Repository path: workshop/experiments/E-20260813c-ennoblement-direction/design.md · rendered 2026-09-09

E-20260813c — where a published translator sits on a register axis that now has both ends

Frozen 2026-08-13 (S174) before any call was dispatched. ARM-ennoblement step 1.

1. The question, and why it needs a new pole

Berman's fifth and sixth deforming tendencies are one axis: ennoblissement upward, vulgarisation downward. The project built the downward end (R21, S124) because a measurement found it empty. The upward end has never existed, so every register figure this project has published — including RS-20260808e's ten published hands across five language pairs, pooled REG +0.9716, no hand lowering where the source lowers — was measured on a scale with a low pole, a plain pole, and nothing above the line. framework/v0.2 §7 records that it cannot make a register-carriage recommendation.

R25 (frozen today, before either translation was written) supplies the upper pole. The question this run asks is the one that becomes askable only once it exists:

On a source whose own register is uniformly low, where do published translators sit on a register axis that has both ends — at the plain pole, or up beside a rule set built to ennoble on purpose?

A subsidiary question the materials force and the design does not dodge: is Berman's ennoblement, executed deliberately, doing something more extreme than what Victorian translators actually did, or is it merely a description of them?

2. Materials

Source. H. C. Andersen, «Flipperne» (1848), 767 Danish words — the project's first text in Danish, and its sixth language pair for this instrument. Chosen because Andersen's register here is spoken and low almost throughout, which is the condition under which ennoblement and any source-ward markedness point in opposite measurable directions.

Five English arms, all of the whole tale:

2.1 The materials gate, run before this design was written, and it constrains the design. tools/dependence_check.py, all pairs:

Three consequences, all binding:

1. R06 is contaminated and may not be read as an independent hand. Its own artifact already restricted it to the lower calibration pole on the ground of declared session priming; the measurement confirms the restriction rather than discovering it. No claim in this design rests on R06's position relative to a published hand — it is used only against R25.
2. The only clean published pair is A ~ BRAK. PAULL is dependence-flagged against both other hands, so the three published arms are not three independent observations, and no pooled published figure here may be read as a three-hand agreement. P1 is therefore stated per hand.
3. R25 is clean against every hand. The ennobled arm, written after the plain one by the same primed translator on the same day, shares no 12-gram with anything.

3. Sites

14 sites, frozen in sites.json before dispatch: 10 marked (the Danish sits below its own neutral written level — bare colloquial exclamations, homely compounds, the particle cluster «jo/vel/nok», the third-person address-drop) and 4 neutral (plain narration at the story's ordinary level). Each site carries the Danish, a literal English gloss of the Danish written by the lead, and the five arms' renderings of that site.

4. Procedure

One call per (site × seat). 42 calls: 14 sites × 3 seats.

Seats: P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, P5 deepseek/deepseek-v4-pro (the standing non-Anthropic jury; config/models.md). Judgment is not parallelised across sites within a seat's reasoning — each call is stateless and sees one site only.

Blinding. Arms are presented under labels V1…V5, scrambled per (site, seat) by sha256(site_id | seat | "E-20260813c") mod 120, so no arm sits behind a fixed label and order is controlled without paying for two orderings. No provenance, no date, no translator name, no mention of a regime, of Berman, of ennoblement, or of this project appears in any prompt.

Each call asks two things:

• (a) The source-side probe. Relative to the ordinary written Danish of an 1848 storybook, does the Danish at this site read below that level, at it, or above it? → -1 / 0 / +1.
• (b) The register-height codes. For each of V1…V5 independently: relative to the Danish at this site, does this English read below its level, at it, or above it? → -1 / 0 / +1.

Answers returned as strict JSON. Caps sized per note (b)/(bmb): max_tokens 1500 with reasoning_effort pinned low, and the re-dispatch guard triggers on a body that fails to parse, irrespective of finish_reason — note (bmz), whose remedy is applied here for the first time.

5. Gates, run before any primary is read

Aggregation, fixed here for every gate and prediction: a figure is the mean of seat-means over seats surviving G3, computed over cells that returned a usable body. Minimum 2 surviving seats. (Critic finding 8.)

• G1 — the calibration gate. At marked sites, mean height(R25) − mean height(R06) ≥ +0.50. R25 was written on purpose to raise the register; if the seats cannot see a movement the translator made deliberately, they cannot see register at all.
• G2 — the construction check. At neutral sites, mean height(R25) − mean height(R06) ≥ +0.25. E1 raises the neutral too, so a movement visible only at marked sites would mean the seats are tracking something other than register. Threshold and aggregation specified per critic finding 3; the lower bar reflects n = 4 and the fact that E1's neutral effect is smaller by construction.
• G3 — inter-seat coherence, NOT agreement with the lead's map. Rewritten per critic finding 4, which is accepted in full. The original form dropped a seat for disagreeing with the lead's own marked/neutral labels — a gate that fires on a competent reader who contests the map. Replaced by: a seat is retained if its source_level codes agree with the majority of the other seats at ≥ 8 of the 14 sites. The agreement between the seat-majority source_level and the lead's marked/neutral labels is reported as a measurement, not used as a gate — it is the only independent check this run has on whether the lead's register map of a Danish text is one other readers share, and turning it into a drop rule would destroy that.

