Repository path: workshop/experiments/E-20260826b-radif/design.md · rendered 2026-09-09

E-20260826b — a repetition and a chime in one slot, taken apart

Design v2, frozen before dispatch. v1 was frozen, sent to two adversarial critic seats together with the exact materials, and rebuilt on their findings: fifteen findings, seven BLOCKING, all accepted, none overruled (critic-response.md; v1 is in git history). The translation limb and the four line-end arms were frozen and committed at 74cc7c83, before v1 existed. No seat has judged anything at the time this is frozen.

1. The question

wiki/goodness-senses.md §style-correspondence defines itself over marked formal features and names repetition among them. Every measured statement on the entry is about chime: device presence (+3.524, S142) and device extent (+0.321 / +0.214, RS-20260824c-run-placement, S219). Nothing on the entry measures repetition, and the run that came nearest (RS-20260821b-matched-heard, S209) established that at a matched shape the subtraction method is not available, closing with the sentence this design answers: what remains available is varying which resource carries the match with the position held fixed, which no run has yet built.

The Persian ردیف is where that is buildable: two resources in one slot on every line — a phrase repeated identically at every rhyming position, and a chime on the word immediately before it.

Which of the two is a reader responding to, and is the pair worth more than the two halves added up?

2. Materials, and what the manipulation actually is

workshop/translations/saadi-ghazals-radif/R51-v1/arms.md v2. Seven loci, four arms each, word-identical within a locus except at the line-ends (one declared mid-line normalisation at L8, applied to all four arms).

L1–L4 are depth 3, L6–L8 depth 2. L5 was withdrawn before any judging: at a comparative radif every English bearer must be a comparative, every English comparative ends in the same unstressed syllable, and the language will not permit a no-chime arm. That is a result, recorded on arms.md.

The confirmatory set is L1, L2, L4, L6, L7 — the five loci where the hand and tools/rhyme_pairs.py agree that the AB bearers chime (≥ 1 STRICT pair). L3 and L8 are a labelled sensitivity set: their AB bearers are blame/claim and lane/slain, full rhymes to any English poet, rejected by the tool's rime-riche guard because the phoneme before the rime is L in both members. Every quantity is reported on the confirmatory five and again on all seven.

What the two factors are NOT. They are not isolated phonological changes. Removing a chime means changing a word, and un-repeating a phrase means changing words; the substitutes differ in shading, and a varied tail is longer than is there or arose however carefully chosen. E_rep and E_chime estimate preference between constructed line-end packages; repetition and chime name the intended manipulations, not isolated causes. This is the critic's own first remedy for its BLOCKING 2, taken in full, and it governs every sentence written from this run.

3. Seats

Three, all non-Anthropic: P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, QR qwen/qwen3.7-max. P4, P5 are out on any task shape (notes (bps), (bne)); P3 is a cost problem; GL is out on long prompts. max_tokens 3000 on every seat from the start, per note (brr).

4. Procedure

Blind forced choice, one pair per call, temperature 0. Each call shows two versions of the same passage, labelled A and B, and asks which reads better as English verse, with one sentence of reason. No call mentions Persian, Sa'di, rhyme, repetition, line-endings, translation, or that anything was manipulated.

Four pairs per locus — the edges of the 2×2: AB–RH · AB–RD · RH–NN · RD–NN. Both orders at every seat: 7 × 4 × 2 × 3 = 168 judging calls.

5. The statistics, registered

For each locus and each pair let w be the fraction of the six judgments (3 seats × 2 orders) preferring the first-named arm.

The unit of analysis is the locus. Seats and orders are pooled inside a locus and are never counted as independent observations (RS-20260809i BLOCKING 3).

Primary reporting is the per-locus effect and its dispersion — median, range, interquartile range over loci — in the shape RS-20260824c used, because five loci do not support an interval estimate. The exact one-sided sign test is secondary and is a descriptive consistency summary over constructed loci, not evidence generalising across texts or readers: the seven loci come from three ghazals and share radifs. Its floor on the confirmatory five is P = 0.031.

