Repository path: workshop/experiments/E-20260811g-carrier/design.md · rendered 2026-09-09

E-20260811g — the carrier, not the content: a 2 × 2 inside one Turkish story

Frozen 2026-08-11, before any call. The translation limb it hangs on (T-kasagi-R06-v1, the whole story) was frozen and committed at d58caa5 before this design was written, and its contamination was measured after that commit and before any locus was chosen.

1. Where the question came from

It came from translating, at log decision D6, and from a control in the previous session's run.

Every second-person form in «Kaşağı» is sen; siz never occurs. D6 decided to compensate for none of it and to log eighteen sites as losses. That is the same decision ARM-dakghar's V5 made about Bengali, and RS-20260811f then measured it: CROSS 0.0000, 60 of 60 cells SAME — the honorific verb ending reached English at exactly zero.

But that run's own added control did reach English, at 0.8333. C-c changed the same social relation by changing an address term — a free word — and the arbiters saw it. Same play, same seats, same call, same question; the only difference was what the Bengali used to carry the mark.

The conjecture this design tests: what predicts arrival is the carrier, not the content. A distinction the source expresses in a free word reaches English; the same distinction expressed in a bound morpheme does not.

Two things stop that being established, and the design is built against both. (a) It has been seen only in the social domain, so English does not do social is a rival explanation of equal standing — so content is crossed with carrier. (b) The free-word arm was a positive control chosen to be easy and was never matched against the bound arm as an equal arm — so here both arms are primaries, in the same windows, against the same baseline rendering.

2. Question

Holding the language, the work, the passage, the translating hands, the arbiters, the question and the statistic constant, does a one-word change to a passage of Turkish reach an independent English hand's rendering more often when the word is free than when it is bound — and does that hold for an epistemic distinction as well as a social one?

3. Materials — six windows, three variants each, one word apart

Every variant is built by exact string substitution from the frozen copy-text workshop/translations/kasagi/source-trwikisource.txt (tr.wikisource revid 172873), by build_items.py, which asserts that each replacement fires exactly once in its window.

Each window has an A (as printed), a GRAM (the mark carried by a bound morpheme) and a LEX (the same mark carried by a free word). A is shared by both contrasts, so GRAM and LEX are measured against the identical baseline rendering.

SOCIAL — the mark elevates the footing between two speakers

EPISTEMIC — the mark makes the narrator's access indirect

[20b] is the sub-span of the printed paragraph [20] from Ben bir gün yalnız başıma kaldım. to Az daha sevincimden haykıracaktım. — declared here because it is the one window that is not a whole printed unit.

C-NEG — the positive control, and it is a bound morpheme

LEXNULL — the added-word control, added on the critic's SERIOUS 3

Together C-NEG and LEXNULL bracket the artifact. C-NEG is a bound edit that must arrive; LEXNULL is a free edit that should not. If both behave, neither bound is invisible nor free is always visible can explain P1.

Both recensions agree at every locus

collation.md establishes that the two reachable Turkish e-texts are different modernisations. All seven edited strings were checked in the MEB school recension before the design was frozen and all seven are present verbatim; only surrounding words differ (Hadi/Haydi, Niye/Neden, haykırdı/bağırdı, birdenbire/aniden). No measurement here turns on a word the two texts disagree about.

The asymmetry this design does not remove, stated before the run

GRAM edits substitute one word; LEX edits add one. So a LEX variant is one word longer and its English will usually be one phrase longer. This is registered as the leading rival explanation of any positive result — arbiters notice added length — and the design's defence against it is G1 below plus C-NEG, not a claim that it has been eliminated. Both arms are one-word edits and both add a mark rather than remove one; that is the extent of the matching.

4. Procedure

Stage T — the independent hands. Two non-Anthropic panel seats, neither the lead and neither an arbiter, are each given one variant at a time, alone, with no mention of politeness, evidentiality, address, or that an experiment exists. The prompt asks for an English translation of the passage and nothing else.

