Repository path: workshop/experiments/E-20260901-inversion-habit/design.md · rendered 2026-09-09

Was the line-end inversion a device Payne bought, or the line he was already writing?

ARM-inversion-price step 2 (T3), 2026-09-01. Frozen before any rate was computed.

1. The question, and why it is not method work

RS-20260829-radif-hands §4 explained the difference between two published English Hafizes in one sentence: Payne carries the Persian radif and Leaf does not, because "Payne writes an archaizing nineteenth-century English in which object–verb order at the line end is available, and Leaf, writing plainer, does not use it." framework/v0.2 §7.41.3 printed that as advice.

RS-20260829b-inversion-price measured it and the register clause failed: the inversion costs 2.407 points of ten in a plain lexis and 1.926 in an archaic one, an interaction of +0.569 against a registered bar of 2.556, and 101 of 106 decisive forced choices reject the inverted order in both registers. §7.41.3 was withdrawn as §7.42. That leaves the handbook recording a negative and nothing else, and leaves the original question open: what did separate Payne from Leaf?

The arm's step 2 names two accounts that the published pages can distinguish without a model:

• HABIT — the inversion is a standing property of Payne's line, present wherever he writes and not bought for the radif at all. Leaf's line does not have it. Then the radif is carried by a hand who was already writing verb-final lines, and refused by a hand who was not.
• ROOM — the difference is metrical: Payne writes a longer line, with room to front a phrase and still reach the end. This is what §7.42.4's uncounted end-weight conjecture would predict.

The subject rule (continue-prompt.md §4.5), in one sentence. This unit teaches a translator of Persian verse whether the word-order move that carries a ghazal's commonest formal device is something you can adopt for one line, or a grammar you have to be writing the whole poem in — and it answers that by counting what two published translators actually put at their line ends. The object is a translation move on published pages; no figure of this project's own is under audit.

2. The wire between the two limbs

The translation limb generates the claim; the study limb tests it on two published books.

T-hafez-daasht-R58-v1 rendered غزل ۷۷ whole in a declared plain modern register, carrying the finite transitive radif داشت as had at 9 of 9 rhyming positions. Its frozen log grades the reach of each inversion under R58 §4 and reports LOCAL 6, SPREAD 3, REBUILT 0, REFUSED 0 — that is, a plain modern hand bought the inversion one line at a time, and never had to reshape a predicate the way Payne reshapes one at sh130 (RS-20260831b-radif-hands §6). The log's closing sentence is the claim under test:

If that generalises, the inversion is something a translator buys where it is needed rather than a grammar the whole poem has to be written in — and a hand who inverts at a radif position should not, on that account, be inverting anywhere else.

The log was frozen and committed (c95add53) before this design existed, and before Payne's ode 69 — his rendering of the same ghazal — was opened.

3. Materials

Both already extracted, both re-used rather than rebuilt:

• Payne 1901, volume 1 entire — runs/RS-20260831b-radif-hands/payne_odes.json, 199 odes segmented at S236, 1,703 bayts = 3,406 hemistichs. Payne prints one hemistich per line.
• Leaf 1898 entire — runs/RS-20260830-leaf-contract/leaf_odes.json, 28 odes, 444 lines = 222 bayts, reconstructed under a layout-and-punctuation rule whose QC read two page scans directly and recovered 22 of 22 lines exactly (RS-20260830-leaf-contract §2). Leaf also prints one hemistich per line, and his own metre table targets 14–17 syllables per line.

The unit is the printed hemistich. Both hands print the same unit, which is what makes the between-hand comparison possible at all.

Cells. A hemistich is OBLIGATED if it is a rhyming position in an ode where that hand carries a repeated line-final tail (coded mechanically at S236's 0.80 bar), and UNOBLIGATED otherwise. The purest unobligated line end in a ghazal is the first hemistich of a non-maṭlaʿ bayt: it carries neither the rhyme nor any repeated tail, in either hand, by the form's own rule.

4. Procedure

Stage A — mechanical, free, whole-book

1. Extract every hemistich of both books with its cell label. Re-run S236's carriage coder on both hands at the 0.80 bar to label odes.
2. Syllables per hemistich, CMUdict via tools/rhyme_pairs.py's loader with a declared fallback syllabifier for the words CMUdict does not hold — which, RS-20260830-leaf-contract §3 warns, is every -eth form, so the fallback is load-bearing here and its rule is printed in the result.
3. A mechanical line-final word-class coder. The lexicon is built from the pooled line-final word types of both books together, with the hand and cell labels stripped, so the lead codes "can this word be a finite verb?" without seeing which hand or which cell it came from. The lexicon and the blinding script are committed.

Stage A is descriptive on its own. Its whole-book rates are reported only if stage B measures the coder's agreement with blind seats at ≥ 0.85 on the sampled items; below that, only the sampled rates are reported, and the coder is reported as failed.

Stage B — blind coding of line ends, bought

Sample. Seeded (random.Random(20260901)), drawn in stage A before any rate is computed:

What a seat sees. One printed couplet, with one of its two lines marked >>. No hand, no book, no ode number, no date, no Persian, no rule and no prediction. The words radif, refrain, Persian, Hafez, ghazal and translation do not appear in the prompt. The couplet rather than the bare line is shown because both hands enjamb the subject across the bayt, and a line judged without its clause would be miscoded; the couplet is the smallest window that contains the clause and it shows no repetition, because a repeated tail recurs across bayts, not inside one.