Tie rule, registered: a locus whose effect is exactly 0 counts as not supporting the one-sided prediction and stays in the denominator.

Registered predictions, one-sided:

• P1 — E_rep > 0. The entry's gap.
• P2 — E_chime > 0. A replication check on the chime package, not a validity gate for P1 (critic finding 8).
• P3 — E_int > 0. The Persian form's own claim. Declared underpowered and reported with its dispersion only.

Depth-2 and depth-3 loci are reported separately for every quantity (critic finding 12).

6. Failure and withholding criteria, registered before dispatch

• F1 — P1's interpretability rests on C1's per-edge gate (§7), not on P2. If C1 fails its planted-error gate, all three quantities are reported descriptively only and nothing is written into wiki/goodness-senses.md.
• F3 — an edge that C1 marks DIFFERENT at a locus is dropped from the quantity it feeds at that locus, and the locus leaves that quantity's denominator. n is reported per quantity.
• F4 — order, per edge. For each of the four edges, the first-position win rate is computed over its 42 judgments. If any edge falls outside [0.35, 0.65], every quantity that edge feeds is recomputed within each order and both reported; if the two orders disagree in the sign of a quantity's median, that quantity is withheld.
• F5 — parse. An empty or unparseable body is re-dispatched once; a second failure drops that judgment, reduces the denominator, and is reported as a count.
• F6 — the arms must be what they claim. tools/rhyme_pairs.py over the declared bearer lists must return, for every live locus, ≥ 1 relating pair in AB and 0 relating pairs in RD. Run before this design was frozen: passes at 7 of 7 (grading-arms-v2.json). Registered here so a later reader can see it was a gate.
• Saturation is a diagnostic, not a withholding rule (critic findings 4 and 13). If a pair returns the same preference in ≥ 95% of judgments the fact is reported beside the number.

7. Control C1 — a per-edge equivalence gate

The substitutions are defensible renderings of the same Persian but are not identical in content, and at six of seven loci the RD bearer is the more literal. A seat preferring RD might be preferring accuracy.

Procedure. Seat P2, one item per call, temperature 0, shown the two texts of one edge and asked: Do these two passages assert the same thing? Answer SAME or DIFFERENT. If DIFFERENT, say in one clause what the second asserts that the first does not, or what it contradicts. The prompt states explicitly that a difference of wording that is not a difference of assertion is SAME (critic finding 15).

Items: the four edges × seven loci = 28 real items, plus 6 planted items in which one side has been altered to change a number, change a referent, or drop a clause outright. 34 calls.

Gate. DIFFERENT on ≥ 5 of the 6 planted items. Below that the control is uninformative, F1 fires, and that is reported rather than papered over.

If C1 fails its gate the judging stage is not dispatched at all.

8. Pre-flight cost

Worst case from max_tokens, not from expected output (note (abc)).

Headroom before this run: $4.783086950 of $5.00 (opening snapshot 144.764653501). Staged so it can be stopped: critic → C1 → judging.

9. What this cannot establish, written before any number exists

1. Nothing about hearing, and nothing spontaneous. This measures explicit comparative preference after a salient line-end substitution. Side-by-side presentation makes the changed words conspicuous however the prompt is worded. No sentence citing this may say a reader hears a repetition or responds to one unprompted.
2. The factors are packages, not isolated causes (§2). A varied tail is longer and heavier than an identical one; a de-chimed bearer is a different word with different shading.
3. Three poems. Seven loci drawn from three ghazals, two of them sharing a radif with another locus. This is not seven independent texts and the sign test is not a general inference.
4. n = 5 confirmatory. The one-sided exact test cannot go below P = 0.031, so this run can support a direction, never a strong claim.
5. Three model seats, no human reader. Tier D is NOT PASSED; internal-judgment-only, provisional.
6. E_int cannot establish the Persian form's claim at this n and is not reported as if it could.