Repository path: workshop/experiments/E-20260810t-drift-source-present/design.md · rendered 2026-09-09

E-20260810t-drift-source-present — is the zero a fact about the sense, or about a source-blind reader?

ARM-terminology-drift step 2 of 2, T2. Frozen before dispatch; the pre-run critic pass is §9 and every amendment it forces is recorded there with the wording it replaced.

1. The question

RS-20260809i (S147) found that a jury scoring consistency on the English alone rates a passage that renders the same realia term two different ways exactly as coherent as a composition-matched passage that renders every term one way: mean difference 0.000, nine of twelve units tied, P = 0.875, against a comparator carrying the identical number of foreign tokens. The same seats, the same scale and the same passages dropped 2.000 points when one proper name was spelled two ways.

That result has a rival reading of equal standing, and the result page registered it as the run's sharpest limit (§5.1). voice was scored with the Ukrainian on the page and consistency was not, because consistency is defined as target-internal. So the finding may be about the sense — terminology drift is not a coherence event — or about the reader — a reader of the English alone has no way to know that polonyna and the high pasture answer to one Ukrainian word, and drift is visible as drift only against a source.

This design separates them. The same operator, the same scale, the same rubric string, the same four seats, in two conditions that differ in exactly one thing: whether the Ukrainian is on the page.

Subject-rule statement (continue-prompt §4.5). What this unit teaches about evaluating translations: whether internal coherence is a property a reader can assess without the source at all, or whether the sense the typology calls target-internal is in fact a source-relative sense wearing a target-internal definition. That is a claim about a sense of good, not about the project's apparatus.

2. Materials

Two units of the same novella, six blocks, 123 marked class sites.

tini-budz is Kotsiubynsky's milking and cheese-making sequence, four pages further on in the same novella — the same institution, the same class of culture-bound vocabulary, translated in this session by the lead under R04 with its R06 draft frozen first. Provenance, the one elided verse passage, the four editorial glosses that were read, and the within-block repetition table are on workshop/translations/tini-budz/README.md. It doubles the units of analysis from 12 to 24 and supplies a second, independently translated drift realization — RS-20260809i §5.3 named the single realization as a limit and could not afford a second.

Arms. Generated mechanically by materials/arms.py (57 assertions, all passing before dispatch) from materials/template.txt, which is itself built by substitution on the frozen R04 text (materials/build_template.py) rather than retyped:

MATCH is the whole design, and it is E-20260809i's pre-run critic's BLOCKING 1: it carries the same number of foreign tokens as DRIFT in every block, from as nearly as possible the same items, and differs from it only in whether the variation falls between items or within them. Residual per-item L1 on the new unit: 10 / 6 / 8 tokens.

Composition on tini-budz, fraction of class tokens in COPY form:

Fifteen items drift within a block (against S147's smaller set): BD diynytsia ×6, vivchar ×5, strunka ×4, kozar ×3, zahoroda ×3, honinnyk ×2; BE vatah ×5, staya ×2, berbenytsia ×2, budz ×2, zhentytsia ×2; BF vatah ×7, spuzar ×3, budz ×3, staya ×2.

Two declared asymmetries between the units.

1. MIXED is class-uniform in BE and BF. On tini-budz the base translation domesticated only two class items (вівчар → shepherd, загорода → pen) and both occur only in BD. So the class-non-uniform exhibit this project has been reading since S012 lives, on this unit, in one block of three. tini-polonyna's MIXED (0.875 / 0.789 / 0.842) remains the exhibit; MIXED here is reported and carries no registered test either way.
2. SPELL exists for BA, BB and BD only. BC has no proper noun occurring twice (S147's finding); BE's single name occurs once and BF contains none. The control therefore runs on 3 blocks × 4 seats = 12 units per condition, against 24 for the primary.

3. The two conditions, and what separates them

The manipulation is information, not question. Both conditions ask the same target-internal question and both explicitly refuse the accuracy reading. What changes is whether the reader can know that two English strings answer to one Ukrainian word. If the sense is genuinely target-internal, that knowledge is irrelevant and the two conditions agree.

