Repository path: workshop/experiments/E-20260805g-printed-switch/design.md · rendered 2026-09-09

E-20260805g — the printed switch: how much of the exclusion is the foreign line carrying?

Frozen 2026-08-05 before any seat dispatch. The translation it is about was frozen first, at a3b9a0b, with its log (charter A4, R05 rule).

This is revision 3. Revisions 1 and 2 each went to an independent adversarial critic before any seat call and each came back NEEDS-REDESIGN — nine findings then thirteen, six then nine BLOCKING, all twenty-two accepted. critic.md records every one and what it changed. The question this page asks is not the question revision 1 asked, and the reason is critic pass 2's F4: the scene contains a route to the same conclusion that does not go through the foreign line at all, and until that route was made into a measured baseline the run could not attribute anything to the Swedish.

1. Question

Canth prints one line of Swedish inside her Finnish at ¶449 — «Kan hon botas?», can she be cured — spoken by the pastor to the doctor over a woman who has just been tied hand and foot on the floor of her own room. She does not gloss it, does not narrate it, and does not mark it typographically. D115 kept it bare and declared a cost: the device survives, and the reader's seat inverts — Canth's Finnish reader could read the line and was placed with the gentlemen; an English reader cannot and is placed with Mari.

But the paragraph is not alone. In the same fifteen paragraphs the doctor avoids Holpainen's questioning eyes and looks out of the window; the pastor's question gets no answer at all; the whole room falls silent while the doctor writes; the pastor says there is nothing more for us to do here; and the landlord has to break in and re-ask, in the language of the book, whether there is any hope. A reader can arrive at "the gentlemen have something between them that the household is outside of" without ever noticing a language.

So the question is quantitative, not binary: how much of this scene's exclusion is the foreign line carrying, over and above what Canth has already built round it? That is a question about what a translator's decision at one line is actually worth, and it is answerable because the counterfactual — the same fifteen paragraphs with the line in English — can be put to the same readers.

2. What this run is NOT

No seat is asked whether anything is well or badly translated, and no quality claim is derived from any body. This is a comprehension probe. The lead never judges its own translation (charter §5). The registered predictions are expectations about readers, frozen before dispatch.

3. Materials

One extract, three arms, differing in exactly one line. Span 7 of T-koyhaa-kansaa-R05-v1, ¶442–456, 15 paragraphs, 253 English words — the doctor entering to Heikura's "They'll not take to listening to that." ¶457, which narrates a second switch, is outside the extract: with it in, every arm would recover the switch from the narration.

Orientation identical across arms; it names no language, mentions no exclusion, and — after critic F8/F10 — no longer calls the pastor "parish", the doctor "town", or Mari "poor".

4. Seats

Five seats × three arms = fifteen independent stateless dispatches, each showing one arm. Seats: P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, P5 deepseek/deepseek-v4-pro, qwen/qwen3.7-max, mistralai/mistral-medium-3-5. z-ai/glm-5.2 and moonshotai/kimi-k3 excluded by note (bhf). Critic x-ai/grok-4.5, not a seat.

5. Procedure — one uncued task, and nothing else in the payload

Critic pass 3's F4 is accepted and it is why this section is short. Revision 3 put an "uncued" retelling in Part 1 and a cued exclusion block in Part 2 of the same payload, and told the model not to revise Part 1. A single-pass model reads the whole prompt before it generates anything, so Part 2's definition of exclusion cued Part 1 however the instruction was worded, and the frozen EXCL lemmas mirrored Part 2's own vocabulary so that keyword scoring would have rewarded prompt-echo.

Part 2 is deleted. The payload is the orientation, the extract, and one instruction: retell the passage in four to six sentences for a reader who has not seen it. The body is RETELL: and END. No question in the payload mentions language, following, understanding, exclusion, or authority. Two predictions of revision 3 (PR-L, the leak audit) and revision 2 (INTENT, RANK) die with it, and neither was load-bearing: critic pass 2's F6 had already shown PR-L true by construction.

6. Scoring — frozen before dispatch, recomputed by analysis/verify.py

Three independent codes over the whole RETELL line, case- and quote-normalised. The split between the first two is critic pass 3's F2, accepted: revision 3 would have scored a seat that simply could not parse the quoted string as having recovered a social exclusion, and those are not the same thing at all.

• OPACITY — the retelling reports something on the page as foreign, untranslated or unintelligible, without naming anyone in the room as unable to follow it. This is the seat's own parsing, not the scene's.
• EXCL — the retelling asserts that a named person or group who is in the room could not follow, was kept from, or was outside what was said. Requires both an exclusion predicate (could not / couldn't / cannot / did not / unable to + understand / follow; shut out; exclud; kept from; over the heads; privately; in private; aside; only the doctor; only the two; between themselves; among themselves; no one else; the others could not; side conversation; left out; not for the others) and a person in the room (holpainen, husband, mari, tiina katri, heikura, family, household, the others, everyone else, no one else, those present). Co-mention of a name with a language earns nothing (critic pass 2 F2).
• VIA-LANG — EXCL and a language marker anywhere: swedish, another language, a different language, foreign language, foreign tongue, in a tongue, a language, not in finnish, untranslated, kan hon botas, code-switch, scandinavian.

Every retelling is quoted verbatim in the result page. The coding is auditable against the bodies, and verify.py carries non-keyword acceptance tests (§11) so the list's brittleness is measured rather than assumed away.

