Repository path: workshop/experiments/E-20260827-declared-play/design.md · rendered 2026-09-09

E-20260827-declared-play — a translator's positive printed promise, and whether a reader gets what it promises

ARM-declared-function step 2. Frozen before any English of this span was read beyond the priming declared in loci-frozen.md, and before a word of the lead's own rendering existed.

1. The question, and why this is a different question from step 1

Step 1 (RS-20260825b-flippancy) tested two printed declarations by translators of al-Ḥarīrī and found both of them right about the reader. So did A-knatchbull-kalila on a third. All three are refusals:

• Chappelow 1767 — "nor shall I imitate the author in my translation … our English tongue will not admit of it."
• Preston 1850 — "Rhyming prose is extremely ungraceful in English, and introduces an air of flippancy."
• Knatchbull 1818 — "it would have been impossible to express the sententious brevity of the Arabic with strict fidelity."

A refusal is cheap to keep. A translator who prints that he will not do something will be found not doing it, and the finding that he was right about the cost is a finding about the cost, not about his self-knowledge. The arm's question — does a translator's account of what his choices do predict what a reader gets? — has never been put to a declaration that could be caught failing.

This session found one, in the last paragraph of a preface nobody in this project had opened:

Edward B. Eastwick, Haileybury College, October 1st, 1852, closing his translator's preface to the Gulistan: "I have also endeavoured to make the metre correspond in some degree to that of the Persian, and I have uniformly done my best to preserve the play upon words which occurs so often, and which is accounted such a beauty in the East."

It is positive, it names a specific device, it is quantified (uniformly), and its author's page is already on this project's shelf (A-gulistan-hands). Two other complete English hands of the same book, from 1806 and 1823, declare nothing about the play upon words — they are the control.

Q1, the primary. At the places where Sa'di puts a play upon words, does a blind reader find a play upon words in Eastwick's English more often than in the English of the two hands who promised nothing?

2. Materials

Copyright. Persian thirteenth-century; all three hands pre-1900. Logged in consulted.md.

3. The instrument, and what it deliberately is not

One call = one block × one arm × one seat. The seat is shown one English passage, told only that it comes from a book of moral tales, and asked to list every play on words it finds, quoting the exact words, or to answer NONE. It is never told the passage is a translation, never told which hand wrote it, never shown the Persian, never shown another arm, and never told that a rhyme, a pun or any particular figure is at issue.

It is not a comparison. Note (brs) — per-cell order consistency is a gate on any which reads better task on minimal pairs — cannot fire here, because no call ever contains two arms. Removing the comparison removes the order artefact entirely, at the cost of a weaker contrast; that trade is made deliberately and stated.

It is scored by location, not by string match — note (bru), whose defect was a rule that under-counted any target longer than the phrase a seat would naturally quote. For each locus × hand, the English span that renders the Persian members is marked by the lead and frozen before the run (materials/alignment.md). A seat's quotation is a HIT on locus J if it overlaps that frozen span; a quotation elsewhere in a تجنیس block is DISPLACED; any quotation in a control block is a FALSE POSITIVE.

What the lead is and is not deciding. The lead marks which English words render which Persian words — a translation-alignment judgment, the same operation A-knatchbull-kalila §3 used. The lead does not decide whether an English passage contains a play upon words. That is the seats', and it is the whole point: the arm exists because ARM-device-function closed when the labeller and the prediction could not be kept apart.

Block-level outcome. A block is DETECTED for a hand when a majority of the three seats returns a HIT on at least one of its loci.

3a. The verse-rhyme confound, and the two things done about it — amendment, 2026-08-27

Written before the run, before the pre-run critic returned, and before a word of any English of this span had been read or written. Recorded as an amendment rather than folded silently into §3, so that the order is inspectable.

A-gulistan-hands §2 records that Eastwick rhymes his verse and Gladwin and Ross do not — 38 of 42 bayt-end rhymes against 0 and 0. Ten of the twenty تجنیس blocks are bayts. If a seat asked for a play upon words quotes an end-rhyme, Eastwick wins the primary for a reason that has nothing to do with his promise about wordplay. Two measures, both registered here:

1. RHYME-ONLY is not a HIT. The seat prompt says in terms that rhyme alone and metre alone are not plays upon words. A quotation consisting only of two line-final words in a rhyming position is coded RHYME-ONLY and scored as no hit, whatever it overlaps.
2. The 10 prose blocks are a pre-registered confirmatory stratum, and on disagreement the prose figure governs. The verse stratum carries a confound the prose stratum does not; where the two answers differ, the smaller and cleaner one is the finding, and the other is reported beside it.