4. Procedure

• One arm per call. No seat is told that any other rendering exists, or that a condition exists (E-20260809i critic, BLOCKING 4).
• Both conditions dispatched in ONE interleaved pass, order shuffled once from frozen seed 20260810 and stored at materials/dispatch-order.json before any call. Running SB to completion and then SP would confound the condition with anything that changes at a provider over the hour the run takes, and the condition contrast is the primary.
• Four seats, the same slugs in the same roles as S147: openai/gpt-5.6-terra, google/gemini-3.6-flash, deepseek/deepseek-v4-pro, x-ai/grok-4.5. Not five or six: the SB arm is a replication of S147's C stage, and comparability is worth more here than the marginal power two more seats would buy.
• Unit of analysis = block × seat. 6 × 4 = 24 for every primary and secondary; 3 × 4 = 12 for the control.
• Cells. 2 conditions × 4 seats × (4 arms × 6 blocks + SPELL × 3 blocks) = 216 calls (DOM, MATCH, DRIFT, FALSE, and SPELL on three blocks). Total re-dispatches capped at 20 (critic MINOR 12); beyond that the run stops and reports incomplete.
• temperature: 0.0, max_tokens: 4000. A body that returns finish_reason: length with no content is moved to runs/discarded/ and re-dispatched once at 8,000 with reasoning disabled; the dead body is kept and ledgered (notes (bkw), (blj)). A body that returns length with readable content is NOT silently accepted — note (bls), S151: it is counted, named in the result, and its score used only if the JSON object is complete.

5. Registered predictions and bars

All tests are exact sign tests by enumeration, one-sided where a direction is registered, ties excluded from the test and reported. Nothing below may be altered after any body is read.

Let, for unit u = (block, seat): d_SB(u) = consistency(MATCH,SB,u) − consistency(DRIFT,SB,u), d_SP(u) = consistency(MATCH,SP,u) − consistency(DRIFT,SP,u).

Every P below indexes stochasticity within these four fixed model seats only (critic SERIOUS 6). None of them is an inference to a population of readers, and none may be reported as one.

Registered expectation, stated so it can be wrong. I expect P1 to fire and P3 to replicate — that is, the zero is about the reader. The alternative outcome that would most change the project's sense list is P1 and P2 both failing while C1 passes in both conditions: that would say terminology drift is not a coherence event for these readers even when they can see it is drift, and the parenthesis's first half would have to be struck rather than qualified.

6. Failure criteria and withholding rules

• F1 — the gate. If C1 fails in a condition, every null in that condition is uninterpretable and is reported as such. If C1 fails in either condition, P1 is withheld entirely: a difference of differences needs both differences to mean something. P3 may still be reported if C1 passes under SB.
• F2 — completeness. If fewer than 95% of the 216 cells return a parseable integer in range, the run is reported as incomplete and every figure carries the completeness rate. Below 90%, no registered test is reported at all.
• F3 — G1 parity. mistralai/mistral-medium-3-5 (not a jury seat) checks, with the Ukrainian shown, six real arm pairs on the new unit plus four planted propositional errors. Bar: at least 4 of 6 real pairs equivalent AND at least 3 of 4 planted errors caught. Below either, the twelve DOM strings written for tini-budz are declared unverified and every budz contrast involving DOM or MATCH is withheld — a DOM string that says something different would sit in MATCH and DRIFT a different number of times and is therefore a live confound, not a cosmetic one. The old unit's DOM forms passed this gate at S147 (8 of 8 pairs, 4 of 4 planted) and are imported unchanged.
• F4 — no post-hoc analysis choice. Test, units, direction, bar and tie handling are fixed above. analyse.py is written before dispatch and verify.py recomputes every reported number by independent routes.

6a. G1 parity — run before dispatch, and it took two seats

Outcome: 4 of 4 planted errors caught, 5 of 6 real pairs equivalent. Nothing withheld.

