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

E-20260827c — the second declarer: does printing a policy leave a trace on the page?

ARM-balanced-period step 2 of 2 · T4 · frozen before any statistic below was computed, except where §6 says otherwise and says why.

1. The question

RS-20260826-balanced-period measured Preston 1850's printed English against the sentence he printed about it, and found that he keeps the length ceiling and not the balance. Its §8 records the limit that makes the finding hard to read:

One declarer, no replication of the declaring. Preston is the only nineteenth-century English translator of al-Ḥarīrī who printed this policy. Chenery is a comparison, not a control.

Chappelow 1767 is the other declarer, and step 1 excluded him because he prints running prose with no colon marking. Under U-COMMA and U-STRONG — one mechanical rule applied identically to every text, which is not any hand's own marking — he can be measured like everyone else.

Do the hands that printed a policy differ, on their pages, from the hand that did not — and on which measures?

Amended after the pre-run critic (§10, P1 finding 3). The question was first written as does printing a policy leave a trace on the page, which is causal and which three fixed hands cannot answer: every contrast between them is also a contrast of person, date, edition and printing-house practice. Nothing below attributes any difference to the act of declaring.

Preston printed two things: a refusal of the rhyme, and a positive rule about length and balance. Chappelow printed only the refusal. So the design separates two questions that step 1 could not:

• the refusal has two declarers and one non-declarer, and is a comparison of like with like;
• the positive rule has one declarer and two non-declarers, which is the strongest form of it the surviving record allows.

2. A correction to A-hariri-hands §4a, established before the design was written

The anchor states, of the second Assembly: "Chappelow 1767 does not contain this Assembly." That is false. The 1767 volume's Assembly II is headed HULWANENSIS, and its Assemblies I, II, III, IV, V and VI are al-Ḥarīrī's first six in the standard order (SANANENSIS, HULWANENSIS, [the third], DAMIATENSIS, CUFENSIS, MARAGENSIS). The claim was made at S223 without opening the volume's table of contents, and it removed the second declarer from a panel he belongs in.

The anchor is corrected, and the correction is what makes the two-panel design below possible.

3. Materials

Three published hands, two Assemblies (Panels A and B); the same three plus the lead, on a third (Panel C).

Panel C's Arabic and the lead's arm are both new this session. Assembly III was the natural third panel and was rejected: the lead had read about six hundred words of Chappelow's Assembly III while establishing the volume's contents, and six hundred words of a published rendering is priming. Seventy words of his Assembly IV had been read in the same check; that is declared on the translation page and measured in §7.

Recovery. slice.py records the line ranges taken from the two volume OCRs; extract_chappelow.py, extract_preston.py and extract_chenery.py separate translation from apparatus, each writing everything it discards to a raw/*-dropped.txt file. Preston's and Chenery's Assembly I and II texts are the files E-20260826-balanced-period measured, reused unchanged, except that Chenery's Assembly I is re-cut whole — step 1 §8 records that the stored S222 comparator ended before the close of the maqāma.

4. Procedure

Identical to step 1 except where stated. metrics.py is imported from E-20260826-balanced-period, not copied, so the measures are the same code that produced the published figures.

Two segmentations, both uniform, neither any hand's own, because Chappelow has no own unit below the paragraph and no unit is invented for him:

• U-COMMA — split at . ; : ? ! — – and ,
• U-STRONG — split at . ; : ? ! — – only

OWN is computed for Preston, Chenery and the lead's arms only, and is reported only for the extraction check.

Two units of length, both reported: WORDS and SYLLABLES. Words are primary throughout, and this is registered here, before measurement, for a reason that is not about the result: Chappelow's 1767 long-s is OCR'd as f at a high rate, which corrupts a syllable count and not a word count. The out-of-dictionary token rate is reported per text.

The shuffle null, as step 1: the observed mean absolute difference between neighbouring unit lengths, divided by the mean over 10,000 random reorderings of that text's own unit lengths, seed 20260827. A hand with uniformly short units gets no credit for smoothness it did not arrange.