And the lead's own arm is held to it: R52 rule 1 forbids rhyming verse for the sake of rhyme, so LED cannot inflate its own ceiling the way the confound would inflate EAS.

4. Procedure, in the order it must run

1. Frozen already: collation.md, loci-frozen.md, the two JSON files, and this page.
2. Pre-run critic. Two seats, NEEDS-REDESIGN permitted, no data. Findings are answered in writing in critic-response.md before anything else runs.
3. The lead translates the span whole from the Persian alone, under R52, without opening any English of this span. The translator's log is written and frozen (charter §3, A4).
4. Only then are the three published hands extracted and block-aligned, and the alignment frozen. No parameter of this design may change at that point, and the LED arm's inclusion is fixed here, in advance, so that adding it later is not a decision made after seeing anything.
5. Contamination measured (tools/dependence_check.py, LED against each of EAS, GLA, ROS). This is a gate on the LED arm only — the Q1 primary is between three published hands and cannot be touched by it. If LED measures DEPENDENT? against EAS, the ceiling figure Q3 is reported with that measurement attached and is not used to support any claim.
6. Pilot, 6 calls: 2 blocks × 3 seats, to fix max_tokens on the actual task shape before any fan-out — note (brt), whose defect was a cap verified on a different shape.
7. Stage P: 32 blocks × 4 arms × 3 seats = 384 calls, dispatched at concurrency ≤ 6 (note (brf)).
8. Stage N, free, lead: the notes census. For each locus and each published hand, does the book's own apparatus — footnote, bracket or gloss — report the play without reproducing it? Coded REPRODUCED / REPORTED / NEITHER. This is a census of print, not a judgment.
9. Verification: an independent script recomputes every reported number from the raw JSON, plus mutation tests.

5. Seats

P1 openai/gpt-5.6-terra · P2 google/gemini-3.6-flash · third seat decided by the pilot: GL z-ai/glm-5.2 preferred (the prompts here are short, and GL is out only on LONG prompts), falling back to QR qwen/qwen3.7-max if GL fails the pilot. Whichever is used is fixed by the pilot and not changed afterwards, and the resolved slugs are logged as provenance.

6. What is registered, before the run

Q1 primary. EAS DETECTED count over the 20 تجنیس blocks, against GLA and ROS, paired by block. Sign test on discordant blocks, one-sided in the direction Eastwick's declaration predicts (EAS higher). Reported as a reference tail under exchangeability of the hand label, not as a Type-I error guarantee — RS-20260825b §2, the critic's finding accepted there, governs here too: these are fixed printed texts, not randomly assigned treatments.

Q2 secondary, free. The notes census of §4.8. Registered question: where a hand does not reproduce the play, does it report it? No direction registered; this has never been measured.

Q3 secondary. LED's DETECTED count — what a translator who takes Eastwick's sentence as an instruction and actually tries reaches on the same 20 blocks. A ceiling, not a comparison: the lead is one hand, working in 2026, with R52 aimed at exactly this figure.

Exploratory, reported and not interpreted. DISPLACED counts (a play somewhere else in the block — compensation, which framework/v0.2 §7.12 item 2 says is available); prose-vs-bayt split; per-seat rates; the two PRIMED blocks in and out.

The lead's own prior, recorded now so it cannot be claimed afterwards. internal-judgment-only: I expect Q1 to fail — EAS at or near GLA and ROS. The ground is RS-20260822c-persian-hands, where Eastwick answered Sa'di's rhymed prose at 3 of 32, the same rate as the hands who promised nothing, and where his declared verse policy was kept at 38 of 42. He keeps what he declares about form; the question is whether he keeps what he declares about wordplay. This prior is not evidence and does not enter any figure.