• T-a = moonshotai/kimi-k3, T-b = x-ai/grok-4.5 (config/models.md P4, P3).
• Per hand: 6 windows × 3 variants (A, GRAM, LEX) + 2 LEXNULL + 1 C-NEG + 4 A-repeats in independent calls = 25 calls; 50 hand calls in all.
• The A-repeat exists to measure the floor: how much one hand differs from itself on identical input. No hand sees two variants of a window in one context.

Stage J — the arbiters. Three non-Anthropic seats, none of them a translating hand, are shown two passages at a time and asked one fixed question, blind to the item's class and to the design.

• J1 = openai/gpt-5.6-terra, J2 = google/gemini-3.6-flash, J3 = deepseek/deepseek-v4-pro (P1, P2, P5). No seat both translates and arbitrates.
• The question is deliberately symmetric across the two contents, so that neither arm is favoured by what the arbiter is asked to look for:

"Below are two versions of the same short passage. Ignore differences of wording, phrasing or style that make no difference to the sense. Do the two versions convey the SAME thing, or something DIFFERENT, about (a) what is asserted or implied about the people and what happens, (b) how the speaker stands socially to the person spoken to or spoken about, and (c) what the speaker claims to know and how they claim to know it? A difference of wording or style alone is not a difference."

Answer as JSON: {"verdict": "SAME"|"DIFFERENT", "what_differs": "..."}. - Item classes, 50 items × 3 seats = 150 calls: - SOURCE — the two Turkish variants (A vs GRAM, A vs LEX). 6 × 2 = 12. - CROSS — hand h's English of A against hand h's English of the variant. 6 × 2 × 2 hands = 24. - LEXNULL — the added-word control, W1 and W5 only, 2 × 2 hands = 4 (critic A3). - FLOOR — hand h's English of A against hand h's English of A again, on W1, W2, W4, W5 only, 4 × 2 = 8 (narrowed from six windows to pay for LEXNULL, critic A3). - CONTROL — hand h's English of W1 A against hand h's English of W1 C-NEG. 2. - Item order is shuffled under a seed fixed in the built file; which passage is shown first is randomised per item and recorded.

Measure. d = 1 if the seat returns DIFFERENT, 0 if SAME. A cell that does not parse is void and reported, never imputed.

5. Registered gates and predictions

Written before any call. Every gate is a withholding gate: if it fails, the primary is not read.

• G1 — source parity, and it is load-bearing here, not a checksum. On SOURCE items, the GRAM and LEX contrasts must both come back DIFFERENT at mean ≥ 0.667, and their means must differ by ≤ 0.25. If the free-word edits change the Turkish more than the bound edits do, an English difference measures the size of the edit and not the carrier. This is the gate RS-20260811f demoted because nothing there turned on it; here everything does.
• G2 — the bound-morpheme positive control. C-NEG on CONTROL, mean ≥ 0.75. A bound morpheme that English must carry has to arrive, or the GRAM arm measures the arbiters' blindness to short edits rather than anything about carriers.
• G3 — the floor. On FLOOR items, mean d must be ≤ 0.40. The run is void only if FLOOR ≥ CROSS overall.
• G4 — the added-word control, ADDED on the critic's SERIOUS 3, and it is the gate that matters most. LEXNULL on CROSS, mean d ≤ 0.50. LEXNULL adds one free word carrying neither a social nor an epistemic mark (— Gel buraya, hadi!; Hemen kasabaya at gönderildi.). If adding any word makes an arbiter call the two Englishes different, P1 measures added-word salience and not the carrier.
• P1 — the primary, and it can fail in either direction. Over the six windows, on CROSS: mean d(LEX) − mean d(GRAM) ≥ 0.40 ⇒ the carrier predicts arrival. A bootstrap 90% interval over windows is reported alongside; if it straddles 0.40 the verdict says so and the threshold is not treated as a boundary the data can resolve.
• P2 — the crossing, which is what makes P1 more than a restatement of RS-20260811f. P1's difference must hold in both contents separately: LEX − GRAM ≥ 0.25 within SOCIAL and within EPISTEMIC. If it holds in SOCIAL only, the rival explanation — English does not do social relations, and this is not about carriers — survives and the result page says so as the primary reading.
• P3 — the null half, registered so that it can fail. Over the six windows, d(GRAM) − d(FLOOR) ≤ 0.20 ⇒ a bound mark reaches English at the rate one hand differs from itself, replicating RS-20260811f in a second language family. If any single window reaches d(GRAM) − d(FLOOR) ≥ 0.50 it is named and the null is stated as predominantly, not universally.
• P4 — carriers named, not just counted. A positive P1 is only reported as relocation if the arbiters' what_differs text names an English device (an adverb, a modal, a vocative, a title, a choice of verb) at ≥ half the LEX windows that drove it. A numerical difference with no nameable carrier is reported as unexplained.

