Repository path: workshop/experiments/E-20260814d-elevation-resolution/design.md · rendered 2026-09-09

E-20260814d — a register axis with three rungs, and a gate that can fail short of ceiling

Frozen 2026-08-14 (S183) before any call was dispatched. ARM-elevation-resolution step 1.

1. The question, and why the third rung is the whole design

framework/v0.2 §7.4: "It says nothing about elevation: the run's Axis E voided on a dead body and where published hands sit above plain English is still unmeasured." That is Q-e's open half.

The predecessor run (RS-20260813c) built the upper pole and could not read the interior. Its gate G1 required the deliberately ennobled arm to read above the plain arm by ≥ +0.50 on a −1 / 0 / +1 scale, and got +1.0000 in 56 of 56 cells, zero variance, while the published hands it existed to measure sat at +0.125 to +0.70. Method note (bnb): a calibration gate that passes at ceiling has not calibrated anything in the range the primaries occupy.

On a source whose own register is uniformly low, how far above it do published translators pitch their English — and can this instrument resolve the interval at all?

The answer to the second half gates the first. A third rung (R26, light ennoblement) is built so that the calibration gate is a one-step movement rather than a maximal one, and can therefore fail.

2. Materials

Source. H. C. Andersen, «Flipperne» (1848), 756 Danish words in 26 paragraphs, copy-text workshop/experiments/E-20260813g-register-quadrants/corpus/DA.txt, re-fetched from Danish Wikisource this session and identical to the stored file apart from one paragraph break. Single-witness and uncollated; two transcription slips recorded on the translation artifacts.

Seven English arms, all of the whole tale, all held in workshop/experiments/E-20260813g-register-quadrants/corpus/:

2.1 The materials gate, run before this design was written, and it constrains it

tools/dependence_check.py over all 21 pairs. The ten pairs not involving R26 reproduce RS-20260813c §2.1 exactly, which is a check on the tool as well as on the texts.

Five consequences, all binding, all written before any code was seen.

1. R26 is not an independent hand and no claim about published practice may rest on it. It is a constructed pole. Q3 is the only prediction that puts R26 beside a published hand and it is secondary and dependence-flagged, with R26~BRAK at 20 shared twelve-grams named at every use.
2. R06 is contaminated with A (run 19) and BRAK (run 15) and is used only as the lower pole, never as an independent hand. RS-20260813c §6's argument carries over: that contamination pulls R06 up, which shrinks every X − R06 difference, so G1 and G2 are conservative gates — the contamination can make them fail and cannot make them pass spuriously. But a second bias runs the other way and the two are stated together, per critic finding 7(b): the crib is a lead-written word-for-word English rendering, and so is R06. A seat anchoring on the crib would push whichever arm most resembles it toward 0, inflating every X − R06 difference — so G1/G2 are conservative with respect to contamination and not with respect to crib anchoring, and the design does not claim otherwise. One handle on it is free and is registered here: the arm that most resembles a word-for-word crib is R08, which calques, not R06, so crib anchoring predicts R08 sitting nearer 0 than R06 does. That comparison is reported.
3. A ~ BRAK is still the only clean published pair. PAULL is dependence-flagged against both others, so the three published arms are not three independent observations. Q1 and Q2 are stated per hand on the clean pair, with PAULL reported alongside as a dependent sensitivity.
4. R26 ~ R06 is the closest pair in the table by a wide margin, so a design using their contrast as its manipulation must show they differ at the grain the gate runs at — critic finding 6, and the paragraph-level check the design first carried does not entail it. Checked per candidate span before dispatch and recorded in pool.json: token-level similarity over the whole texts is 0.720 against 0.388 for R06~R25; of the 27 candidate spans, 1 is identical (S23a, «Forlovet!»), 3 differ only in terminal punctuation (S10a, S18a, S20a), and the remaining 23 differ lexically. So 4 of 27 spans carry no lexical manipulation at all, and if they are sampled they contribute cells where G1 measures nothing. G1 is therefore computed twice and both figures are reported: over all marked sites, and over the marked sites where R06 and R26 differ lexically. The first is the registered gate; the second is what says whether a failure is the instrument's or the materials'.
5. R26 ~ R25 is clean at 0 shared twelve-grams, so the two rungs above the baseline are not each other's copies.

