Repository path: workshop/experiments/E-20260802d-deference-carry/design.md · rendered 2026-09-09

E-20260802d — is the deference loss forced by the pair, or chosen by this translator?

Frozen 2026-08-02 (S090), before any call. ARM-atelier-cycle step 2, study limb, arm 3 of 3. The span-2 translation and its log were frozen at commit 4a87026 before this design was written, and no seat in this experiment is shown the lead's English.

1. Question

Finnish marks the class line in «Köyhää kansaa» grammatically, twice in span 2:

• ¶132 — Mari, speaking to the merchant's wife, addresses her in the third person: «Kyllä minä ihan varmaan rouvalle toimittaisin» (I would certainly get it to the lady).
• ¶104 — Hellu, speaking to her own mother, uses the honorific plural: «Viivyttekö kauan?» — while her mother addresses her and every other child in the singular.

The lead's span-2 log declares both losses total (register.md V13, V14) and attempts a compensation at only one of them (a vocative, "ma'am"). A declared loss is honest only if the alternative really was unavailable. So:

Do independent translators, given this Finnish cold and no instruction about register, mark the deference at these two sites — and if so, by what device?

2. What this teaches about translating literature — rewritten, amendment A5

Whether a grammatically-encoded social relation survives into a target that lacks the grammar is one of the oldest practical questions in literary translation, and it is normally answered by assertion. This is a small version of it on one text, and its reach is small in exactly the way §9 says. A positive is decisive and a negative is weak: if any independent attempt marks the deference by a device the lead declared unavailable, register.md V14 is wrong and is corrected this session. If none does, the only licensed sentence is three independent machine attempts, under no instruction, did not find a device the lead overlooked — which says nothing about English and nothing about what a competent human translator would do.

The remaining six spans are not bound by a null result here.

3. Materials

Two passages, Finnish only, taken verbatim from the stored source (workshop/translations/koyhaa-kansaa/source-fi-full.txt), stored under materials/ and hashed:

Amendment A4 extended Site A by one paragraph so that «äiti lähti kaupungille» is inside the passage and the relation is recoverable from the materials, which §3 previously claimed and did not deliver. The superseded cut was 39 words, sha256 32c96de05617454c. Residual limitation, stated now: the passage establishes a mother and a girl in one room, not that the girl is her daughter, and Finnish te is general polite address rather than a parent-directed form.

Each site carries enough context around the target line for the relation to be inferable — who is speaking to whom, and their relative standing — and nothing else. No English, no gloss, no mention of class, deference, register, address, politeness, or of this project.

Bibliographic frame — amendment A3. The frame (author, title, 1886) primes period English, of which the thou/you contrast is a feature, so a grammatical code in a framed arm could be a response to the date rather than to the Finnish. Site A is therefore run twice on the same passage with the same seats: framed (author, title, year) and unframed (no bibliographic information at all). A GRAMMATICAL code appearing only in the framed arm is date-priming; one appearing in both is a response to the source. Site B keeps the frame only, and its result is read with the confound named rather than removed.

4. Seats

Three non-Anthropic panel seats as contrast subjects, not judges (charter §3; nothing here is a quality judgment and no calibration state is invoked):

P4 is excluded on note (bhf): its reasoning appetite burned $0.503 for zero characters at S087.

5. Procedure

1. Stage 0 — independent pre-run critic on this frozen design (P5), before any subject call. Findings accepted or refused in writing in critic.md; the design is amended before dispatch.
2. Stage 1 — nine calls (A3): Site A × 3 seats × 2 frames, plus Site B × 3 seats. Stateless, temperature 0, max_tokens 4000. The prompt is identical across seats but for the passage and the frame. This is not judgment, and the nine calls are independent by construction.
3. Stage 2 — coding by the lead, against the rules in §6, which are fixed now.

6. Coding rules, fixed before any output is seen

Site A — code the rendering of «Viivyttekö kauan?» (the girl → the mother). The Finnish contains no vocative at this line, so any English address term is an addition, not a carry.

Amendment A2: KIN and GRAMMATICAL both count as marking the deference. The earlier scheme folded kin terms into a neutral VOCATIVE bucket, which made prediction 1 unfalsifiable by the one device English actually uses. The evidence that this matters is already in hand: the only human translator of this text on record, Hertzberg 1886, dropped the honorific plural and marked the relation with the kin term — "Blir mamma länge borta?".

Site B — code the rendering of «rouvalle» in «Kyllä minä ihan varmaan rouvalle toimittaisin». Amendment A1: what is measured is where the address noun ends up when the construction is converted — not whether the seat "compensated", which no outcome here could show.

Codes are read off the seat's own words; every coded string is quoted verbatim on the result page so the coding can be checked without the raw files.

7. Predictions, registered before dispatch (1 and 2 restated by A2 and A1)

1. Site A: 0 of 6 (site × frame) cells mark the deference at all — i.e. no GRAMMATICAL and no KIN. (Any mark falsifies register.md V14's claim that no compensation exists; F3 fires, the erratum is written this session, and the result page leads with it.)
2. Site B: ≥ 2 of 3 seats code 2P-VOC — the address noun survives, relocated out of its grammatical slot into a vocative. (This licenses a claim about where the noun goes, and explicitly not that the vocative is an independently-reached compensation — A1.)
3. Site B: 0 of 3 seats code 3P.
4. Site A: ≥ 1 of 6 cells codes NONE — at least one translator lets the deference vanish, which is what V14 does deliberately.
5. Frame check (A3): the framed and unframed arms of Site A agree on every seat. (Disagreement means the 1886 frame, not the Finnish, is producing whatever marking appears.)

8. Failure criteria

• F1 — a body that is not a translation of the passage (refusal, commentary, partial) is a seat failure; one retry, then the seat is dropped and every count is reported out of the seats that returned, never renormalised to 3.
• F2 — fewer than 2 seats returning on a site: that site reports no result, and the prediction on it is neither confirmed nor refuted.
• F3 — prediction 1 fails: V14 gets an erratum this session; the probe's headline becomes the falsification.
• F4 — prediction 2 fails at 0 of 3: the lead's "ma'am" is an addition, not a compensation, and D27 is restated on the translation page.

9. What this cannot show, stated before it runs

• Three language models are not a survey of English, and they are not human translators. Convergence among them is weak evidence (charter §4) and is weaker still for being convergence among systems with correlated training.
• A negative bounds one thing only: that three independent attempts, under no instruction, did not find a device the lead overlooked. It does not establish that no such device exists.
• Convergence on the vocative is not evidence that the vocative is good. It is evidence about what gets reached for, nothing more.
• Every judgment in the result page about the quality of any rendering is internal-judgment-only and provisional (Tier D NOT PASSED, config/models.md).

10. Budget

Declared worst case from max_tokens, per note (abc) — built from the cap the request permits, not from an assumed output length:

Against $3.024465590 headroom on UTC day 2026-08-02 (four prior sessions, $1.975534410). Routing is read off every response (note (x)) and every body checked against its cap (note (bgk)).