Repository path: workshop/experiments/E-20260901-inversion-habit/design-v2.md · rendered 2026-09-09

Design v2 — the operative design, after 28 critic findings

This supersedes design.md, which stays frozen as the record of what was proposed. Everything below is unchanged from v1 unless the disposition in critic-response.md says it changed. §1–§3 of v1 (the question, the wire, the materials) stand; the sections here replace v1 §4–§9.

4. Cells, redefined from the Persian rather than from Payne's English

C1-A3 / C2-A3: obligation must not be defined by the outcome. It is now read off the source.

The eleven matched odes are Leaf III, IV, V, VI, VII, VIII, IX, X, XI, XII, XIII = Brockhaus 6, 8, 43, 44, 52, 79, 84, 121, 123, 151, 198, read from the numbers Leaf himself prints at the head of each ode (runs/RS-20260830-leaf-contract/extract.py, LEAF_BROCKHAUS). Payne's volume 1 is Brockhaus I–CC, which is why eleven and not twenty-six.

Cells are labels, not partitions: an item may sit in more than one (an INT-P line is also SPILL-Y or SPILL-N). Items are drawn once, deduplicated by uid, and coded once. Sampling is seeded random.Random(20260901) and is done before any rate is computed.

5. What a seat sees and returns

One printed couplet with one line marked >>. No hand, no book, no ode number, no Persian, no rule, no prediction; the words radif, refrain, Persian, Hafez, ghazal, translation never appear.

The decision procedure, given to the seat verbatim (C1-A6, C2-A7):

1. Find the finite verb of the clause that the marked line's last word belongs to.
2. Say what role the last word plays in that clause.
3. Ask: writing these same words as ordinary modern English prose, would that word still be last?

Returned per item:

• ORDER — INVERTED (ordinary modern prose would not leave it last) / CANONICAL / UNCLEAR.
• WHAT_MOVED — OBJECT_OR_COMPLEMENT_BEFORE_VERB / SUBJECT_AFTER_VERB / ADJUNCT_OR_ADVERBIAL_LAST / MODIFIER_AFTER_HEAD / NONE / CANNOT_TELL.
• CLASS — VERB / NOUN / ADJ / ADV / PRONOUN / PREP / OTHER.
• WHY — at most eight words.

Seats, fixed at three in advance (C1-A13, C2-A15): P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, P3 x-ai/grok-4.5. Non-Anthropic (charter §5). P4, P5 and z-ai/glm-5.2 are out under notes (bps), (bne), (brt).

Batching 22 items per call, shuffled under the seed, calibration interleaved. The cap is probed per seat at 22 items — note (bsf), five firings, most recently $0.301 wasted.

6. Two gates, and the second is the one that decides

Floor — 24 synthetic pairs. Each pair is one real printed line whose printed order is canonical, plus a counterpart made by moving that line's finite verb to the end and changing no word. Exactly one member is INVERTED, known by construction; members never share a batch. Gate ≥ 21 of 24. This shows a seat can detect the contrast at all. It shows nothing about natural verse displacement, which is C2-A4's point and is why it is a floor and not the gate.

Validity — 40 real-item gold. Forty study items, drawn by the seed, printed by gold_blind.py with hand, ode and cell stripped and the order shuffled, coded by the lead on ORDER alone, frozen and committed to gold.json before any seat is dispatched. Gate: a seat votes only at ≥ 0.75 agreement with gold on ORDER. Gold items are study items, so this costs no extra calls. The circularity is declared: the lead wrote the definition, so this measures whether the seats apply this definition, not whether it is the right one.

Positive instrument control (C1-A12, C2-A5). Cell CTRL+ is Payne's rhyming hemistichs in odes whose carried tail is itself a finite verb: lines that visibly end on a verb with its object in front. If CTRL+ returns below 0.70 INVERTED on the voting seats, the instrument is declared broken and no primary is reported at all, whatever the two gates say.

Hand-identification probe (C1-A5, C2-A6), dispatched before the coding run: each seat sees four labelled couplets per book, then twenty held-out couplets, and guesses the book. Accuracy is reported per seat. It measures the leak; it cannot measure whether the leak biased ORDER.