3. Stage 1 — the site list, three annotators, source only

The repair this stage exists for: RS-20260813c §8 measured its own site list against two independent readers and six of ten lead-marked sites were read as unmarked. The list was one annotator's, and that annotator was the lead. RS-20260808e's three-annotator majority vote is the procedure being copied.

3.1 The candidate pool is built mechanically, with no lead judgement in it. Frozen rule, applied to the copy-text before any annotation:

• Split each of the 26 Danish paragraphs into its quoted and unquoted material. The quoted runs of one paragraph merge into a single candidate (they are one speech turn interrupted by its tag); the unquoted remainder is a second candidate, kept only if it is ≥ 4 Danish words. The floor applies to narration only, and critic finding 8 is right that this needed a written reason: a one-to-three-word unquoted remainder is a syntactic remnant (a speech tag, «sagde Saxen»), which has no register content of its own, whereas a one-to-three-word speech turn — «Snærpe!», «Las!», «Forlovet!» — is a complete utterance and is exactly where this source's register lives. The finding's other half is accepted as a cap rather than as a floor: see §3.3.
• Drop any candidate longer than 40 Danish words — a whole-paragraph judgement of register is a different measurement and the four dropped blocks (59, 42, 159 and 96 words) are named in pool.json.

Realised pool: 27 candidates, 20 speech and 7 narration, frozen in pool.json before dispatch.

3.2 The annotation. Three seats — P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, P3 x-ai/grok-4.5 — each receive the whole Danish tale and the 27 numbered spans, and return one code per span:

Relative to the ordinary written Danish of an 1848 storybook, does the Danish at this span read BELOW that level, AT it, or ABOVE it? → below / at / above.

No English appears anywhere in a stage-1 prompt. Output is a closed form — one code per numbered span, 27 codes — per note (bng).

3.3 The majority rule and the sample, both frozen here.

• A span is marked if ≥ 2 of 3 annotators code it below; neutral if ≥ 2 of 3 code it at; otherwise it is excluded (no majority, or a majority of above).
• From each class, order by int(sha256("E-20260814d|" + span_id).hexdigest(), 16) ascending and take the first 10 marked and the first 5 neutral. Fewer than that available → take all.
• At most 3 of the 15 sampled sites may be spans shorter than 4 Danish words (critic finding 8). Once three such spans have been taken, later short spans are skipped in hash order and the next eligible span of that class is taken instead. Six candidates are affected — S12a, S14a, S15a, S23a (1 word), S20a (2), S18a (3) — and coding seven renderings of a single interjection is a materially harder judgement than the design's other cells.
• The lead's role in stage 1 is confined to building the mechanical pool and extracting each arm's aligned text. The lead does not mark, vote, or break ties.

3.4 The dependence this creates, declared rather than discovered. The three annotators are the same three seats that judge in stage 2. Calls are stateless and no stage-2 prompt contains a stage-1 answer, so the dependence is at the level of a model's dispositions, not of memory — but a seat that reads a span as low may be likelier to read an English rendering as above it, and that is stated as a limit rather than argued away. The alternative — annotating with the reserve seats — was rejected because their Danish is entirely unprobed and P5 is excluded outright by note (bne).

4. Stage 2 — the height run

One call per (site × seat), 3 seats, ≤ 15 sites: at most 45 calls. P5 is not seated, on note (bne). Judgment is not parallelised: each call is stateless and sees one site.

Each site payload carries, for the site the annotators selected:

• the Danish span;
• a crib written by the lead: word-for-word rendering plus dictionary senses, and nothing else. Machine-checked before dispatch against a banned register vocabulary — colloquial, homely, blunt, everyday, literary, elevated, decorous, insult, shout, spoken, plain, formal, register, low, high, ornate, dignified, slang, vulgar — because RS-20260813c's critic finding 1 caught exactly this leak, and the prompt says in terms that the crib is not English prose and is not a yardstick of level;
• the seven arms' renderings of that span, under labels V1…V7, scrambled per (site, seat) by sha256(site_id | seat | "E-20260814d") mod 5040, so no arm sits behind a fixed label.

