Repository path: workshop/experiments/E-20260810c-register-reach/design.md · rendered 2026-09-09

E-20260810c — the powered cell: how much of English's way down is bought by the place, and how much by the spelling?

Frozen 2026-08-10 (S150) before dispatch. ARM-register-reach step 1. The translation limb (T-botchan-R22-v1, its log, and its contamination measurement) was committed at 915bf89 before this file existed.

1. Question

framework/v0.2 §7.1 refuses a register-carriage recommendation and names, in writing, the one thing a successor needs:

the register comparison needs more sites per language cell than fourteen across two […] A cell of ~20 marked narration sites in one pair, with sites chosen for device availability rather than only for source markedness, is what would make P1a readable.

This is that cell.

At narration sites where a source drops below its own neutral written register, how much of English's downward movement is bought by a LOCATED IDIOM and how much by a NONSTANDARD SPELLING — asked where there are enough sites for either answer to be legible?

2. What is carried from E-20260809g and is not re-asked

1. P3 is settled and is not re-tested. Nonstandard spelling is not placeless: loc(B) = 0.5714 / 0.5357 against a registered 0.25 on two independent judges, respelling placed more often than located idiom in all four judge × cell blocks. This run measures loc only as a manipulation check (G2), not as a hypothesis.
2. A21 — where an arm's English at a site is byte-identical to that hand's ∅, d is 0 by construction and the seats' codes are not differenced.
3. A19's conclusion matrix is carried in §11 with the cell count changed and nothing else.
4. Note (bkr) — no content-parity check gates anything here. The err and add questions are asked inside the reproduced loc prompt and are reported descriptively only.
5. REG is a direction, not a magnitude, and is batch-sensitive by a factor of three (RS-20260806g). Every difference is read within one call: the rating is split by site, with all eight arms of a site kept together in one call, and absolute per-block means are never pooled across blocks.

3. Five registered changes from E-20260809g, each with its reason

No bar carried from S145 is loosened. G2, G3, G6′ and every prediction threshold are the S145 values verbatim.

4. Materials

Why this span is the right material and not a convenience. It is 3.66× the Japanese the S145 JA cell was cut from, it is continuous, it is the part of chapter 1 with no deferential speech in it, and it is clause-shaped throughout — Sōseki's 俺-narration is finite predicates end to end, which is the syntactic property framework/v0.2 §7.1 identifies as what gives a respelling anything to attach to.

5. Stages

Stage 0 — critic. One adversarial pre-run pass over this frozen file plus R23, openai/gpt-5.6-terra, reasoning off (note (bkw)), cap 16,000. Findings accepted or overruled in writing, before dispatch.

Stage 1 — the site census, three independent annotators. Each is given the source span, the frozen criterion verbatim, and nothing else: no English, no mention of translation, no hypothesis, no indication that register direction or any device is at issue. Each returns a JSON array of stretches with src and type.

• B1 mistralai/mistral-medium-3-5 · B2 z-ai/glm-5.2 · B3 nvidia/nemotron-3-ultra-550b-a55b
• The vote, frozen in code.py before dispatch. B1 is the spine by dispatch order and not by merit. A B1 stretch is endorsed by B2 or B3 if that census lists a stretch overlapping it under vote_sites.py's overlap() copied verbatim from E-20260808e. A stretch is admitted at ≥ 2 of 3. type is the majority of the votes cast on it; a tie takes the spine's.
• Only type: N sites go forward. If more than 24 are admitted, the first 24 in source order are taken — registered here, before the census exists. If fewer than 14 are admitted, G5 cannot be met and §10 F7 fires.
• |B2 \ B1| and |B3 \ B1| are reported as the coverage the spine missed.

Stage 2 — generation, R23's four cells, two hands. H1 x-ai/grok-4.5 · H2 qwen/qwen3.7-max. Each hand writes its own ∅ and then three minimal revisions of its own ∅, one call per cell, using E-20260809g's BASE/CLAUSE/REVISE/TAIL prompt strings word for word. A hand is given the source stretches, the rule text for its cell, and (for a revision) its own ∅. It is not told the hypothesis, that other cells exist, or that any comparison will be made. 8 calls, cap 6,000.