A new null, for the chime. The chime rate at adjacent unit ends is tools/rhyme_pairs.py's verdict — STRICT, IDENTICAL or any NEAR — counted over adjacent pairs. It is compared against a bearer-permutation null: the text's own unit-final words are shuffled 2,000 times (seed 20260827) and the chime rate recomputed, which holds the hand's vocabulary and unit count fixed and destroys only the adjacency. A hand whose English happens to be full of -tion endings gets no credit for a chime its word-list produces by itself.

5. Registered predictions

All four are associations among three fixed hands. None is a causal claim about declaring (§10, P1 finding 3).

P1 — the punctuation-defined unit length. In words, under U-COMMA only, on Panels A and B, Preston's 95th-percentile unit length is below both non-declarers'. Two cells; support requires 2 of 2; any failure contradicts it.

Amended after the critic (§10, P2 finding 3). As first written, P1 was registered on both segmentations and carried a sentence saying where it was expected to fail — which is a fallback authorised before the run and is exactly the defect §8 forbids. U-STRONG is removed from the registration and its figures are reported descriptively. The measure is renamed: it is a 95th-percentile punctuation-defined unit length, not a ceiling (§10, P1 finding 4). The observed maximum is reported beside it, because Preston's printed rule is about a maximum.

P2 — an equivalence claim about the shuffle ratio, and nothing wider. In words, in each of the 4 (panel × segmentation) cells, the three published hands' shuffle ratios span less than 0.15. Contradicted by any cell whose spread is 0.15 or more. Differences below 0.005 are ties and are reported as ties.

Amended after the critic (§10, P1 findings 5 and 9). The first version predicted that Preston would not rank lowest in 3 of 4 cells, which is the expected outcome under chance and is not evidence of anything; and it had no tie rule. P2 now states an equivalence and can fail. It is about the registered shuffle ratio only and is not evidence that no hand's prose is balanced in any other sense of the word — paired-clause symmetry and syntactic parallelism are not measured here.

P3 — the chime at adjacent unit ends. For each text, the ratio of the observed chime rate to its own bearer-permutation null mean. In each of the 4 (panel × segmentation) cells, both declarers' ratios are below the non-declarer's by more than 0.10. A cell in which the largest difference is 0.10 or less is a tie and counts as neither support nor contradiction. A cell in which any text's null mean is 0 is void and is reported as void.

Amended after the critic (§10, P2 finding 2 and P1 finding 6). Adjacent-clause chime in unrhymed English prose can be 0 for every hand, in which case the first version's strict rank test would have failed on arithmetic rather than on anything about the pages. The zero rule and the minimum difference are registered here, before the counts are computed.

P4 — how an expanding translator expands. Registered on the one uncomputed quantity: Chappelow's mean U-COMMA unit length in words is within 20% of Chenery's on both panels. Contradicted if it fails on either panel.

Amended after the critic (§10, both seats, BLOCKING). The first version added a second clause — units per Arabic colon exceeding Chenery's by at least 50% — presented as independent confirmation. It is not independent: it is entailed by the first clause together with the word totals §6 says were already visible. Both seats derived the entailment. The clause is deleted. Units per colon are still reported, as arithmetic and not as evidence.

P5 — the lead's arms are descriptive and predict nothing. R53 on Panel C is reported beside Chappelow's Assembly IV as the observed shape of a hand doing his page-policy on purpose, exactly as step 1 §6 reported R50 beside Preston. No prediction is registered on it, because the hand that made it is the hand reporting it — see R53 §What the rule deliberately does not say.

6. What was already visible when this design was frozen

The extraction scripts print word counts, so the following were known before the predictions above were written, and none of them may be read as a prediction:

• prose words recovered: Chappelow 2,461 / 2,886 / 3,280 on Assemblies I, II, IV; Preston 1,032 / 1,157 / 1,258; Chenery 1,165 / 1,091 / 1,369; the lead's R53 1,747.
• so Chappelow's Assembly IV expansion is 5.40 English words per Arabic prose token against Preston's 2.07 and Chenery's 2.25, and the lead's R53 2.88.
• C4 was run before the critic pass and its numbers are therefore also known; see §7.