Each call asks two things, both closed-form:

• (a) source_level — 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) heights — for each of V1…V7 independently: relative to the Danish at this site, does this English read below its level, at it, or above it? → -1 / 0 / +1.

Two instructions carried over from the predecessor's amended prompt, because they were accepted findings there: length is not level — judge the words chosen, not the number of them (critic finding 7), and no rendering is described as "independent" (finding 9).

Blinding assertions, split per note (bnm). The scaffolding is fixed text and carries the ban absolutely: no prompt contains any of the 19 provenance strings Andersen, Paull, Brækstad, Braekstad, Berman, ennobl, regime, foreigniz, resistancy, Venuti, vulgaris, Gutenberg, R06, R08, R25, R26, 1888, 1900, 1597, matched as substrings (several are stems). 1848 is deliberately NOT on this list — critic finding 1, and it was a BLOCKING error: the date is the period anchor the question itself needs ("the ordinary written Danish of an 1848 storybook"), so the frozen assertion was unsatisfiable the moment it was written, exactly as note (bnm) describes. 1848 identifies no translator and no arm; every English arm in the run postdates it. The crib is lead-written data and carries the 19 register words ban — colloquial, homely, blunt, everyday, literary, elevated, decorous, insult, shout, spoken, plain, formal, register, low, high, ornate, dignified, slang, vulgar — matched on word boundaries, not as substrings, because low is inside below and plain is inside explain (critic finding 10). The verifier counts any breach rather than assuming none.

Caps, per note (bnl). Reasoning is never disabled on a judging seat. One probe call per seat is dispatched on a real stage-2 prompt, reasoning_tokens is read off the response, and the cap for the remaining dispatches is set at measured reasoning + 400 for the answer, floored at 1,200 and ceilinged at 3,000. The probe's own cap is 2,500. Stage 1's cap is 4,000 for all three seats and is not probed, because the stage is only three calls.

Re-dispatch, per note (bnn). A body is usable only if it parses as JSON and carries its stage's key (levels for stage 1, source_level and heights for stage 2) with the right arity. Anything else is re-dispatched once at 2× the cap but never above 4,000 tokens, and the stored record is marked usable: false so the analysis and the F3 count key off the same fact. Stage 2's re-dispatches are capped at 6 in total (critic finding 4): beyond the sixth, an unusable body is recorded usable: false without a retry, and the number not retried is reported. Without that cap the design's own rules permit 42 extra calls and the declared ceiling is not a ceiling.

5. Gates, run before any prediction is read

Aggregation, fixed here for every gate and prediction: a figure is the mean of seat-means over seats surviving G4, computed over cells that returned a usable body. Minimum 2 surviving seats.

• G1 — the resolution gate, and it is built so it can fail. At marked sites, mean height(R26) − mean height(R06) ≥ +0.25. This is note (bnb)'s registered remedy: the gate separates the one-step arm from the baseline, not the maximal one. A failure means the instrument does not resolve a movement the size of the one published hands make, and that is a reportable outcome, not an accident.
• G2 — the gross-movement check. Reported; gates nothing. At marked sites, R25 − R06 ≥ +0.50 — the predecessor's G1 verbatim, kept so that a G1 failure beside a G2 pass is interpretable as sees a maximal movement, not a one-step one rather than as sees no register at all.
• G3 — the crib check. Not a model call. Every crib passes the banned-register-vocabulary screen before dispatch, or the site is re-cribbed until it does. Recorded as pass/fail per site.
• G4 — inter-seat coherence, NOT agreement with anybody's map. Rebuilt per critic findings 2 and 3, both BLOCKING and both correct. With three seats, "the majority of the other seats" is a majority of two and does not exist when they disagree, so the statistic was undefined on exactly the sites that matter; and a fixed bar of 8 over an eligible set that can fall below 8 would have voided the run by arithmetic — note (bna)'s failure reintroduced one clause after the note was cited. The rebuilt gate: a seat is retained if its source_level code agrees with at least one other seat's at ≥ 2/3 of the sites where at least one other seat returned a usable source_level. If fewer than 6 sites are eligible, G4 is indeterminate: it is reported as not having run, every seat returning ≥ 80% usable bodies is retained, and the result page says the coherence gate did not fire rather than pretending it passed. Agreement between the stage-2 seat majority and the stage-1 annotator majority is reported as a measurement, never used as a drop rule; it is this run's only check on whether a site list survives being re-asked in a different frame.