Failure criteria. The run is void if: fewer than 90% of cells parse; any hand returns a body that is not a translation (a refusal, a commentary, a transliteration); or FLOOR ≥ CROSS overall.

6. What this design cannot establish

1. Two hands is two hands, three arbiters is three arbiters. Any null is about these seats on this material, not about English.
2. Six windows. The denominator is small and P1's margin is not a confidence interval.
3. The arbiters share a training distribution with the hands. "An independent reader cannot tell them apart" means these arbiters.
4. Variant B is not attested text, in either arm. G1 is what stops that being fatal.
5. One language. Even a clean crossing licenses a statement about Turkish→English, and the arm page binds step 2's wording to that.
6. The added-word asymmetry of §3 is bounded, not removed.
7. Nothing here scores the lead's own translation, which no seat sees; the lead never judges its own translation (charter §5).
8. Free-word carriage and explicit marking are not separable here (critic A4). A free word that carries evidentiality or deference is by that fact more explicit than a suffix. "The carrier predicts arrival" and "explicitness predicts arrival" are two readings of the same result and this design does not choose between them.
9. Every GRAM edit is a pragmatically loud violation and every LEX edit is pragmatically consistent (critic A1), because siz occurs nowhere in this story. This biases against the registered prediction and is registered here before the run.

7. Declared deviations

• No senses: are invoked. This is a descriptive measurement of what survives into English, not an evaluation. Any evaluative sentence on the result page carries internal-judgment-only.
• The work was chosen for its grammar. Turkish was selected because it grammaticalises both an epistemic and a social distinction that English lacks. The windows and the edits were chosen after the translation and its log were frozen and committed at d58caa5, from the log's own D6, D7 and D8; the lead is not blind to them. The design's protection is that the lead neither translates for the measurement nor arbitrates it.
• EPISTEMIC-LEX uses one device three times (anlaşılan); SOCIAL-LEX uses three (ağam, abla, beyim). The social device has to fit its speaker pair and no single Turkish word does all three. Recorded, not corrected.

8. Pre-flight cost estimate

Built from max_tokens, not from expected answer length (note (abc)).

Declared ceiling: $1.46. UTC day 2026-08-11 stands at $3.497487166 of $5.00 before this run, so the ceiling leaves $0.04 of the day's cap unspent even if every call bills its worst case. A re-dispatch that would carry the run past the ceiling is not made: it is deferred to the next UTC day and reported as a gap, never imputed. max_tokens is 400–1,200 because a cap prices the estimate and must also be large enough to hold an answer (S148, S157, S160 all truncated a body).

9. Pre-run critic — adjudication

One pass, qwen/qwen3.7-max (reserve slug, neither a hand nor an arbiter), cap 12,000, finish_reason: stop, $0.020465625. Verdict NEEDS-AMENDMENT, 8 findings — 1 BLOCKING, 4 SERIOUS, 3 ADVISORY. Full text at critic.md. Five accepted, three accepted in part with the overrule written.

What the critic bought. One withholding gate the design did not have (G4), a registered direction-of-bias statement on the whole social arm (A1), a per-content reading of G1 (A5), and a limit (A4) that changes how a positive result may be worded. The declared ceiling does not move: A3 was paid for out of FLOOR breadth, not out of budget.