Nothing else — no unit count, no length distribution, no shuffle ratio, no chime count — had been computed for any text when this file was frozen, and nothing else was computed between the freeze and the amendments in §10.

7. Controls and checks

• C1a extraction check, on unchanged inputs. Run on the files E-20260826-balanced-period measured — its own Preston Assembly I and II texts and its own (truncated) Chenery Assembly I text — the pipeline must reproduce that result's published figures exactly: 107 and 123 printed lines, longest line 14 and 13 words, CV 0.198 and 0.177, and the six OWN/U-COMMA/U-STRONG shuffle ratios of its §5 table. A mismatch halts the reading of every new figure.
• C1b the re-cut Chenery Assembly I is a new text. Its figures are reported as new, with the difference from the truncated file stated. A page-image audit of it was not done — the critic asked for one and no page images were consulted this session, which is the same limit step 1 recorded at its §8.
• C2 OCR damage, in two parts. (i) Out-of-dictionary rate per text, reported; words primary for every registered prediction. (ii) A declared long-s repair (§10, P2 finding 4): before the chime and syllable measures only, an out-of-dictionary token whose f→s variants yield exactly one dictionary word is replaced by it. Applied identically to every text; the repaired-token count is reported per text; word counts are computed on the unrepaired text and are therefore untouched by it. C1a is run with the repair off, so the reproduction check is exact.
• C3 extraction residue. The count of catchwords the rule leaves standing in each Chappelow text, and the Preston Assembly IV OCR loss — page 384 of his book is not in the OCR, about six printed lines covering p126–p131.
• C4 contamination — a descriptive provenance check, not a control (§10, P1 finding 7). tools/dependence_check.py on the lead's R53 against all three published Assembly IV texts and on the three published hands against each other, run before any of them was read. It can detect copied strings and nothing else: it cannot detect the declared priming from the seventy words of Chappelow's Assembly IV the lead had read, and it is not treated as though it could.
• C5 punctuation density (§10, P1 finding 5-b). Commas and strong stops per 100 words, reported for every text. U-COMMA and U-STRONG units are made by punctuation, so a difference in unit length between two hands is in part a difference in how they point their prose, and the reader is given the quantity that difference runs through.

8. Failure criteria

Each prediction fails on any single contradicting cell, and a mixed outcome is reported as mixed and never as support. Ties and void cells are as defined in §5 and count as neither. If C1a fails, nothing else in the run is reported as a result. No sentence of the result page may attribute a measured difference to the act of declaring a policy.

9. What this design cannot show

• Nothing here is a reader measurement. It is a measurement of pages.
• Chappelow's 1767 OCR is the worst text in the study, and every conclusion involving him has to survive in words.
• Three hands are not a sample of translators, and two declarers are not a replication in any statistical sense. What the second declarer buys is that the positive rule now has two hands who did not print it, and the refusal has two who did.
• The hands are not independent: Chenery cites Preston and De Sacy (A-hariri-hands §5), so a result in which the two declarers sit together and Chenery apart is an association among three historically linked hands and not a policy effect.
• Panel C is an incomplete descriptive case study, not a third panel (§10, P1 finding 8): Preston's Assembly IV is missing a printed page in the OCR, and the lead's R53 arm is not an independent observation of anything. No registered prediction is evaluated on Panel C.
• No page images were consulted, so the OCR's punctuation — which is what makes the units — is unaudited. The critic named this and the remedy it asked for is not affordable here; C5 reports the quantity the threat runs through, which is less than an audit.

10. Amendments after the pre-run critic

Two non-Anthropic seats, one round, raw/critic-round1b.json, critic-response.md. Both returned a BLOCKING finding and both found the same one. Twelve findings, all twelve accepted, none overruled; two remedies were accepted in part because a page-image audit is not available this session, and that is said where it applies. Every amendment above is dated to this pass and no statistic was computed between the freeze and the amendments.