Repository path: workshop/experiments/E-20260806d-first-span-again/design.md · rendered 2026-09-09

E-20260806d — what has the first span learned?

ARM-atelier-cycle step 8, the arm's closing step. Frozen 2026-08-06 before any word of the retranslation was written and before any span-1 English was read. Every prediction in §4 was written against the Finnish and the closed register alone.

1. The question

«Köyhää kansaa» is translated whole — 20,276 English words from 13,517 Finnish, seven spans, S085 to S115. The arm's own question is what a long work teaches that short pieces cannot, and its closing step owes a craft report. A craft report can be written as recollection. This design tries to make one number of it instead.

Span 1 was translated first, and it was translated by a translator who did not yet have: the copy-text (V21, span 4); eighteen of the register's twenty-six live rules; the whole-work censuses of te/sinä, of the zero person, of the divine capital, of the honorific-to-a-parent; the completed collation; or the book's ending. Span 7's translator had all of it.

So: put span 1 back through the finished apparatus and see what moves. If the length taught the translator things, the first span is where the teaching should show — and if it does not show, the craft report has to say that the learning lived in the register and the collation and not in the prose.

2. The design in one paragraph

Re-render ¶2–75 whole, from the copy-text, under the closed register, blind to R05-v1's span-1 English. Freeze it. Then diff v1 against v2 mechanically, code every divergence by cause, and put a stratified sample of the paragraph pairs to three blind seats who are asked one question — do these two English passages differ in what they say, or only in how they say it? The seats never see the Finnish, never see which is which, and never rate quality.

3. Materials, and what the translator was and was not shown

The blindness is partial by construction, and the direction matters. The register quotes v1's own lexis — right enough, not a rap, mother dear, waah, a body, and every row of the terminology table. That is what a register is: the carrier of the earlier spans' decisions into the later ones. So v2 is blind to v1's sentences and not to v1's rules, which biases the experiment towards agreement at exactly the sites §4 predicts agreement. A prediction of no divergence at a rule-governed site is therefore cheap; a prediction of divergence is not. §4 registers three of each.

4. Predictions, registered before translating

4.1 Five named sites

Written against the Finnish and the register, with no span-1 English in view. Three predict a null.

4.2 The count

Every divergence is coded to exactly one cause, by the precedence (E) > (A) > (C) > (B) > (D):

• (A) COPY-TEXT — the 1886 reading differs from the 1917 one v1 worked from, and it changes the English.
• (B) LATER RULE — a register rule, terminology row or erratum created after span 1 determines a reading v1 does not have.
• (C) WHOLE-WORK — the change depends on something outside span 1: a recurrence, a census, a motif, the ending.
• (D) FREE VARIATION — none of the above. Two defensible renderings of the same thing.
• (E) CORRECTION — v1 is wrong against the Finnish and v2 is right. Every (E) is listed with the Finnish so a reader can check it.

P1 — the attributable count. (A)+(B)+(C)+(E) together come to ≤ 12 sites. Point expectation 8.

P2 — the shape. (D) is the largest class and is at least three times the attributable count.

P3 — corrections. (E) is ≥ 1 and ≤ 5. Point expectation 2.

P4 — rule discipline. At sites governed by a rule that already existed when span 1 was written — V2 (mies/vaimo), V4 (äiti/Mari), V5 (paragraph boundaries), V6 (double quotes), V7 (yhyy = waah), V9 (vaivainen = cripple) — v2 reproduces v1 at ≥ 0.85, and the paragraph count is 74 to 74 with no split or merge (V5).

What P1 and P2 are for. The arm has claimed for seven visits that the length was making things visible. P1 and P2 say what that is worth in the prose: if the attributable count comes in at 8 against a free-variation count of 40, then the honest sentence in the craft report is the register carried the learning and the rendering barely moved, and the arm should say so rather than tell a story about accumulated mastery. A high attributable count would be the more flattering result and is the one predicted against.