Reliability: mean pairwise agreement and unanimity on ORDER across voting seats. Non-unanimous items are resolved by majority; no majority ⇒ dropped, with the drop rate printed per cell.

7. Registered predictions

• PR1 (matched-poem, primary). Over the eleven odes both hands rendered, Payne's INVERTED rate at unobligated line ends (INT-P) exceeds Leaf's (INT-L) by ≥ 0.20. This is a descriptive rate difference on matched material. It does not by itself license a causal reading, and the decision matrix no longer asks it to (C1-A1, C2-A1).
• PR2 (spill, primary, and not definitional). Among Payne's unobligated line ends, the INVERTED rate in odes where he carries a tail (SPILL-Y) exceeds the rate in odes where he carries none (SPILL-N) by ≥ 0.10. Refuted below 0.10 or reversed. This is the generalised form of the translation limb's LOCAL / SPREAD grades.
• PR2b (rhyming slot, secondary). O-P exceeds N-P by ≥ 0.15. Reported with its definitional component named: where Payne carries a finite-verb radif, the line ends on a verb because he carried it.
• PR3 (the ROOM account). Payne's median hemistich syllable count exceeds Leaf's by ≥ 2. Reported only after the fallback syllabifier is validated on 30 hand-counted lines, 15 per hand (C2-A10). The lead's prior, registered: this fails.
• PR4 (the priced move). Among items coded INVERTED, WHAT_MOVED is OBJECT_OR_COMPLEMENT_BEFORE_VERB in > 0.70, computed excluding CTRL+ and O-P, where a carried verbal tail could force the answer (C1-A4, C2-A12).
• PR5 is withdrawn. The mechanical whole-book coder is removed from the design entirely (C2-A14); no whole-book rate is reported anywhere, including in the journal.

Statistics (C1-A11). Every rate carries a Wilson 95% interval, every difference a Newcombe 95% interval, and both primaries additionally an ode-clustered bootstrap interval, 5,000 resamples of odes with replacement. Outcomes are three-way: supported (interval excludes the bar on the predicted side), contradicted (excludes it on the other), indeterminate (spans it).

8. The decision matrix, rewritten

PR3 is orthogonal and is reported either way.

9. Failure criteria

1. CTRL+ below 0.70 INVERTED ⇒ instrument broken, no primary reported.
2. Any seat below either gate (≥ 21/24 floor, ≥ 0.75 gold) is dropped and reported. Fewer than two voting seats ⇒ no ORDER figure at all.
3. Mean pairwise agreement on ORDER below 0.70 ⇒ the construct is reported as not reliably codeable and the primaries are withheld.
4. Drop rate above 0.20 in any cell ⇒ that cell is reported as a range over the dropped items.
5. Final-token damage rates differing by more than 0.10 between the hands ⇒ PR1 is reported as a range, not a point.
6. Denominator below 25 ⇒ counts, not rates (C2-A13).
7. Sensitivity analysis, always run: every primary recomputed excluding UNCLEAR items and damaged-final-token items.

10. Limits carried into the result page

1. Two hands, and register is still confounded with hand. The matched-poem restriction removes poem selection; it cannot remove the fact that Payne archaizes and Leaf does not. The within-Payne primary PR2 is where register is constant, and that is why it is a primary.
2. Eleven poems. Payne's volume 1 stops at Brockhaus CC and most of Leaf's odes are past it.
3. No page images (C1-A9, C2-A9). §6's damage rate, the printed couplets and the sensitivity analysis are what stands in for an audit, and they are weaker than an audit.
4. The gold set is one coder, blinded to hand and cell, who also wrote the definition.
5. Prose order is the yardstick. Inversion is ordinary in verse of both hands' periods; this measures displacement from prose order, which is what §7.42 priced. It is not a claim that a reader notices anything.
6. Tier D is NOT PASSED. No sentence here may say a reader hears, prefers or wants anything.

11. Pre-flight cost

UTC day 2026-09-01: no prior row, the whole $5.00 available. Declared ceiling $3.50 (C1-A13, C2-A15), three seats fixed.

The translation limb, all extraction, the eleven-ode intersection, all syllable counting, all sampling, the gold coding and every interval are the lead's own and are not ledgered (charter §3, A4).