What a seat returns, per item:

• ORDER — INVERTED if ordinary modern English prose would not leave the marked line's final word last in its clause (most often: the final word is a finite verb whose object or complement stands before it); CANONICAL if ordinary prose would leave it last; UNCLEAR if the text is too damaged to tell.
• CLASS — the grammatical class of that final word in its clause: VERB / NOUN / ADJ / ADV / PRONOUN / PREP / OTHER.
• WHY — at most eight words.

Seats: P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, P3 x-ai/grok-4.5 (config/models.md). Non-Anthropic, as the charter requires. P4 and P5 are out under notes (bps) and (bne); z-ai/glm-5.2 is out under note (brt).

Calibration, and it is certain by construction. Twelve pairs. Each pair is one real printed line of one of the two books whose printed order is canonical — its finite verb precedes its complement and its final word is the head of a noun phrase — together with a counterpart made by hand-moving that same verb to the end of the same line, changing no word. Exactly one member of each pair is INVERTED and it is known which. Members are never adjacent and never in the same batch. Registered validity gate: a seat votes only if it returns ≥ 21 of 24 calibration bits correct. A seat below the gate is reported and dropped, and the primaries are computed on the seats that pass.

Reliability: mean pairwise agreement and unanimity on ORDER across the voting seats are computed and reported. Items on which the voting seats are not unanimous are kept and resolved by majority; where there is no majority (two seats split, or one seat voting) the item is dropped and the drop rate is reported per cell.

Batching: 22 items per call, order shuffled under the same seed, calibration interleaved. The max_tokens cap is probed per seat at the batch size that will actually be dispatched — note (bsf), which has now fired five times, most recently costing $0.301 for nothing when a cap probed on a ten-item batch was used for a larger one.

5. Registered predictions

Written before any rate was computed. Each names the bar and what refutation looks like.

• PR1 (the HABIT account, primary). On UNOBLIGATED line ends, Payne's INVERTED rate exceeds Leaf's by ≥ 0.20. Refuted if the point difference is < 0.20 or reversed.
• PR2 (the local-purchase account, primary). Within Payne, the INVERTED rate at O-P exceeds his U-P rate by ≥ 0.15. Refuted if < 0.15 or reversed.
• PR3 (the ROOM account). Payne's median hemistich syllable count exceeds Leaf's by ≥ 2. Refuted otherwise. The lead's prior is that this fails: both hands print 14–16 syllable hemistichs, and saying so in advance is what makes the failure worth anything.
• PR4 (the lead's own log, generalised). Among Payne's O-P line ends coded INVERTED, the CLASS of the final word is VERB in > 0.70. Refuted at ≤ 0.70. This is the check that "inversion" at an obligated line end is the object–verb move the framework prices, and not some other displacement.
• PR5 (coder validity). The stage A mechanical coder agrees with the adjudicated seats on CLASS = VERB at ≥ 0.85 on the sampled items. Refuted below, and then no whole-book rate is reported.

6. The decision matrix, written before the run

PR3 is orthogonal to both and is reported either way.

7. Failure criteria, declared in advance

1. Any seat below the calibration gate is dropped, and if fewer than two seats pass, no ORDER figure is reported at all and the session reports the instrument failure.
2. If mean pairwise agreement on ORDER is below 0.70, the construct is reported as not reliably codeable and the primaries are withheld — the fate RS-20260828-forced-half records for a construct the seats could not hold steady.
3. If the drop rate in any cell exceeds 0.20, that cell's rate is reported as a range over the dropped items rather than a point.
4. If stage A's coder fails PR5, no whole-book rate is printed anywhere, including in the journal.
5. The OCR is not audited against page images, as at S236 — the project has the text layer only. UNCLEAR returns are the instrument's own report of damage and are counted and printed per cell.

8. Limits known before the run

1. Two hands. Nothing here says what English can host; it says what these two books do. Payne archaizes throughout and Leaf does not, so HABIT and register remain confounded between hands — which is exactly why PR2's within-Payne contrast, where register is constant, is the second primary and not an afterthought.
2. Different poems. Leaf chose 28 odes; Payne translated the book. RS-20260829-radif-hands §6 already refuses a three-hand comparison on this ground. Here the comparison is of a hand-level rate over hundreds of line ends, not of matched poems, and the material confound is named rather than removed.
3. Verse, not prose. "Ordinary modern English prose order" is the yardstick the seats are given; inversion is ordinary in verse of both hands' periods, so the measure reports displacement from prose order, which is what §7.42 priced and is not a claim that a reader notices it.
4. Tier D is NOT PASSED. Nothing here is a judgment of quality, and no sentence may say a reader hears, prefers or wants anything.

9. Pre-flight cost estimate

UTC day 2026-09-01 opens with no prior row: the whole $5.00 is available. Declared ceiling for this experiment: $3.00.

Worst case is built from the max_tokens cap the request permits, not from an expected length (note (abc)).

The estimate exceeds the ceiling in its worst case, so the run is staged: seats are dispatched one at a time and the running total is checked against the ceiling after each seat. If two seats fit and the third does not, the design runs on two seats and says so — the calibration gate needs a seat to pass, not three seats to exist. The translation limb, all extraction, all syllable counting, all sampling and all arithmetic are the lead's own and are not ledgered (charter §3, A4).