Stage 3 — loc. Two judges over every ⟨site, arm⟩ pair (8 arms). The prompt is E-20260809g's LOC_HEAD word for word, including its two companion questions. L1 nvidia/nemotron-3-ultra-550b-a55b (S137's and S145's checker) · L2 z-ai/glm-5.2 (S145's). Split by site into two blocks; cap 16,000.

Stage 4 — REG. The signed −3…+3 register-direction scale, reproduced word for word from E-20260806b via E-20260806g, E-20260808e, E-20260808f and E-20260809g. Three seats: J1 openai/gpt-5.6-terra · J2 google/gemini-3.6-flash · J3 deepseek/deepseek-v4-pro — S145's seats, same instrument, no new competence claim is made. Split by site into three blocks with all eight arms of a site together; independent shuffle per seat under a fixed seed; cap 12,000.

Role double duty, declared. The panel and reserves reachable here are nine models and this design needs eleven roles. B2/L2 and B3/L1 are the same models in two stages. A census annotator saw the Japanese only and was told nothing about translation; a loc judge is asked about English placedness. No generating hand, and no REG seat, is a census annotator. openai/gpt-5.6-terra is both critic and J1, as at S145.

Not dispatched, deliberately: moonshotai/kimi-k3 — note (bkh-corr).

6. Quantities

For arm x and site s, REG(x,s) is the mean over the returned seats of the register-direction code. Positive means the English is higher than its source. For a generating hand h:

• d_I(h,s) = REG(h∅,s) − REG(hA,s) — what the located idiom buys.
• d_S(h,s) = REG(h∅,s) − REG(hB,s) — what the respelling buys.
• d_AB(h,s) = REG(h∅,s) − REG(hAB,s) — what both buy.

Positive d = the arm went lower than its own ∅. Site-level values pool the two hands.

7. Gates, registered before dispatch

8. Predictions, registered

P1a — which device, policy contrast (reach included). Over all admitted sites, hands pooled within site: mean(d_I) − mean(d_S), tested by an exact two-sided sign-flip permutation blocked at the site, α = 0.05.

Registered in both directions, and the direction the record points is stated so it cannot be claimed afterwards. S145's unread JA margin was −0.5833 (mean d_I 0.2083, mean d_S 0.7917) — the spelling, in all four hand × cell blocks and under every leave-one-seat-out. P1a′ (spelling buys more) is therefore the expected direction and is the weaker claim to make; P1a (idiom buys more) would overturn a consistent margin. Either is licensed only at |difference| ≥ 0.50 and P < 0.05 and the same sign in both hands.

P1b — conditional contrast. The same test restricted to sites where, for that hand, both A ≠ ∅ and B ≠ ∅. P1a is about a policy; P1b is about the devices where both were applied, and the two answer different questions (S145 A1+A2+A6).

P2 — the bundle. mean(d_AB) ≥ +0.50, pooled over hands and sites, de-linked from S137's magnitudes.

P4 — reach, mechanical, no jury. For each hand and device, the share of admitted sites at which that device's cell differs from that hand's own ∅ at all.

P5 — the translator's claim, tested against hands that never read it. T-botchan-R22-v1's frozen log states that at 19 of its 21 low sites (0.905) the readiest English was locatable and had to be refused. Registered: reach(I) pooled over the two hands ≥ 0.75. If the independent hands find a located option far less often, the log's claim is a fact about the lead's English and not about English. This is the wire between the two limbs and it costs nothing to run.

Also reported, descriptively and with no test: the share of independently censused N sites that fall inside the lead's 21 rows, in both directions.

9. Failure criteria

No bar in this table moves after it fires. If a gate fires, the margin is reported and not exploited.

10. Limits, declared in advance

1. Site-level generation and site-level rating. Arms render and are rated on short stretches out of context. A device's effect over a whole paragraph is not measured, and T-botchan-R22-v1's own log argues the low register of this narration lives in its speed, which is a paragraph property. This design cannot see that, and a null here does not reach it.
2. One work, one author, one pair. Sōseki's 俺-narration is a particular kind of low, and nothing here generalises past Japanese narration of this shape.
3. Two hands. Both are LLMs.
4. No human rater exists in this project. The standing BLOCKING from E-20260808e's critic is unrepaired and is a property of the project, not of this run.
5. The seats are three LLMs and the loc judges are two more, two of which also censused the Japanese (§5, role double duty).
6. T-botchan-R22-v1 is the lead's and the lead knew the arm's question while writing it. It is rated by nobody here. P5 uses only the counts in its frozen log, not its English.
7. Tier D is NOT PASSED. Nothing here is a quality judgment; the seats code a source–target register relation and no goodness sense is scored.

