Repository path: workshop/experiments/E-20260808a-set-forms/design.md · rendered 2026-09-09

E-20260808a — do the fixed forms of a language mark where two hands part?

ARM-two-hands step 2, translation limb. Frozen before any API call. The renderings and the locus classification were frozen first, at b583be6 and ea3531e.

Question

RS-20260807c-two-hands §4 found four loci the lead had labelled plain and at which two independent hands nevertheless parted, and three of the four contained a set form the lead had not noticed when it labelled them. That is n = 3, post hoc, on one chapter.

The transportable claim, and the one under test: set-formhood predicts where two independent hands part, even when nothing about the stretch looks difficult. A translator scanning a source for hard places looks for the class this design calls MARK — puns, culture-bound objects, code-switches, things English visibly has no slot for. If SET — archaic morphology, frozen idiom, proverb, fossilised politeness formula, Latin tag, reduplicative, frequentative — behaves like MARK rather than like PLAIN, then the advance list every translator makes is systematically short, and short in a nameable way.

What this teaches about translating literature (subject rule): it names a class of place where translations of the same source will differ, and it is a class defined on the source language's inventory of fixed forms rather than on the translator's sense of difficulty.

Materials

• Source. Mikszáth Kálmán, «Szent Péter esernyője» (1895), Part I ch. 2 «Glogova régen», whole, 30 paragraphs, 726 Hungarian words (span-2-source.txt, MEK-00954).
• LEAD — T-szent-peter-esernyoje-R04-span2-v1, whole chapter, 995 words. Frozen ea3531e.
• WORS — B. W. Worswick, 1900, ch. II (Project Gutenberg #31945, public domain), 757 words. Worswick's chapter renders Hungarian ¶1–¶22 only; ¶23–¶30 have no counterpart in his text.
• M1, M2 — two unbriefed panel renderings of the whole chapter, generated for this run, P5 and P4 (config/models.md). Each is told only that it is translating Hungarian literary prose of 1895 into English; no regime, no locus list, no sight of any other rendering.
• Loci. 37, classified by the rule frozen in materials/loci.md: 8 MARK, 19 SET, 10 PLAIN, PLAIN drawn with no selection at all (first finite clause of every paragraph containing neither other class).

Contamination, measured before the design was written (materials/contamination.json): LEAD × WORS over ¶1–22 gives 12 shared 7-grams, 0 twelve-grams, longest run 11 tokens, against 160 / 27 / 24 for an independent published pair. Clean. (Span 1 of the same work and the same comparator: 2 / 0 / 8.)

Pairs

The frozen design's blocks A and C were WORS × M1 and LEAD × WORS. Both are gone from the inferential structure; the human comparator survives as block D and as the D1 census.

Procedure

1. Generate M1 and M2, unbriefed, one rendering each, temperature 1.0.
2. Three seats — P1, P2, P3 — none of which produced any rendering in this run. Each seat receives each block as one call: both full English versions, labelled only Version A and Version B, side order randomised per seat per block, authorship never stated.
3. For each locus the seat is given the Hungarian stretch and its paragraph number, and returns: verdict ∈ {same, different}; quote_a and quote_b — the stretch it aligned in each version; and where different, direction ∈ {both_carry, a_drops, b_drops}.
4. Judgment is never parallelised across loci within a seat; each block is one call.

Predictions, registered

Amendment A3 binds all of these: the inferential unit is the LOCUS, and a locus's verdict is the majority of the three seats. Seats reduce measurement noise; they are not replicates.

• P1 — PRIMARY. On block A, over the 49 loci, on seat-majority verdicts: SET − PLAIN ≥ 0.25 in the proportion called different. The threshold is transported verbatim from E-20260807c P3 and is not chosen here (see §Limits). Reported separately against PLAIN-A and PLAIN-B; if the two PLAIN subsets disagree, the primary is read against PLAIN-B alone, which is the within-paragraph control.
• P2 — replication. The same criterion on block B, the independent second draw.
• P3 — permutation. Exact one-sided permutation over the 49 frozen class labels of block A on the statistic SET − PLAIN, enumerated or Monte-Carlo'd to a stated precision. P < 0.05.
• P4 — secondary, descriptive. SET against MARK. No threshold is registered, because any number chosen here would have been chosen after the comparator was read. Reported as a difference with its seat-level spread and nothing inferred from its size.
• P5 — control on the instrument. MARK − PLAIN ≥ 0.25 at ≥ 2 of 3 seats on block A. If P5 fails, P1 is uninterpretable rather than confirmed or refuted: a design that cannot see hands part at a pun cannot be trusted when it says they part at an idiom.

Failure criteria, registered

• F1 REPEAT. Four loci in block C show the same rendering on both sides. A seat fires if it calls ≥ 2 of its 4 different; if ≥ 2 of 3 seats fire, the primary is withheld. (Amendment A6: the frozen bar was any of the four, which the critic computed would withhold the primary ~9% of the time against a perfect instrument.)
• F2 WRONG. Four loci in block C carry a deliberately damaged side (intents frozen in loci.md). Detection means different and the damaged side named as the one that drops something. Below 9 of 12 across seats, the primary is withheld (amendment A6).
• F3 quote verification. Every judgment must quote the stretch it aligned in each version; quotes are checked mechanically against that version. Failing judgments are dropped from every rate. A seat below 0.60 verification on a block has that block withheld.
• A withheld primary is not re-derived from a weakened bar after the fact.

Limits, declared before running

1. The lead read WORS before these thresholds were written. The loci and their classes were frozen blind (ea3531e, before the comparator was opened), and P1's threshold is transported verbatim from the previous experiment rather than chosen here — but the choice to restrict the primary to ¶1–22 was made after reading Worswick, because his chapter stops there. This is the run's principal bias exposure and it is stated here, not in a footnote.
2. P4 carries no threshold for the same reason.
3. One chapter, one work, one language pair, one human hand. Worswick is not a sample of human translators, and M1/M2 are not a sample of translators at all.
4. Class sizes are unequal and small — 8 MARK, 19 SET, 10 PLAIN, of which 8 MARK, 11 SET, 7 PLAIN fall inside Worswick's coverage.
5. Tier D is NOT PASSED. Nothing here judges quality, ranks a translation, or scores a sense. Different is not worse.
6. The direction field inherits RS-20260807c §6's confound whole: all seats are 2026 models, two of the four hands are 2026 models, and a shared prior favouring machine prose is a live alternative to any reading of who is said to drop things. Reported descriptively only.

Descriptive census, no prediction attached

D1 — what Worswick has no text for. Of the 37 loci, which have no counterpart in his chapter at all. This is computed from his text and needs no seat. It is reported because his chapter's omission of ¶23–¶30 was discovered in the course of this run and is a fact about the translation, not about the instrument.

Cost, pre-flight

Worst case built from max_tokens, per note (abc).

Headroom at declaration: $5.00 (UTC day 2026-08-08 opens with no rows); $4.974405 at the moment A8 raised the ceiling, the critic call having already been paid.