7. Failure criteria — written before the run, and binding

• F1, the control gate. Pooled over the three published hands, the false-positive rate on the 12 control blocks must be strictly below the pooled HIT rate on the 20 تجنیس blocks. If it is not, Q1 is withheld in full and the run reports an instrument that cannot see the thing.
• F2, discrimination. At least one arm must show a DETECTED rate at least 0.25 above its own control rate. If no arm does, the instrument floors and no comparison between arms is licensed.
• F3, missingness. A cell that returns no parsable body after one re-dispatch is dropped and counted. If more than 10% of stage-P cells drop, the run is void.
• F4, alignment coverage. If a locus cannot be aligned in a hand because the hand does not render that material at all, the block is coded ABSENT for that hand and excluded from the paired test for that pair, with the count reported. If more than 4 of 20 blocks are ABSENT in any hand, that hand is reported separately and not pooled.
• F5. If LED measures DEPENDENT? against EAS, Q3 carries the measurement and supports nothing.
• F6, the verse-rhyme confound (§3a). RHYME-ONLY quotations are not hits, and the 10-block prose stratum governs on disagreement with the 20-block figure. Each hand's bayt-end rhyme rate on this span is counted mechanically and reported beside the verse stratum, whatever the outcome.

What no outcome of this run licenses. Three model seats are not Victorian readers and no sense here is Tier-D calibrated; every figure is provisional. A null on Q1 would mean these seats cannot be shown to find more play in Eastwick's English, not that Eastwick broke his promise — RS-20260826c-register-cost is the standing example of that distinction and its lesson is imported, not re-derived.

8. Cost

Pre-flight, built from max_tokens and the worst plausible provider (note (abc)):

Expected, at S221/S225 observed rates on short prompts: $0.90–$1.30. The UTC day 2026-08-27 opens at $0.00 of $5.00. Translation, alignment, the notes census, the contamination check and verification cost $0.

9. DESIGN v2 — what the pre-run critic changed, and what actually runs

Everything in §§1–8 above is the frozen v1 and is kept for the record. Where this section differs, this section governs. The full finding-by-finding answer is critic-response.md; two seats returned 21 findings between them and both independently found the same two BLOCKING defects. Nothing had run.

9.1 The pool

The hand-built inventory is discarded. enumerate_loci.py enumerates candidates from the Persian by a rule fixed before it was run; adjudicate.py records a disposition for every candidate, on the Persian alone, before any English of the span was opened, and fails if one is missing. 62 candidates → 22 loci in 14 prose blocks. loci-frozen.md v2 carries the tables.

9.2 What runs

14 تجنیس prose blocks + 7 prose control blocks = 21 units × 4 arms × 3 seats = 252 calls.

The verse blocks are dropped entirely. Eastwick rhymes 38 of 42 bayts on this book and the two comparators rhyme none; a reader asked for a play upon words in a rhymed quatrain quotes the rhyme. No scoring rule inside a verse block removes that, so the verse blocks are enumerated, adjudicated, published and not scored. §3a's RHYME-ONLY rule survives as a belt-and-braces check on the prose.

Arms: EAS Eastwick (1880 ed. of the 1852 translation) · GLA Gladwin (1865 reprint of 1806) · ROS Ross 1823 · LED T-gulistan-bab2-R52-v1.

Seats, frozen now, not chosen by the pilot: P1 openai/gpt-5.6-terra · P2 google/gemini-3.6-flash · GL z-ai/glm-5.2. The pilot's only role is mechanical: if GL returns fewer than 5 of 6 parsable bodies, or any empty body at finish_reason: length, the seat falls back to QR qwen/qwen3.7-max. No content of any pilot answer may enter the choice.

Exposure, frozen: main text only. Footnotes, note markers, bracketed glosses, running heads, page numbers and italic markup are stripped; paragraphing is preserved. The apparatus is the object of the separate free stage N.

9.3 The instrument and the hit rule

The seat sees one English passage, is told only that it is from a book of moral tales, and is given a frozen definition with positive and negative examples. It must return, for each item it finds, both expressions and the relation between them, in a fixed parsable format, or NONE.

For each locus × hand the lead freezes two minimal member spans of 1–3 words — the English rendering of each Persian member — in materials/alignment.md, committed before dispatch.

Only HIT counts. A block is DETECTED for a hand when a majority of the three seats returns a HIT on at least one of that block's loci.

Alignment statuses, frozen per locus × hand: PRESENT · OMITTED (a member is not rendered) · COLLAPSED (both members rendered by one English word) · RELOCATED (rendered in another block) · ABSENT (the material is not in the English at all). Only PRESENT loci can produce a HIT; the rest are counted and reported, and a block all of whose loci are non-PRESENT in a hand is excluded from that hand's paired comparison.

