Repository path: workshop/experiments/E-20260824b-hariri-hands/design.md · rendered 2026-09-09

E-20260824b — does the English tradition's answer to rhymed prose hold outside Sa'di?

ARM-persian-hands step 2. Frozen before any published hand was opened at Assembly I beyond the four lines declared on ../../translations/maqamat-sanaa/loci-frozen.md §7, and before the grading script was written.

1. The question

RS-20260822c established, on four English Gulistans spanning 1806–1899 at 32 Persian rhymed-prose loci, 128 cells: not one full rhyme inside the prose, twelve near-misses that are mostly English inflection, and the genre escape — lifting the passage into English verse — taken at 4 of 128. framework/v0.2 §7.27 tells a practitioner that dropping a saj' is what the tradition does.

That is four hands who read each other, on one author, in one language. Eastwick footnotes Gladwin and Ross by name on those very pages. The finding's whole force is that it is about the tradition, and its evidence is Sa'di.

Does the record hold on a second author, in a second language, whose prose is rhymed from end to end and whose English hands are a century older?

2. Materials

Source. «المقامة الصنعانية», the first of al-Ḥarīrī's fifty Maqāmāt (c. 1110) — the canonical monument of Arabic saj'. 590 tokens; 139 prose cola of which 137 stand inside a rhyme figure; 9 verse lines in two mono-rhymed pieces. Copy-text and its single-witness declaration: ../../translations/maqamat-sanaa/collation.md.

Loci, frozen from the Arabic before any English: ../../translations/maqamat-sanaa/loci-frozen.md — 67 PROSE-SAJ loci, primary pool 65 (two total puns excluded by rule, S67 added post-freeze and reported both ways), 9 VERSE lines.

Hands. All public domain, all fetched as page scans and OCR from the Internet Archive.

Steingass 1898 translates Assemblies 27–50 and cannot supply this locus set. It is named here so that its absence is a fact of the materials and not an omission.

Three hands, 1767–1867 — a hundred years, against the Gulistan's four hands over ninety-three. The two censuses are close enough in size to be read against each other and that is deliberate.

3. Codes

Imported verbatim from ../../translations/maqamat-sanaa/loci-frozen.md §6: FORM (VERSE / PROSE / OMITTED, read off the page image, not the OCR), CHIME (graded by tools/rhyme_pairs.py between the English words rendering the Arabic bearers), COLA-KEPT, and the derived ANSWERED and GENRE-CHANGED.

The per-bearer alignment — which English word renders which Arabic bearer — is committed to raw/census.json before analysis.py is written or run. This is the round-2 and round-3 BLOCKING finding of the Gulistan critic and it is not re-litigated here.

STRICT and NEAR are reported separately and never summed in a headline. The reason is recorded on the translation limb and holds for the hands too: rhyme_pairs.py returns NEAR for any two words sharing a suffix — running / singing grades NEAR by rule N1 — so a NEAR total is an upper bound on anything a reader would hear.

3a. Amendments after the round-1 critique (2026-08-24, before any hand was opened)

Fourteen findings, five BLOCKING; the response and what was accepted is critic-response.md. The amendments that change what gets measured:

• CHIME-WIDE, a second registered measure (B1). Every adjacent pair of clause-final words inside the English rendering of the locus, graded by the same tool — not only the two words aligned to the Arabic bearers. It is reported beside the bearer measure, and the headline is stated in the bearer measure's terms and never beyond them. It also answers S3: CHIME-WIDE does not use the alignment, so if it too returns zero STRICT, the alignment is not what produced the zero.
• P1b is evaluated on STRICT + NEAR-N2 + NEAR-N3 (P2-S1). NEAR-N1 — the stress-or-suffix match that faithful parallel translation produces for free — is reported separately and never folded into a headline.
• The inferential unit is the hand (B3). Hand-level rates are the primary; pooled counts are descriptive. The conclusion is bounded to these three hands read beside the Gulistan's four: seven hands, two authors, two languages.
• OMITTED and UNSCORABLE are excluded from every denominator (S1, S6) and counted in the open. No hand is dropped for OCR; if any hand exceeds ten UNSCORABLE, no pooled figure is reported at all.
• COLA-KEPT, operationally (S4). The English rendering of the locus contains two or more members in the Arabic's order, each rendering one Arabic colon, neither absorbed into the other and neither reduced to a modifier of the other. Two failure shapes: collapse (the two cola become one clause — Lane's «in ancient times, a King» for a doublet) and deletion (one colon gone). If it cannot be coded reliably it is reported inconclusive, never withdrawn.
• SOURCE-DIVERGENT (B4). A locus where a hand's English implies a different Arabic bearer than the copy-text prints is flagged, and the primary is reported with and without.
• A rhyme-detection positive control (S5). Ten constructed English locus-renderings carrying a full rhyme at the bearer positions go through the identical alignment-and-grading pipeline. If fewer than 9 of 10 grade STRICT, a zero in the real data is not interpreted.
• F1 is demoted (B5, S2). The CHANCE figure is a descriptive background rate, not a control and not a gate. The study therefore has no true control and says so; note (bqr) fires a third time.
• S1 is descriptive (M1); P3 reports printed as verse and printed as verse and rhymed separately (M-P2); stage L gains four subtle plants and eight faithful negative controls (S7).

4. Predictions, registered

P1 — the record generalises (the primary). Over the 65 primary loci × 3 hands = 195 cells:

• P1a: the pooled STRICT count is 0.
• P1b: the pooled ANSWERED rate (STRICT + NEAR) is below 0.15.

P1a is the Gulistan result's strong form, which was not registered there and is the one worth having. P1a fails on a single full rhyme anywhere in 195 cells.

P2 — the genre escape is not taken. Pooled GENRE-CHANGED below 0.10, and no single hand above 0.30. (Gulistan: 0.031 pooled, maximum 0.125.)

P3 — verse is answered where prose is not. All 9 verse lines are scored for every hand, whatever form the hand prints (B2). For each hand: (verse ANSWERED) − (prose ANSWERED) ≥ 0.30. Printed as verse and printed as verse and rhymed are reported separately. If no hand prints verse at all, P3 is reported as no cases — a null, never as satisfied. (Gulistan: +0.811 and +0.921 for the two hands that print verse; the two that print none scored 0 and 0.)

P4 — the doublet survives where the sound does not. Pooled COLA-KEPT ≥ 0.70. RS-20260814 §1 found that Lane and Burton lose the saj' of the Nights' frame tale without losing a single colon; that was two hands, eight loci, and coded by the lead's reading. This is the first time it is registered in advance, and it is this study's one positive prediction. It fails below 0.70.

S1 — descriptive, no claim attached (amended, M1). The STEM and AFFIX answer rates are printed. No decision rule, because none was registered that could bear one. (Gulistan: 0.133 against 0.059.)

F1 — a background rate, and not a control (amended, B5/S2). Each hand's locus ANSWERED rate is printed beside its own CHANCE rate (loci-frozen.md §4: every adjacent pair of clause-final words across the whole of that hand's rendering). It is descriptive and it gates nothing. As first written it disqualified exactly the hands that confirm P1 — a hand that does not rhyme cannot exceed chance — which the P2 seat caught and this design had not. So the study has no true control, and says so: al-Ḥarīrī's unrhymed prose in this maqāma is six words long. Note (bqr) fires a third time and is named in the result.

5. The lead's column, and what the money is for

T-maqamat-sanaa-R43-v1 renders the same maqāma under R43, which requires an attempt at a prose chime at every locus and forbids the escape into verse. It is a labelled subject and is not part of the primary. Its verdicts are the hand's own: TAKEN 33, AVAILABLE-REFUSED 1, UNAVAILABLE 31, of which only 4 are STRICT.