6. Predictions, registered before any call

• Q1 — the census. PRIMARY, on the clean pair only. At marked sites, each of A and BRAK has mean height > +0.20. This puts RS-20260808e's ten-hand result to a sixth language pair, and it is the first time it will have been read on a gated interior. PAULL reported alongside as a dependent sensitivity, not part of the primary. Conditional on G1.
• Q2 — is a published hand distinguishable from a deliberate ennobler? PRIMARY. At marked sites, R25 − A ≥ +0.30 and R25 − BRAK **≥ +0.30. **The outcome of interest is the null**: a Victorian hand arriving level with a rule set built out of Berman's complaint would say the complaint is a description rather than an exaggeration, which is the substantive questionARM-ennoblementwas constituted on and never answered. Conditional onG1`.
• Q3 — where do published hands sit relative to the ORDINARY pole? SECONDARY, dependence-flagged at every use. At marked sites, |A − R26| < 0.30 and |BRAK − R26| < 0.30. If it holds, a Victorian published hand and a modern one-step-up policy are the same height above this source, which is the practically useful form of the elevation question. R26~BRAK shares 20 twelve-grams and R26~A shares 2; the flag travels with the number.
• Q4 — policy discrimination at neutral sites. SECONDARY, n small and stated. R25 E1 raises the neutral by rule; R26 N1 explicitly leaves it alone. Registered: at neutral sites, R25 − R06 ≥ +0.25 and |R26 − R06| < 0.25. This asks whether the seats track two declared policies or a single global impression of fanciness, and it is the cheapest test of that this project has been able to build.
• Q5 — the direction question. EXPLORATORY. At marked sites, R25 − R08 ≥ +0.30: markedness toward the source is not coded as elevation. RS-20260813b §7 found a target-only style question could not tell the two kinds of markedness apart; this asks the same question source-relative, which is the form R25's regime page says is answerable. R08 is dependence-flagged against R06 and R26.

7. Failure criteria, registered

• F1 — G1 fails → Q1, Q2 and Q3 are withheld. G2, Q4, Q5, the stage-1 site list and the full descriptive table are still reported, as instrument facts, with the resolution failure stated as the run's result.
• F2 — G4 leaves fewer than 2 seats → every prediction is withheld.
• F3 — more than 10% of the bodies belonging to seats retained by G4 are unusable after one re-dispatch → the run is reported as incomplete and every prediction is withheld. A seat dropped by G4 cannot also void the run through F3 — this is note (bna)'s generalisation applied to the criterion that fired at S174, and it is registered here before dispatch rather than re-read afterwards.
• F4 — stage 1 yields fewer than 6 marked or fewer than 3 neutral spans by majority → stage 2 is not dispatched and the run reports the site-list measurement only. The money not spent is recorded as not spent.
• F5 — fewer than 3 annotators return a usable stage-1 body after one re-dispatch → the majority rule is undefined and stage 2 is not dispatched.
• No override of F1–F5 is available to this session. RS-20260813b §6 refused an override that would have rescued its own headline and RS-20260813c §5 refused a second; those refusals are the precedent and they bind here.

8. What this run cannot establish

• Nothing about quality. Register height is a position, not a verdict. Tier D remains NOT PASSED; every figure is internal-judgment-only and provisional.
• Nothing about three independent published hands (§2.1(3)).
• Nothing attributing a R25 − R26 gap to lexical height alone. The two regimes differ in height and in whether expansion and clarification are permitted; R26's own limitations section says the reduction is compound.
• Nothing about Danish generally, and nothing about any period's norms: R26's ceiling is ordinary now, and a Victorian hand's ordinary sits above a modern one's by an amount this design does not measure and cannot subtract.
• Nothing about readers. Three language models coding height are not readers, certified or otherwise, and their Danish is unprobed at 1848 orthography.
• Nothing that survives F1. If the interior cannot be resolved, the census numbers are not evidence and the run says so.