9.4 What is registered

Q1, descriptive. The paired differences EAS − GLA and EAS − ROS in DETECTED count over the 14 blocks. Both are reported. A reference tail is computed for each by exhaustive enumeration over discordant blocks and Holm-adjusted across the two — labelled a reference tail under exchangeability of the hand label, not a test, because these are fixed printed texts. If the two comparators point in opposite directions, the result is reported as a null.

The asymmetry, registered in advance. A positive difference is confounded with everything else that differs between an 1806, an 1823 and an 1852 page — diction, literalness, verbosity, verse policy, editorial modernisation — and licenses nothing about Eastwick's declaration. A null is harder to explain away: a translator who printed that he uniformly did his best to preserve the play upon words, whose page a reader cannot tell from two hands who promised nothing, has left no trace of the promise on this instrument. Neither outcome shows that he kept or broke it, and the design does not claim it will.

Q2, free. The notes census: where a hand does not reproduce a play, does its apparatus report it? REPRODUCED / REPORTED / NEITHER.

Q3, descriptive. LED — one instructed contemporary rendering, not a ceiling. What a hand reaches with Eastwick's sentence as its brief.

Denominators, fixed. DETECTED rate = detected blocks ÷ blocks with ≥ 1 PRESENT locus in that hand. Background rate = control blocks with a majority OFF-TARGET ÷ 7. Dropped cells are excluded from their block's majority; a block with fewer than 2 live seats in an arm is excluded from that arm and counted.

9.5 Failure criteria, restated

• F1 — the pooled background rate over EAS, GLA, ROS must be strictly below the pooled DETECTED rate. Otherwise Q1 is withheld in full.
• F2 — EAS must exceed its own background rate by ≥ 0.25. Otherwise no comparison between hands is licensed. (v1 allowed any arm to satisfy this; P2 was right that it let LED license a run in which the historical text showed nothing.)
• F3 — > 10% of stage-P cells dropped voids the run.
• F4 — alignment statuses as §9.3.
• F5 — a DEPENDENT? contamination measurement on LED attaches to Q3 and supports nothing.
• F6 — RHYME-ONLY is never a hit.

9.6 Cost, revised

252 stage-P calls, 6 pilot, ≤ 30 re-dispatches. Declared ceiling $2.20, of which $0.115140750 for the critic is already spent. Expected $0.70–$1.00.

9.7 The contamination measurement, run before dispatch — and it hits the comparators, not the lead

tools/dependence_check.py over the twenty frozen unit texts, all six pairs, before any stage-P call was dispatched. Data: materials/dependence.json.

The shared twelve is "plunder as soon as he had got out of sight of the", and it is not a coincidence: Eastwick's own footnotes on these very pages cite Gladwin and Ross by name (A-gulistan-hands §4 already records this, and his note 143 on this span says "Gladwin and Ross translate as above, and I am content to follow them").

The consequence, registered here before the run rather than discovered after it.

1. EAS − GLA is a comparison between dependent texts. Where Eastwick follows Gladwin's wording he cannot differ from him, in either direction, for reasons that have nothing to do with any policy. That comparison is reported and is not the one that decides anything.
2. EAS − ROS is the comparison between independent hands, and on disagreement it governs.
3. LED is measured clean against all three, so F5 does not fire and Q3 stands as reported. Its longest run against EAS is 11 tokens, at the top of the clean band and below this project's own record for an independent pair.

And this is a finding in its own right, not only a caveat. Two of the three published English hands of the Gulistan on this project's shelf are not independent witnesses to what an English translator does with Sa'di's wordplay. Any future use of A-gulistan-hands that pools the four hands has to carry it.

Erratum 1 — the Arnold exclusion was wrong (added 2026-08-31, S235; the design above is not edited)

The materials table declares: "Arnold 1899 (frompersianguli00arnogoog) is a verse selection and does not cover this span. Declared, having been checked, not assumed." The check was wrong. That volume — From the Persian: the Gulistan … the first four bábs — prints باب دوم whole, headed CONCERNING DARWEESHES, from its tale I, which is Sa'di's حکایت ۱, onward. Read and verified at S235 while coding E-20260831-arabic-in-persian; the confusion appears to be with gulistanbeingro00arnogoog, a different Arnold book that is a verse selection.

No number this design produced is affected — RS-20260827-declared-play reported three hands and said three. The study is under-powered by one hand rather than limited by the archive, and an extension should code Arnold. Recorded here rather than by editing the frozen text, per R05 rule 2's principle and charter §8.