed lines — so the joiner rule is fixed in advance: a fragment is joined to the previous line unless it begins with a capital letter and the previous line ends in terminal punctuation or a rhyme-word. A random sample of 20 reconstructed lines is checked against the page scans and the disagreement rate printed; above 0.10 the measure is withheld.

C3 — bayt-for-bayt. English rhyming positions against the matched Persian bayt count. A hand keeping the form should give B + 1.

C4 — L-REPEAT, and it is Leaf's own comparative claim. For each matched ode, the repetition rate = 1 − (distinct rhyme-bearing words ÷ rhyming positions), computed the same way for Leaf's English and for Hafez's Persian qāfiya list from the census. L-REPEAT predicts Leaf's rate below Hafez's, ode by ode. Reported as the paired difference over the 23 matched odes, with the exact sign test.

C5 — L-CHOICE. From Leaf's printed table: Spearman correlation between a metre's frequency in Hafez and the number of Leaf odes in it; and the count of the 24 metres his 28 odes represent. L-CHOICE predicts a near-zero or negative correlation and broad coverage; a translator picking by taste would track frequency, because the commonest metres carry the most poems.

5. Predictions, registered

• H1 — P-FORMS, the sweeping clause, is kept at a higher rate than L-METRE, the narrow one. Rationale: Payne's contract names no measure and can be satisfied by any regular English verse; Leaf's names the Persian measure and can fail. This is the prediction the run is most likely to embarrass, because it says the vaguer promise wins.
• H2 — C1 is at or near 1.000 for both hands. Monorhyme is the clause a Victorian verse translator can always keep, at the cost of vocabulary; both books are set as monorhymed odes.
• H3 — L-REPEAT fails. Hafez rhymes in a language with, in Leaf's own words, "eight or ten rhymes to most words"; English does not, so the hand under the tighter constraint should repeat more, not less, and Leaf's belief about himself should be wrong.
• H4 — L-CHOICE holds, at Spearman ρ ≤ 0.2 and ≥ 15 of 24 metres represented.

Failure criteria, registered. C2 is withheld entirely if the line-reconstruction disagreement rate exceeds 0.10. C4 is withheld if fewer than 18 of the 23 matched odes yield a complete rhyme list on both sides. C1 STRICT/LEAF-LICENCE disagreements go to §6 and no C1 figure is reported until they are adjudicated.

6. The one thing bought, and why it is bought

The author of this audit is the party claiming the clauses fail. C1's load-bearing judgment — do these two English words rhyme — is exactly where that matters, because Leaf's own printed licences widen the category and the widening is in his favour. So every rhyme pair on which STRICT and LEAF-LICENCE disagree, plus every pair STRICT calls a non-rhyme, goes to three blind seats (P1, P2, P3 per config/models.md), shown the two words alone — no poem, no hand, no ode, no rule, no prediction, and no indication that anything turns on the answer. Each seat returns two bits: do these rhyme in ordinary present-day English and would a nineteenth-century English poet have rhymed them. A pair is coded a rhyme at a 2-of-3 majority on the second bit, which is the historically correct standard for both books.

Deduplicated pairs only. Dispatch order is randomised against a seed printed in the result, and the two words are presented in randomised order within each pair, so no position artefact can carry a verdict (note (bqz) shape).

Gate G1. If inter-seat agreement on the second bit falls below 0.80, C1's adjudicated figures are withheld and only STRICT is reported, labelled as one party's coding.

7. Ceiling

$0.60. Worst case: ≤ 90 disputed pairs × 3 seats = 270 calls at max_tokens 400. Note (bsf) binds — the cap is probed on three calls of the task shape and re-probed at the tenth and the fiftieth item, not once at the start. Note (brx) binds — a truncated body carrying both bits is re-parsed, not re-bought. If the disputed set exceeds 90 pairs the run is cut to the pairs STRICT calls non-rhymes and the reduction is declared.

Everything else — the census reuse, the matching reuse, all extraction, all counting, all scansion, the translation limb and its contamination measurement — is the lead's own and is not ledgered (charter §3, A4).

8. The translation limb, and the wire

Written and frozen before this design's coding was run and before any English of the poem was opened. T-hafez-bekonad-R57-v1: Hafez غزل ۱۸۷ «دلا بسوز که سوز تو کارها بکند» rendered whole, twice, under R57 — the first instantiation of Leaf's printed contract by a hand other than Leaf's.

The poem is Leaf's own ode XV (Brockhaus 234), and it is the one item in RS-20260829-radif-hands §4's forbidden cell that no published hand carried: Leaf declined the radif here, and Payne's vol. 1 does not reach ode 234, so the record has no counter-instance.

The two arms differ in exactly one clause:

The wire, in one sentence: the audit measures which clauses of a printed formal contract two published hands actually kept, and the translation manipulates the one clause the audit cannot manipulate — releasing Leaf's metre on the very poem where Leaf let the radif go — so that what the audit can only observe as a correlation between a strict contract and a dropped device is put to a within-hand, within-poem test.

What the translation limb may and may not conclude. It is one hand on one poem. It can show that a clause is unkeepable together with another by exhibiting the failure, and it can name the mechanism. It cannot establish a rate, and no number from it enters C1–C5.