6. Predictions, registered before any call

§2.1's dependence table governs which of these can be primary. Critic finding 5 is accepted: PAULL is dependence-flagged against both other published hands, so it cannot supply an independent observation, and A~BRAK is the only clean published pair.

• P1 — the census. PRIMARY, on the clean pair only. At marked sites, each of A and BRAK has mean height > +0.20. This is RS-20260808e's ten-hand result put to a sixth language pair and a new source. PAULL is reported alongside as a dependent sensitivity and is not part of the primary.
• P2 — the new question. SECONDARY, and deliberately restated. Critic finding 2 is accepted: "R25 sits above the published hands" is close to true by construction, since R25 was written to raise the register and G1 already requires the seats to see it. What is not true by construction, and is what P2 now registers, is: is PAULL distinguishable from R25? Registered prediction: R25 − PAULL ≥ +0.30 at marked sites. The outcome of interest is the null — a Victorian published hand arriving level with a rule set built out of Berman's complaint would say the complaint is a description rather than an exaggeration. P2 is conditional on G1 and attributes nothing to E1/E6 rather than E3/E4 (§8).
• P3 — the ordering among published hands. EXPLORATORY, not primary. PAULL > BRAK at marked sites. Demoted per critic finding 5: these two hands share a 16-token run and 7 twelve-grams, so an ordering between them is not an ordering between independent observations. Reported with that caveat attached at every use.
• P4 — the marked/neutral contrast. EXPLORATORY, not primary. Each published hand's mean height is higher at marked than at neutral sites. Demoted per critic finding 3: n = 4 neutral sites cannot carry a primary. It is the substantive form of the floor claim and is reported as a direction with its n stated.

Why G1/G2 may rest on the contaminated R06 arm, stated rather than assumed (critic finding 5's last clause). R06's contamination is with A (run 19) and BRAK (run 15), both of which sit above the Danish on the hypothesis under test. Contamination therefore pulls R06 up, which shrinks R25 − R06. G1 and G2 are consequently conservative gates: the contamination can cause them to fail, and cannot cause them to pass spuriously. No shared n-gram between R06 and a published hand raises R25, which shares none with anything.

7. Failure criteria, registered

• F1 — G1 fails → every primary is withheld. The run reports the gate and nothing else.
• F2 — G3 leaves fewer than 2 seats → every primary is withheld.
• F3 — more than 10% of the 42 bodies are dead or unparseable after one re-dispatch → the run is reported as incomplete and every primary is withheld.
• No override of F1–F3 is available to this session. RS-20260813b §6 refused an override that would have rescued its own headline; that refusal is the precedent and it binds here.

8. What this run cannot establish

• Nothing about quality. Register height is a position, not a verdict. No sense is calibrated; Tier D remains NOT PASSED; every figure is internal-judgment-only and provisional.
• Nothing about three independent published hands (§2.1(2)).
• Nothing attributing R25's movement to raised vocabulary rather than to expansion and clarification — R25's own limitations section says the arm is compound and this run does not decompose it.
• Nothing about Danish generally. One tale, one author, one register policy.
• Nothing about the direction question in its full form. Separating "marked but fluent" from "marked and estranged" needs a source-ward marked arm (R08) that this run does not build; that is ARM-ennoblement step 2, and the brief is written in §10 of the result page.

9. Money

Pre-flight, priced from config/models.md at max_tokens 1500 (note (abc): the worst case is built from the cap, not from an expected length), with a 4× routing surcharge on every seat:

Session ceiling $1.40 including the $0.160873565 already spent on the D-20260813-17 gate. Day headroom at design time: $2.04 of $5.00 remaining.

10. Verification

analysis/verify.py, importing nothing from tools/, recomputes every reported number from the stored bodies and asserts: every prompt is free of arm names, provenance, dates and the words Berman, ennoble, regime, foreignizing; the scramble is a bijection at every (site, seat); each reported mean recomputes from the stored per-cell codes; and every individual cell code matches the record, not merely the aggregate — RS-20260813b §8's repair, where a flipped cell survived a majority-level check.

11. Amendment before dispatch: the pre-run critic's findings, accepted

Critic: x-ai/grok-4.5 (P3), no role in this run's jury (P1/P2/P5). Verdict NEEDS-AMENDMENT, 5 BLOCKING and 4 NON-BLOCKING findings. All 9 accepted; the design above is the amended one, and nothing had been dispatched. Cost $0.0177364. Raw body: run/critic.json.

One finding was NOT the critic's and is recorded so the amendment list is honest: the site sha256 scramble keys on site_id only, so rewriting the glosses did not change any arm's label assignment, and the blinding is unaffected by the amendment.