11. Conclusion matrix — binding on the result page

Read down the first row that applies. Nothing below the line a gate draws may be claimed.

12. Pre-flight cost

Built from the caps the requests permit, not from expected output (note (abc)).

Declared ceiling: $2.40. UTC day 2026-08-10 stands at $0.071580820991 of $5.00 before this session, so the ceiling fits the day's headroom of $4.928419 with $2.53 to spare. Actuals recorded from usage.cost per response and reconciled against the key-usage delta.

13. Amendments from the pre-run critic (2026-08-10, before dispatch)

critic.md carries the twenty findings, the adjudication, and the six overruled remedies in full. openai/gpt-5.6-terra, NEEDS-REDESIGN, 20 findings — 4 BLOCKING, 11 SERIOUS, 5 MINOR — of which 17 accepted, 3 accepted-in-part with the overruled half written out, and 6 individual remedies overruled with the reason. Amendments A1–A20 are in force and supersede §§4–11 above wherever they conflict. The originals are left standing rather than rewritten, so that what was frozen and what was changed are both visible.

The three that changed the experiment:

• A1 (BLOCKING 1). The estimand is a permission-policy effect, not a device effect. A opens a broad lexical/idiomatic/grammatical class and B opens one narrow orthographic class; they are not matched doses. Every sentence reads permitting X, never device X buys.
• A3 (BLOCKING 3). The word exact is struck from the test. Nothing is randomized; the sign-flip reference distribution rests on an assumed within-site exchangeability of the two permission labels under the sharp null, that assumption is printed wherever the p-value is, and inference is conditional on these sites, these two hands and these three seats.
• A16 (MINOR 16). The census vote is symmetric: candidates are the union of all three annotators, clustered under the verbatim overlap() rule, admitted at ≥ 2 of 3 distinct annotators. There is no spine, so a stretch two annotators found and the third missed is admitted.

Also in force, in brief. A2 — P1b demoted to exploratory and barred from every claim, being post-treatment conditioning. A4 — the licensed sentence replaced with procedure-level wording. A5 — cap 24 → 30, and above it a paragraph-stratified fixed-seed sample replaces first-n. A6 — non-tie defined in code at the analysis level, G5 evaluated once. A7 — the whole analysis algorithm frozen in analyse.py before dispatch. A8 — G6′ renamed a minimal internal manipulation check. A9 — G3 strengthened: ≥ 10% negative overall and ≥ 10% among ∅/A items alone. A10 — every REG figure is an LLM-panel-perceived relation. A11 — naturalness struck from senses:, nothing here collects it. A12 — site-level compliance coded; P1a declared intention-to-treat. A13 — G1 split into per-site purity and aggregate application. A14 — content parity reported per arm plus an add-excluded sensitivity, and no parity check gates anything (note (bkr), overruled in writing). A15 — the sample is labelled an enriched feasibility sample. A17 — the outcome is a comparative panel judgment within a call. A18 — P5 loses its threshold and is a coarse cross-procedure comparison. A19 — the request map fixed: every seat rates every block. A20 — the word powered struck; this is an enlarged measurement run.

14. Two amendments registered during the run, before any English existed

• A21. A16's union-and-cluster rule, implemented as single-link clustering, chained 312 stretches into 26 blobs. The unit of agreement becomes the sentence — character-level coverage per annotator, 111 sentences, admitted at ≥ 2 of 3. Symmetric, order-independent, cannot chain, and it is the coarsest unit the criterion's own exclusion 3 permits. Registered with no English and no rating in existence.
• A22. The census admits 97 of 111 sentences (0.874). The criterion was built to find islands of low register and this narration is low throughout. The phrase "marked narration sites" is struck: the sample is 30 of the 82 criterion-positive narration sentences of this span, drawn paragraph-stratified at seed 5501, and every figure says so.

Both are stated in full in critic.md.