Repository path: workshop/experiments/E-20260901b-leaf-pattern/design.md · rendered 2026-09-09

E-20260901b-leaf-pattern — pattern or count?

ARM-leaf-contract step 2 (T4), 2026-09-01 (S238). Frozen before any figure of the primary was computed. Tier D is NOT PASSED; nothing here is a quality claim.

1. The question

RS-20260830-leaf-contract §4 established the necessary condition of Walter Leaf's L-METRE clause — his English line has the syllable count of the Persian measure his own table names — and said in terms that the clause itself was not measured:

"A syllable count is a necessary condition of 'one measure'. Whether Leaf holds the long/short pattern line by line is the clause itself, and it is not measured here."

So: does an English translator who signs a quantitative contract for Persian verse deliver the pattern, or only the count? That is a fact about carrying quantitative verse into a stress language, not a fact about this project.

And there is a second question the same instrument answers, which is Leaf's own printed diagnosis of when the thing is impossible. Introduction p. 12:

"many Persian metres so abound in long syllables that the English language will not supply stress enough to reproduce them. In the metres numbered in the Table 4 and 6, for instance, there are three longs, and in 17 four, to every short."

That is a third printed self-report by a translator, of a kind different from the two A-leaf-hafez §4 already tested: not a claim about his frequency (L-REPEAT) and not a misgiving about one ode (XXI), but a claim about which measures the target language can and cannot hold. It is checkable on his own 28 odes.

2. What the contract actually says, and why the naive test would be wrong

Leaf's clause is asymmetric, and he says so in the same paragraph:

• "I have therefore accepted all stressed syllables as long" — a stressed syllable stands at a long position. This is the constraint. It can fail.
• "It is not possible to treat all unstressed syllables as short … in a large majority the quantity is uncertain." — a long position need not carry a stress. This is not a constraint and testing it as one would test a contract nobody signed.

So two rates are computed and only the first is a test of the clause:

3. Materials

4. The instrument, and when it was frozen

runs/RS-20260901b-leaf-pattern/scan.py, written and frozen before this design was drafted and before any scheme was read. A syllable is stressed iff its word is polysyllabic and CMUdict marks that syllable primary, or its word is a monosyllable outside a declared closed-class list — which is Leaf's own licence to "dock the smaller parts o' speech". Unresolvable words make their whole line unusable; nothing is guessed. Proper names are deliberately not in the lexicon.

Measured before any metre was consulted: 387 of 444 lines resolve (0.872). Of those, 301 have a syllable count exactly equal to their metre's, in all 28 odes, and only those enter the primary — position i of the line against position i of the scheme, with no alignment judgement anywhere.

Known limit, stated now rather than after: word-level lexical stress is not line stress. A monosyllabic content word that a reader would throw away still counts as stressed here. This is why the primary is a contrast and not a level.

5. The gate, and it has already run

runs/RS-20260901b-leaf-pattern/gate_metres.py. The whole step rests on 24 schemes read by one hand at S233 that nothing has ever checked. Two independent checks:

• (A) Ganjoor. For the 20 ode-rows with a Ganjoor slug and a named metre, expanding Ganjoor's classical rukn names by the standard arud table gives a scheme built by nobody here. Result: 20 of 20 agree with Leaf's transcription at every position, covering 14 of the 24 metres. sh115's own metre is among them.
• (B) Leaf's arithmetic. His prose says metres 4 and 6 carry three longs to every short and 17 carries four. Transcribed: metre 4 = 2.75 : 1 OK, metre 6 = 3.00 : 1 OK, metre 17 = 1.33 : 1 FAILS.

Metre 17 is therefore excluded from the rival pool, with that reason, before the primary runs. It is a 14-syllable scheme and would otherwise be a rival for eight odes.

6. Procedure

1. Code every line with the frozen coder. Keep the 301 that resolve and match their metre's length.
2. For each kept line compute VIOL and FILL against its own metre.
3. The primary is a within-line contrast. For the same line, compute VIOL against every rival scheme in Leaf's table of the same syllable count (excluding its own and excluding metre 17). Any bias in the coder — the function-word list above all — applies identically to own and rival and cancels. Aggregate to the ode, then test across the 28 odes.
4. Matched-shorts sub-analysis. A rival with more short positions is a harder target. Repeat the contrast restricted to rivals with the same number of short positions, where any exists.
5. H2, Leaf's diagnosis. Spearman ρ across the 28 odes between the long-fraction of the ode's metre and the ode's mean FILL.
6. The prose floor. Leaf's own Introduction prose, cut at word boundaries into windows of exactly N syllables, coded and scored against every scheme of length N. This is what a non-metrical English text scores.
7. The translation limb. The same coder over T-hafez-derakht-R57-v1's 14 lines against metre 6, reported beside Leaf's ode IV, the only metre-6 ode he made.
8. Sensitivity: rerun the whole primary with FUNCTION_WIDE (23 more monosyllables treated as unstressed) and with CMUdict secondary stress counted as stressed.

7. Registered predictions

P1 and P4 are the ones that carry the session. P2 and P3 describe; P5 calibrates; P6 is a check on the hand, not on Leaf.

8. Failure criteria, declared

9. What this cannot show

• Nothing about 1890s stress. CMUdict is a present-day American lexicon. Leaf's own footnote says his r is silent before a consonant and that stress "is largely rhetorical". Every figure here is what a modern coder reads, not what Leaf heard.
• Nothing about Payne. His OCR remains unusable (A-leaf-hafez §5) and no figure of his is computed.
• Nothing about quality, of either hand. Tier D is NOT PASSED.
• Nothing about the 57 lines the coder drops, which are disproportionately the ones carrying Persian proper names and OCR corruption. Coverage per ode is reported so the reader can see where the book is thin — ode XVII keeps only 3 lines.