The first seat failed the task, not the materials. mistralai/mistral-medium-3-5 returned non-equivalent for all ten pairs — the six real ones and the four planted ones alike — citing in every case exactly the lexical differences the prompt excludes ("A uses 'milking gate' vs B uses 'strunka'"). A gate that never returns equivalent has no discriminative power and cannot bar anything; its three bodies are kept in runs/discarded/ and ledgered at $0.030216. The seat was a poor choice on my part: config/models.md already records it as the panel's weakest and not panel material.

The second seat is moonshotai/kimi-k3, which passed this identical task at S147 (8 of 8, 4 of 4) and is not a jury seat. $0.047766.

Item-level inspection, as the amended F3 requires at 5 of 6. The single flag is on drobyeta, whose DOM form is head of sheep at the BE site («Мосійчук має штирнадцять дробєт»). That rendering is denotationally correct, and the "difference" the seat named is the foreign-versus-English difference the prompt explicitly excludes — the same error the first seat made ten times out of ten, made once here. The DOM string is upheld and nothing is withheld, and the flag is recorded rather than argued away. It touches no MATCH/DRIFT contrast in any case.

7. Pre-flight cost estimate

Worst case is built from the caps the requests actually permit, not from expected output (note (abc)).

Expected, from S147's realised $0.0053/call over 128 calls with half of them source-bearing: ≈ $1.50. Today's UTC ledger stands at $0.867877 of $5.00 before this run, so the ceiling fits the headroom with $1.27 to spare.

8. Declared limits, written before the run

1. The translator knew the design. tini-polonyna was translated before E-20260809i existed; tini-budz was translated in this session with RS-20260809i §8's successor design already written down. The consequence is bounded and stated: MIXED on the new unit is the lead's own class handling and carries no weight in any registered test (§5, D1). MATCH and DRIFT are generated from the marked sites and are invariant to which form the base chose at any site, so the primary is untouched by what the translator knew.
2. Four model seats, not readers. Tier D is NOT PASSED; every figure is provisional and internal-judgment-only.
3. One work, one class, one translator, two passages. The class is Hutsul pastoral realia and nothing here generalises past it without being shown to.
4. Two drift realizations, not many. One seeded assignment per unit. Better than S147's one; not a distribution.
5. SP may buy its effect with an accuracy halo. The prompt refuses the accuracy reading explicitly, and refusing it in words is not the same as preventing it. If P1 fires, the honest statement is the source's presence changes the score, and the mechanism — drift becoming visible, or accuracy leaking in — is not settled by this design. That is registered here so it cannot be quietly dropped from the result.
6. C1's 12 units are half the primary's 24, so the gate is the weakest-powered test in the run and a C1 failure could itself be a power failure. If C1 fails, that reading is reported beside the withholding.

9. Pre-run critic — NEEDS-AMENDMENT, 12 findings, 3 BLOCKING

One independent adversarial pass over the frozen design and its predecessor result, moonshotai/kimi-k3 — not a jury seat and not S147's critic. 11 accepted, 1 accepted in part, none overruled. Two calls: the first returned finish_reason: length with zero content, the whole 16,000 cap spent on hidden reasoning ($0.266061, dead, ledgered); the re-dispatch with reasoning disabled cost $0.053280. Note (bkw)'s remedy, fifth occurrence.

The amendment that matters, and it changed what was dispatched: BLOCKING 3. §8.5 conceded that the source-present condition might buy its effect from an accuracy halo, and then §5 registered an interpretation — "the zero is about the reader" — that assumed it had not. Nothing in the frozen design separated the two. The remedy is the FALSE arm: MATCH, with no drift at all, plus one planted propositional error per block that contradicts the Ukrainian and is invisible from the English alone (materials/false_arm.py, 30 assertions; the six edits and the reason each is source-blind-invisible are tabled there). It is registered as C4. MIXED was dropped to pay for it — 48 cells out, 48 cells in, the total unchanged — and that is the right trade twice over, because the critic's MINOR 11 showed MIXED was weaker than §8.1 had admitted.