Repository path: workshop/experiments/E-20260805-heredia-rhyme-flavour/design.md · rendered 2026-09-09

E-20260805 — one trade-off, two resolutions: testing the 1897 Nation's split verdict on Egan and Taylor

Frozen before any call was dispatched. The translation limb (T-recif-de-corail-R16-v1, -R17-v1, T-trebbia-R16-v1, -R17-v1) and its logs were frozen and committed at d2cfb47, before this file existed.

1. Question

The Nation, 16 December 1897, p. 481, reviews two simultaneous American translations of Heredia and splits its verdict by dimension:

Of the two translations, Prof. Egan's is the superior, as showing the more trained poetic touch … On the other hand, Mr. Egan refines more upon the vigorous and daring dialect of the poet, and therefore gives us a little less of his flavor. Each is responsible for some feeble or lax lines, and in respect of rhymes Mr. Taylor is frequently faulty…

Are these two complaints one trade-off with two resolutions? The hypothesis this experiment tests is that a rhymed-verse translator of Heredia must choose, at the line ends, between the rhyme and the poet's specific vocabulary; that Taylor resolves it toward the vocabulary and pays in rhyme; and that Egan resolves it toward the rhyme and pays in vocabulary — which is exactly what the reviewer describes without connecting the two halves.

The subject-rule sentence (continue-prompt.md §4.5), written before the unit was designed: this unit teaches what a rhymed verse translation costs and where the cost falls, by measuring two 1897 translators' opposite resolutions of the same forced choice against the French they both worked from. The object of study is three poets' English and one French poet's line ends.

2. Materials — frozen, and the F3 gate discharged first

Sample: six sonnets, fixed before any coding by a neutral rule — the first six, in Heredia's own book order, of the ten sonnets Egan translated (Egan's ten are the whole of the comparable set, because Taylor translated 118 and Egan only these). They are, by Lemerre page: Suivant Pétrarque (96), Sur le Livre des Amours de Pierre de Ronsard (97), Épitaphe (99), Les Conquérants (111), Midi. L'air brûle… (121), Le Samouraï (126). 84 line-end positions.

Why six and not ten. Every line in the sample must be verified against the page pixels — the arm's own binding rule (ARM-reception-claims §Constraints, F3: reporting a difference between an OCR'd translation and its source without going to the pixels is reporting the scanner). Six sonnets is what that verification covers. The restriction was fixed before any measurement.

The F3 gate, run and discharged before this design was written. Every one of the 84×2 English lines was cross-checked against a second independent scan (Taylor: sonnetsjosmaria00compgoog; Egan: libraryofworldsb13warnuoft), and all six Taylor pages and the two disputed Egan pages were read from the page images. Two substantive OCR errors were found and corrected — Taylor Conquerors l. 12 read then- white carvels for their white carvels, and Epitaph l. 5 Quelus for Quelús — plus five diacritics in Egan (bîva, Hélène, lovèd, antennæ). Had the gate not been run, one of the two would have entered the measurement as a Taylor error that does not exist.

3. Design

Independent variable — specificity of the French line-final content word. 84 positions. Coded by three seats shown the French sonnets and nothing else: SPECIFIC (a technical, local, material or proper term that a general English vocabulary does not supply — the reviewer's "technical and local vocabulary") vs GENERAL. Majority of three governs.

Dependent variable 1 — flavour retention, per arm. For each of the 84 positions, three seats are shown the French sonnet and one English sonnet, unlabelled, and asked whether that English rendering carries the French word's specific sense anywhere in the poem: PRESERVED / GENERALIZED / DROPPED. Line-by-line alignment is not assumed — in verse a word migrates — so the question is asked of the whole sonnet. Seats never see the two arms together and are never told that two translators exist.

Dependent variable 2 — rhyme fault, per arm. Mechanical. Each English sonnet's rhyme groups are enumerated and each group coded PERFECT (all members identical from the stressed vowel onward, in one stated accent) or IMPERFECT, by the lead under a rule written into analysis/rhyme_rule.md before coding, and independently re-coded by one seat shown the groups with arms unlabelled.

