Repository path: workshop/experiments/E-20260829b-inversion-price/design.md · rendered 2026-09-09

E-20260829b — the price of the line-end inversion, with the register held fixed

v2, amended before dispatch after two NEEDS-REDESIGN critic verdicts and 30 findings (critic-response.md). v1 is in git at dcb25c13.

ARM-inversion-price step 1 (T3). The materials were written and frozen first (T-hafez-darad-R56-v1, commit dcb25c13) and this design was written after them. The one change the design has since made to the materials is the one the critic required: the archaic arms are now generated by script from the plain arms rather than hand-written (critic-response.md, P1-2 / P3-4 / P3-9). The rendering of record and the plain arms are untouched.

1. The question, and the sentence it is about

framework/v0.2 §7.41.3, published this morning:

"That order is available in an archaizing nineteenth-century verse English and not in a plain one, and Leaf, writing plainer, never uses it and drops every one of those radifs. … So the instruction a translator can act on is not 'you cannot', it is 'you can, at this price, in this register'."

The evidence for it is two published hands who differ in everything at once — era, verse form, diction, and which poems they chose. RS-20260829-radif-hands §11.2 registers that the design which separates capacity from register does not exist. This is that design: one hand, one source, one sense, word order and lexis crossed.

2. Materials — frozen, and machine-checked before this page was written

T-hafez-darad-R56-v1: 18 items, one bayt each, from Hafez sh124 and sh118, both radif دارد. Every item exists in four arms:

The minimal-pair constraint is mechanical, not editorial. Within a lexis, D and I are the same word multiset in a different order. Across a lexis the archaic arm is a positional one-for-one substitution of the plain arm — same token count, every differing position licensed by the frozen map in archaize.py — which is checked position by position, not as a multiset. pair_check.py v2: 926 checks, 0 failures.

The frozen archaism map, applied by script with no editorial rescue: has→hath, your→thy, you→thou/thee (role declared per site), pass→passest, shows→showeth, spills→spilleth, means→meaneth, lacks→lacketh, holds→holdeth, have→hast (2sg auxiliary). Nothing else.

Frozen item codings (R56 §8): what has to move in the I arms — OBJ 11 · ADV 4 · MIXED 3; and whether the D arm is itself already marked in its order — CANONICAL 17 · MARKED 1 (G2-6).

3. Seats

Panel v1, three seats, config/models.md: P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, P3 x-ai/grok-4.5. P4 and P5 are excluded by standing notes (bps), (bne). Temperature 0. Judgment is not parallelised (charter §6); calls are dispatched one at a time.

The lead wrote all four arms and judges none of them (charter §5, R56 §7). Every body reaches a seat unlabelled, with no source, no author, no arm name and no mention of Persian, radif, inversion or register.

4. Stage S — the within-idiom well-formedness rating

Amended after the critic (P1-1, P3-2). The DV is no longer the naturalness sense. Seats rate well-formedness within the passage's own idiom and are told in the prompt not to reward or penalise a passage for which kind of English it is written in. wiki/goodness-senses.md allows naturalness three register anchors and none of them is passage-indexed (D-20260803-15, ratified A), so this measure is declared outside the six senses, named within-idiom well-formedness, and is internal-judgment-only. It is the only measure that can answer §7.41.3, whose whole content is a claim about availability inside a register.

Each of the 72 arm bodies (18 × 4) plus 12 planted-fault controls (6 plain, 6 archaic) is shown to each seat once. Each seat gets an independently shuffled sequence (seeds 20260829 / 20260830 / 20260831, drawn before dispatch, written to order.json), so order effects do not couple the three columns (P1-6, P3-16).

Below is a two-line passage of English verse.

Rate it on ONE thing only: how well-formed the English is WITHIN ITS OWN IDIOM. Some passages here are written in present-day English and some in an older, more formal English. Do not reward or penalise a passage for which of those it is. Ask only this: taking the passage's own kind of English as given, is this a well-formed, idiomatic sentence of that English, or is it forced, awkward, or ill-formed?