7. What is reported — descriptive counts, and NO confirmatory test

Critic pass 3's F1 and F7 are accepted, and the run's confirmatory ambition is withdrawn before dispatch rather than after seeing the numbers. At n = 5 per arm, with a baseline the design itself expects to be non-zero, the only tables that could have cleared a Fisher gate were (5,0), (5,1) and (4,0); a design whose rejection region is nearly unreachable is pre-rigged, and running it as a test would have dressed a descriptive count as an inference. Revision 3's PR-PRIMARY and its rejection region are deleted. No p-value is computed anywhere in this run.

What is reported, as counts out of five, with all fifteen retellings quoted:

• B — the baseline, and the number the run exists to get. EXCL in A2. A2 has no foreign line at all, so any exclusion recovered there is carried by what Canth built round ¶449: the doctor avoiding Holpainen's questioning eyes, the pastor's question getting no answer, the room silent while the doctor writes, the pastor's "nothing more for us to do here", and Heikura having to break in and re-ask about a cure in the language of the book. B is interpretable whatever it is, and it is not a test of anything.
• D — the difference. EXCL in A1 minus EXCL in A2, reported as two counts and never as an effect. If B is high, the decision at ¶449 is worth less than D115 supposed, whatever A1 does.
• R — the route. VIA-LANG among EXCL bodies, by arm. A reader who reports the exclusion without naming a language recovered it from the scene and not from the translator's choice.
• O — the parsing code. OPACITY by arm, reported separately and never added to EXCL.
• L — the label. A3's counts. A3 is a two-word insertion and nothing causal is claimed of it.

8. Controls, gates, and the limits that travel with the result

• C1 — the locality control, and it is weak, and the design says so (critic pass 2 F7). Four events: the doctor arrives, Mari is bound, a prescription or paper is written, the landlord asks about a cure or raises his other tenants. Registered: ≥ 3 of 4 in ≥ 4 of 5 retellings in every arm. C1 failing shows gross non-locality; C1 passing shows very little.
• G1 — contamination. Any retelling naming Swedish or Finnish with no warrant on its own page — which includes every A2 body — is inspected as contamination, not scored (critic pass 3 F9). contamination.md records that no English rendering of any Canth prose exists on three documented search routes, so recognition would be news about the panel.
• G2 — void bodies. A body quoting text not in the extract it saw is void; one re-dispatch; a seat voiding twice is dropped, the arm reported at n = 4, and the drop named.
• G3 — the floor/ceiling gate, the RS-20260805b lesson. If all fifteen retellings score identically on EXCL, the instrument separated nothing and the result says so in its first line rather than its last.

The findings that survive this revision are carried into the result page's limits section under their own numbers, and are not answered here. They are, in the critic's terms: the scene's four non-language routes cannot be stripped without rewriting Canth (pass 3 F3); keyword coding of free prose is brittle in both directions and blinded dual coding is what would fix it (pass 3 F5); n = 5 supports no estimate of magnitude (pass 3 F1); B has no failure mode and always yields a number (pass 3 F6); A3 adds a narrator cue the source never gives (pass 3 F3). Every one of these is true, none is repaired by this revision, and the result page states them as limits rather than as things the run overcame.

9. The dispatch policy, and why the run proceeds

Three critic passes have returned NEEDS-REDESIGN; thirty-one findings across the three, all accepted, and critic.md records each with what it changed. The policy frozen before pass 3 was: proceed after the third pass whatever the verdict, unless a finding shows that no outcome could bear on the question. No finding shows that. Pass 3's F4 was decisive and is repaired at zero cost by deleting Part 2; F1, F2 and F7 are repaired by withdrawing the confirmatory claim and splitting OPACITY from EXCL; F3, F5 and F6 are not repaired and travel as limits.

A fourth pass is not dispatched. A gate that can be re-run until it passes is not a gate, and a design revised until a model approves of it has been fitted to its critic. What the three passes bought is on the record and is large: the run that will be dispatched asks a different, smaller and answerable question, and its confirmatory ambition was withdrawn before a single seat saw a word.

10. Pre-flight cost — line-itemed per model, per critic pass 2 F12

Rendered seat payload after Part 2 was deleted: 367 words ≈ 500 tokens. max_tokens 1,500 per seat (revision 3 declared 4,000 for a payload that no longer exists).

Every seat is priced at a $15/M ceiling, 2–2.5× the two list prices that are known, per the routing caution in config/models.md (a call has billed 3.8× list through provider routing). Seats out: $0.338. Input: 15 × 650 tok at $3/M = $0.03.

Declared worst case for the seats: $0.37. The three critic passes are spent and cost $0.0602832 against a $0.09 ceiling. Run total worst case $0.43, of which $0.0603 is already spent. Today's UTC ledger before this run: $3.019472636 of $5.00, headroom $1.980527364.

11. Verification

analysis/verify.py recomputes every reported number from the stored raw bodies, imports nothing from tools/, checks the three extracts differ in exactly one line, re-sums per-request costs, runs the exact Fisher probability by enumeration, and carries mutation tests against its own scoring — including non-keyword acceptance tests (critic F11): retellings that express exclusion in words absent from the frozen list must be shown to score EXCL: NO, so the coding's brittleness is measured and reported rather than assumed away. It must exit 0 with zero failures.