Stage L — the outside check, and the only thing in this design that costs money. R43 rule 2 says a chime bought with a word the Arabic has not got is a refusal. The hand cannot adjudicate that about itself. So each of the 33 TAKEN renderings goes to two seats (P1, P2) reading the Arabic colon and the English, asked one question: does the English state anything the Arabic does not? A locus both seats mark as adding is removed from the hand's TAKEN count.

Controls, both directions (amended, S7 / P2-S2). Eight planted items — four with a conspicuous three-word addition, four subtle (an explicitation, a strengthened evaluation, a connective doing semantic work, a licensed-looking gloss) — and eight unedited faithful items as negative controls, all shuffled among the 33. If the two seats do not catch at least 6 of 8 plants, stage L is void. If they mark more than 2 of the 8 faithful items as adding (a false-positive rate above 0.25), the gate is reported as unreliable and the hand's column stands ungated with that said. Stage L is 49 items × 2 seats = 98 calls.

Blind: the seats are not told which loci are chimes, are not told the prediction, and see the items in a shuffled order.

6. Procedure

1. Extract each hand's Assembly I text from its OCR; transcribe every bearer word by eye against the page scan, because 1767 long-s and 1850/1867 Indic-scan OCR are unreliable at exactly the word the study depends on.
2. Build raw/census.json: per hand, per locus — the located English, the two (or more) bearer words, FORM, COLA-KEPT, and the coder's own reading of whether it chimes. Commit.
3. Run analysis.py, which calls tools/rhyme_pairs.py on the committed alignment and computes every rate in §4. It re-asserts that each quoted English string occurs in the stored comparator file and exits non-zero if any does not.
4. Stage L, per §5.
5. verify.py recomputes every number that reaches the result page from raw/.

7. Failure criteria, declared

• If fewer than two hands render this maqāma in full, the census is under-powered; the result says so and ARM-persian-hands closes retired with the reason, exactly as the arm page provided for.
• ~~If a hand's OCR cannot be made to yield reliable bearer words at more than 10 of the 65 loci, that hand is dropped~~ STRUCK by S6, which is right: the hand most likely to refute the headline is the 1767 one whose type is hardest to read. Every bearer is transcribed by eye against the page scan; an unrecoverable locus is UNSCORABLE and excluded from denominators; if any hand exceeds ten UNSCORABLE, no pooled figure is reported at all.
• If P4's COLA-KEPT cannot be coded reliably at more than 10 loci, P4 is reported inconclusive — amended by S4; the original said withdrawn, which is an escape hatch and not a failure criterion.
• If the rhyme-detection positive control (§3a) scores below 9 of 10, no zero in this study is interpreted as evidence of anything.
• Stage L's control failing voids stage L, not the primary.

8. What would refute the headline

One STRICT chime, in any of the 195 cells, refutes P1a and is the most interesting thing this study could find. Chappelow 1767 is the hand to watch: he is a century before the others, he prints the Arabic beside his English, and his book is a commentary on the rhetoric. If any hand tried the figure, it is him. The study is designed so that his succeeding is a clean result and not a nuisance.

9. Budget

Pre-flight, against the UTC-day ledger in config/budget.md: stage L is 33 + 8 = 41 items × 2 seats = 82 calls, plus the pre-run critic. max_tokens 1200 / 400, built from the cap the request permits and not from an assumed output length (note (abc)).

Declared ceiling: $0.90, raised from $0.60 when the critique took stage L from 82 calls to 98 and added a round of critique that was itself unbudgeted. Worst case at P1 + P2 list prices with the caps above is well under that. If the probe puts the run above $0.90, stage L drops to one seat (P1) and, under that fallback, a locus is removed if that one seat marks it as adding — the stricter gate, stated here because P2-S3 pointed out that the fallback had no removal rule at all — and the result says which gate was used.