4.3 Divergence sites, defined mechanically

Align v1 and v2 paragraph by paragraph, then word by word (difflib.SequenceMatcher over whitespace-split tokens, case- and punctuation-preserving). A divergence site is a maximal run of non-matching tokens, merged with the next when the two are separated by ≤ 2 matching tokens. Sites differing only in punctuation or only in capitalization are recorded and excluded from the coded set. If the coded set exceeds 200 sites, the site-level counts are reported as descriptive only and every claim moves to the paragraph level (§4.4).

4.4 The paragraph is the unit the seats see

Each of the 74 paragraphs takes the highest-precedence class of any site inside it, or SAME if v1 and v2 are byte-identical.

5. The blind coding run — is the translator's causal coding recoverable by anyone else?

The threat this run exists to answer. Every count in §4.2 is coded by the translator, about the translator's own two renderings. RS-20260806c closed on exactly this limit — "both censuses are one locus each and coded by the translator". The (D) class is where it bites: (D) asserts that nothing changed except the wording, and that is checkable by someone who has never seen the Finnish.

The instrument. Three seats, each shown pairs of English passages labelled A and B in a per-item randomized order, are asked for one letter per item:

• F — A and B say the same things; they differ only in wording.
• S — A and B differ in what they say: some fact, act, relation or attribution is different.

The Finnish is deliberately withheld. The seats are not being asked which rendering is right — they are being asked whether the two make the same assertions, which needs no Finnish and does not depend on a panel Finnish competence this project has never probed (config/models.md, S015 note: Russian, French and Japanese only).

Strata, drawn after the coding and before dispatch:

Registered tests.

• RT1 — among D-stratum items, the majority-of-seats verdict is F at ≥ 0.70.
• RT2 — among EA-stratum items, the majority verdict is S at ≥ 0.70.

Failure criteria, frozen before dispatch.

• FC1 (REPEAT) — if a majority of seats returns S on any REPEAT item, the instrument fires where nothing differs and the whole coding run reports nothing.
• FC2 (WRONG) — if a majority returns S on fewer than 3 of 4 WRONG items, the instrument cannot see a proposition change and the whole coding run reports nothing.
• FC3 (power) — if the D stratum has fewer than 8 items, RT1 is not reported.
• FC4 (EA power) — if the EA stratum has fewer than 3 items, RT2 is not reported and the §4.2 coding stands as translator-coded only, declared in those words.
• FC5 (seat) — a body with fewer answer lines than the block has items is a seat failure, enforced in the runner, never repaired by reading around it.

What each outcome licenses. RT1 and RT2 both holding licenses one sentence and no more: the boundary the translator drew between "changed what it says" and "changed only how it says it" is recoverable by readers who never saw the Finnish and did not know which rendering was which. It licenses nothing about whether either rendering is good. No sense is scored and no translation is judged; Tier D is NOT PASSED and every sentence of the result is provisional.

RT1 failing is the interesting failure and is not a defect of the run: it would mean the retranslation changed what the text says at sites its own translator believed were stylistic, and the craft report would have to lead with that.

6. Panel, cost, and the pre-run critic

Two blocks per seat, max_tokens 4,000; critic one call, max_tokens 24,000.

Pre-flight worst case, built from max_tokens and not from an expected length — note (abc). Critic 24,000 out × $6/M + ~12k in × $2/M = $0.168, ×1.5 routing margin = $0.25. Seats 2 × 4,000 out each: P1 $0.048, P2 $0.060, P5 $0.007, plus input ≈ $0.035; ×1.5 routing margin and ×2 for the runner's two-attempt cap = $0.58. Declared worst case $0.90, against a UTC-day headroom of $3.678 at the time of freezing. Provider read off every response; per-request usage.cost primary, key-usage delta as the cross-check, read after a pause — note (bil).

7. What this design cannot do

1. v2 is not independent of v1. The same lead, the same register, the same book. It is a reviser's pass, and it is labelled one. Nothing here measures what a different translator would have done.
2. The (E) class is the translator judging its own earlier work. Every (E) is published with the Finnish beside it; the blind run tests only that (E) sites differ in content, not that v2 is the correct one.
3. The register's purchase is bounded by V17, which forbade writing a rule for an unread site. A register built span-by-span is expected to be local; this design measures how local, on one span, of one work, in one language pair.
4. No sense is scored, no jury sits, nothing is judged. Tier D NOT PASSED.