Repository path: workshop/experiments/E-20260811b-realia-channel/design.md · rendered 2026-09-09

E-20260811b — the two channels a translated page is located by

ARM-realia-channel step 1. Frozen before any judge call and before any material was built. Tier D is NOT PASSED: every jury figure this design can produce is an LLM-panel-perceived judgement about English and no goodness sense is scored.

1. Question

A translated page is located twice. It is located by the world it describes — roubles, a samovar, a queue, a yamen, Akaki Akakievitch — and it is located by the English it is written in — colour against color, towards against toward. The translator controls the first almost completely and the second hardly at all.

Can a reader tell them apart?

Specifically: does removing the source-culture words change a reader's judgement of whose English the page is written in, when not one letter of spelling has changed? And the mirror: does changing the spelling change a reader's judgement of where the story is set, when not one culture-bound word has changed?

2. Why it is a translation question and not an instrument question

The subject rule (wiki/tracks.md, continue-prompt.md §4.5) requires one sentence saying what this teaches about translating literature or evaluating translations. The sentence:

It measures what the oldest decision in translation — keep the source's word or replace it — does to a reader, on a channel the translator did not intend to touch.

RS-20260811-floor §6 met this phenomenon as a contaminant of its own instrument and filed it as a limit. This design turns it round and makes it the object. What comes out lands on wiki/goodness-senses.md's cultural-mediation, whose entire content is culture-bound items and whose only evidence to date is about whether raters sort such decisions into the right box, not about what the decisions do.

3. Design — a 2×2 crossed within passage

Two factors, both applied to the same passage, so that everything else about it — its author, its subject, its period, its difficulty, and any recognition a judge has of it — is a constant and cannot produce an effect.

Four forms per passage: KB KU MB MU. Plus two controls (§5).

3.1 Passages — eight, selection rule stated before selection

From the eighteen texts of E-20260811-floor/materials/corpus.json, already on disk and already public domain, plus the translation limb T-fengbo-R06-v1 frozen this session at 79835e3.

Selection rule, applied in this order and mechanically where it can be:

1. The text must be translated narration (or, for the two originals in the corpus, excluded — an English-original passage has no source-culture world to mute).
2. A candidate window is a contiguous run of 110–150 words beginning at a paragraph boundary.
3. The window must contain ≥ 2 orthographic spans that are members of code.py's _PAIRS invertible list, so that P is manipulable.
4. The window must contain ≥ 4 realia spans, so that W is manipulable.
5. The earliest qualifying window in the text is taken. No window is chosen for what it says.
6. One passage per hand, and no more than two per source language.

Realia spans are annotated by the lead, because no word list can find them. This is declared as a non-blind step and is bounded three ways: the annotation is committed as data before any judge is called; it is made without knowing which of QP/QW any judge will answer first; and every substitution is mechanically checked against §5's conditions, which forbid the annotation from touching the orthographic channel at all.

3.2 The two elicitations, in separate calls

The whole design turns on the two questions not contaminating each other, so a judge never sees both about the same passage in the same context. One item, one call, stateless.

QP — the prose question. Which national variety of English is this passage written in? BRITISH, AMERICAN or CANNOT-TELL. Then quote up to three words or phrases from the passage that decided it. Then rate 0–6 how strongly the writing itself — spelling, idiom, phrasing — is marked as one national variety. The prompt states, in terms, that subject matter and recognition of the text are not grounds and that the answer in that case is CANNOT-TELL — the same instruction RS-20260811-floor used and which did not hold; it is repeated here so that the comparison to that run is like for like, and its failure is not this design's primary.

QW — the world question. Where is this passage set? Name a country, or CANNOT-TELL. Then quote up to three words or phrases that decided it. Then rate 0–6 how strongly the passage's subject matter locates it in a particular country or culture.

3.3 Judges

Three seats, three labs, per config/models.md. L1 x-ai/grok-4.5 (P3) · L2 deepseek/deepseek-v4-pro (P5) · L3 google/gemini-3.6-flash (P2), the last with reasoning: {"effort": "low"} — note (bmb), which cost S156 45.7% of its spend in dead bodies and is applied here from the first dispatch. The pre-run critic is openai/gpt-5.6-terra (P1) and judges nothing. z-ai/glm-5.2 is excluded as a seat on the measurement in RS-20260811-floor §6: it flipped its verdict on three of four sentences it saw twice unchanged.

4. Predictions, registered

P1 is the primary. If P1 fails, the finding is that the two channels are separable and the translator is free; that is a first-class result and is written as such.

5. Controls and gates

6. Materials, counts and cost

Pre-flight ceiling. Worst case at max_tokens on the most expensive seat's list price, times a 4× routing allowance (the provider caution in config/models.md): 240 × (450 × $3/M + 700 × $15/M) ≈ $0.09 × 4 ≈ $0.35 for the dearest seat if it were all three; the realistic mixed figure is well under that. Plus one critic pass at a 12,000 cap — S156's first critic pass truncated at 6,000 and had to be re-run, so the cap starts at 12,000 here — at ≈ $0.10.

Declared ceiling: $1.20. Today's UTC ledger stands at $0.329161470 of $5.00 (S156), so headroom is $4.67 and this fits without deferral.

7. Procedure

1. code.py builds materials/passages.json: windows selected by §3.1, orthographic spans located by importing E-20260811-floor/code.py's _PAIRS and marks(), realia spans read from a committed annotation file, all forms generated, G4 asserted.
2. Pre-run critic (openai/gpt-5.6-terra, cap 12,000) over this design and the built materials. Findings adjudicated in writing in critic.md; amendments committed before any judge is dispatched.
3. run.py dispatches 240 calls, one item per call, raw request and response JSON preserved per call, usage.cost and provider read off every response.
4. analyse.py computes every figure in §4 and §5.
5. verify.py recomputes every number the result page reports, from the raw files, by a different route, and includes mutation tests that must be caught.

8. Failure criteria, written before the run

• G2 fails on two or more seats → the run cannot support P1 and the result says so.
• G3 fires → P1 withheld.
• P4 < 0.70 on all three seats → the design measured seats that do not read orthography; P1's direction is uninterpretable and is reported as such.
• P1 fails → the channels are separable, reported as the finding, not as a disappointment.

9. What this design cannot do

1. It is a forced-choice task, not reading. RS-20260811-floor §9.5 applies unchanged: a judge asked whose English is this is doing something no reader of a novel does.
2. Eight purposive passages, five source cultures, two of them Russian. Every figure is corpus-specific.
3. MUTE is a caricature of domestication. A real domesticating translator does not replace every culture-bound noun with a generic; the manipulation is the extreme end of a decision that is normally made site by site.
4. The realia annotation is the lead's, and the lead also wrote one of the eight passages. G5 bounds the second; nothing bounds the first except the mechanical checks in G4.
5. LLM seats are not readers. Nothing here is licensed as a claim about human reading.

10. Order of operations, recorded because S156's was not clean

T-fengbo-R06-v1 and its translator's log were committed at 79835e3, before this file existed. This file is committed before code.py runs, before any material is built, and before any judge is called. Everything that changes after this commit is recorded as a numbered amendment in critic.md with its direction of bias, and the amendments are committed before dispatch.