4. Predictions, registered

• PR1 — the reviewer's own comparative claim. Egan's loss rate (GENERALIZED + DROPPED) exceeds Taylor's across the 84 positions.
• PR2 — Taylor's IMPERFECT rhyme groups are enriched at groups whose French counterparts are majority-SPECIFIC, relative to majority-GENERAL groups.
• PR3 — Egan's loss positions are enriched at SPECIFIC French line ends.
• PR4 — the lead's frozen log speaking. T-recif-de-corail-R16-v1 D10 and T-trebbia-R16-v1 D11 predict that rhyme-driven loss concentrates in the octave (ll. 1–8) rather than the sestet (ll. 9–14). The Trebbia log already records this prediction failing on its own poem and leaves it as written. It is tested here on both arms.

5. Controls and failure criteria, registered

• FC1 — if a 2-of-3 majority on the French specificity coding is reached on fewer than 70% of the 84 positions, PR2 and PR3 are WITHHELD: the independent variable has no usable signal.
• FC2 — if either arm's loss rate is below 10% or above 90%, PR1 is WITHHELD for lack of range.
• FC3 — if the lead's rhyme coding and the independent seat's disagree on more than 25% of rhyme groups, PR2 is WITHHELD.
• FC4 — positive control (canary). A seeded arm is run alongside: Taylor's Le Samouraï with four line-end renderings replaced by generic hypernyms, the four positions registered in materials/canary.json before dispatch. If fewer than 3 of the 4 are coded GENERALIZED or DROPPED by majority, PR1 and PR3 are WITHHELD — the instrument cannot see what it is for.
• FC5 — blindness. Arms are labelled only A/B, order permuted per sonnet by sha256(design-id|sonnet) mod 2. If any seat names a translator anywhere in its output, that seat's cells are discarded and the fact is reported. S108 (RS-20260804i) measured that the blind holds at small grain and fails on whole loci; a whole sonnet is a locus, so this control is not decorative and its result is reported whichever way it goes.

6. Panel and procedure

Seats from config/models.md panel v1: P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, P5 deepseek/deepseek-v4-pro. Coders, never jurors: nothing in this experiment judges the quality of a translation, and Tier D is NOT PASSED, so every sentence it produces is provisional and internal-judgment-only.

1. Stage 0 — seat probe. One minimal call per slug before anything else, to confirm the slug answers and returns a body. Note (bhf) has asked for this since S087 and S108 lost 60% of its spend for want of it. A slug that fails the probe is replaced from config/models.md before the run, not during it.
2. Stage 1 — French specificity. 3 seats × 2 calls (three sonnets each) = 6 calls.
3. Stage 2 — flavour retention. 3 seats × 4 calls (three (sonnet, arm) items each; 12 items from six sonnets × two arms, plus the seeded canary sonnet as a 13th item) = 12 calls.
4. Stage 3 — rhyme re-coding. 1 seat × 1 call = 1 call.
5. Raw request/response JSON preserved for every call; every reported number recomputed by analysis/verify.py, which imports nothing from the runner.

Judgment is never parallelized. Calls are dispatched one at a time.

7. Pre-flight budget estimate

19 calls. Worst case built from max_tokens, not from an assumed output length (note (abc)): max_tokens 4,000 per call, at the most expensive seat's completion price ($15.00/M, P4 rates as a routing-margin ceiling) → $0.06 per call worst case, 19 calls → $1.14, plus one pre-run critic call at max_tokens 24,000 → $0.36. Declared worst case $1.50. UTC day 2026-08-05 opens with the full $5.00.

8. What this design cannot establish

• Nothing about which translation is better. No sense is scored, no jury sits, Tier D is NOT PASSED.
• Nothing about Heredia's other 112 sonnets, or about Taylor's renderings of them.
• Nothing about whether the reviewer was right to prefer Egan: "trained poetic touch" is not a located claim and this design does not pretend to test it.
• The SPECIFIC/GENERAL coding is a machine reading of a French line end by uncalibrated seats. It is an operationalisation of the reviewer's phrase, not a measurement of what he meant.