This is not a question about whether the passage is a good poem, whether it is beautiful, whether it is accurate to anything, or whether you like it.

0 = ill-formed or badly forced, even for the kind of English it is written in. 10 = entirely well-formed and idiomatic for the kind of English it is written in.

---

{BODY}

Reply with exactly two lines and nothing else: SCORE: REASON:

max_tokens 300, temperature 0. Note (bsf) binds: the cap is not treated as verified by a probe on one item; every body's finish_reason is checked and a truncated body already carrying its SCORE: line is re-parsed, not re-bought (note (brx)).

Repeat control. Six bodies are shown a second time to each seat, at a different position in that seat's differently shuffled sequence: 18 extra calls. This estimates presentation-order sensitivity plus sampling noise, which is what it is called; at temperature 0 it is not a test–retest reliability estimate and is not reported as one (P1-10).

Stage S total: 90 bodies × 3 seats = 270 calls.

5. Stage F — the within-lexis forced choice

For each item and each lexis the two orders are shown together. A/B assignment is randomised per (item, lexis, seat), balanced inside each front-type stratum, from the same frozen seeds (P1-7, P3-14) — so every item is seen in both orientations across the three seats and a position effect is estimable. The prompt no longer announces what differs between the versions, which primed the dimension under test (P3-12), and it asks the same within-idiom question as stage S.

Below are two versions, A and B, of the same two lines of English verse. Both are written in the same kind of English, so which kind it is cannot be the answer.

A: {A}

B: {B}

Which of the two is the better-formed, more idiomatic sentence of the kind of English it is written in? Ignore beauty, meaning and accuracy.

Reply with exactly two lines and nothing else: CHOICE: REASON:

• The archaic factor is morphology, not Payne's register. After the critic, the map is purely substitutive — hath, thy, thou/thee and the inflections they force — so "archaic register" and "the -th form of the final verb" are the same manipulation here and cannot be separated (P3-6). A null therefore refutes the morphology-carries-it version of §7.41.3 and not the whole sentence; a hold does not establish that Payne's diction, metre or verse convention add anything further.
• Tier D is NOT PASSED. Nothing here is a claim about what a human reader perceives, prefers or wants. Three model seats rating English well-formedness is internal-judgment-only.
• The naturalness sense is not measured. The DV is passage-indexed by necessity (§4), so no figure here is in this project's naturalness units and none may be compared to one.
• One hand, and the order factor is still hand-written. The mechanical map removes the translator from the lexis factor entirely; it does not remove him from the order factor, where the multiset check constrains the words but not the felicity of the arrangement. An inverted line the lead could not write well is indistinguishable from an inverted line English does not permit (P3-4, partly unfixed and named). The design removes the between-hand confound that §11.2 named and replaces it with a within-hand one: an inverted line the lead could not write well is indistinguishable from an inverted line English does not permit.
• No chime. R56 §5 refuses the rhyme, so this prices the alignment of the tail and not the full device (D7 in the translator's log).
• The D arms are not the radif dropped. They keep the word and lose its position (D9). What is priced is alignment, not repetition, and the result page may not widen it.
• Two poems, one radif, one poet, one language pair.

9. Pre-flight cost estimate

Worst case is built from the caps, not from expected output (note (abc)).

UTC day 2026-08-29 stands at $0.197706300 of $5.00 after S231, so $4.802293700 is available and a $2.20 ceiling fits with room. The opening key snapshot for this session reads 156.738407873, which is $0.872971860 above S231's close — non-project spend on the key between sessions, recorded in config/budget.md and not this project's. Key usage is snapshotted before and after and the delta cross-checked against the per-request sum.

10. Verification

analyse.py --mutate recomputes every number reported on the result page from raw/, and plants deliberate faults to confirm the checks can fail. Nothing is reported that the verifier does not recompute.