9. Money

Pre-flight priced from config/models.md as corrected 2026-08-14 (P2 $0.75/$3.75), from max_tokens and not from an assumed answer length (note (abc)), with re-dispatch in the arithmetic (note (abc)'s S079 firing).

Session ceiling $1.50. UTC-day headroom at design time: $2.763776 of $5.00 remaining (2026-08-14, three prior sessions at $2.236224). A run that does not fit is split or deferred; F4 and F5 are both routes to spending materially less than the ceiling and that is a normal outcome.

Lead translation of «Flipperne» under R26 is $0 and is not ledgered (charter §3, A4), as are the copy-text re-fetch, the dependence measurement, the mechanical pool and the verifier.

10. Verification

analysis/verify.py, importing nothing from tools/, recomputes every reported number from the stored bodies and asserts:

1. every stage-1 and stage-2 prompt is free of all 19 banned provenance strings;
2. every crib is free of all 19 banned register words, and the count of breaches is reported rather than assumed zero (note (bnm));
3. every arm span in every payload is a contiguous substring of that arm's stored corpus file, normalising whitespace only — the guard against a rendering being paraphrased into the payload;
4. the label scramble is a bijection at every (site, seat) and agrees with the runner's record;
5. every individual cell code matches the stored body, not merely the aggregate — RS-20260813b §8's repair, where a flipped cell survived a majority-level check;
6. every reported mean recomputes from the per-cell codes;
7. the site list recomputes from the stage-1 bodies under §3.3's majority rule and hash order;
8. mutation sensitivity: flipping one stored code, and separately corrupting one payload span, each change at least one asserted figure.

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

Critic: moonshotai/kimi-k3 (P4), no role in this run's annotation or its jury (P1/P2/P3), and not a seat in it at all. Ten findings, six BLOCKING, four NON-BLOCKING. All ten accepted; two of them in an amended form with the reason written. The design above is the amended one, and nothing had been dispatched. Cost $0.112383000, provider Chutes, 4,034 reasoning tokens of a 6,000 cap. Raw body: run/critic.json.

The verdict line did not arrive: the body hit finish_reason: length mid-finding-10, so the critic's own summary verdict is missing and finding 10 is truncated after its first sentence. The sentence is complete enough to act on and was acted on. Note (bnk) fired again, on the seat it names, with its own remedy already applied — effort pinned low in the first dispatch and the cap sized from the observed 6,000-token pass at S181 — and it still truncated, because ten findings is a longer answer than eight. Note (bng) is the more exact diagnosis: an open-ended enumeration is the one output shape whose cap cannot be guessed, and a critique is exactly that shape. The finding is recorded rather than repaired: a second dispatch at a larger cap would have bought a verdict word this session does not need, at the price of a second $0.11.

One thing that is not the critic's and is recorded so the amendment list is honest: the site hash keys on span_id only, so none of these amendments changes which spans the sample will draw or which label any arm sits behind.

12. Pre-dispatch amendment A1 — after stage 1, before any stage-2 call

Written and committed 2026-08-14 after stage 1 returned and before a single stage-2 call was dispatched. Stage 1's three annotators returned 3 of 3 usable bodies and produced a majority site list far thinner and far less agreed than the design anticipated: 7 marked, 13 neutral, 0 above, and 7 spans with no majority at all (every one of those seven split three ways, above/below/at). F4 does not fire — its bar is 6 marked and 3 neutral by majority, and both are met.

But two frozen rules then interact in a way nobody foresaw at freeze time. Six of the seven marked spans are shorter than four Danish words, and §3.3's short-span cap (critic finding 8) allows at most three such spans in the sample. The frozen sample is therefore 4 marked sites, of which — by §2.1(4)'s own per-span table — S23a is lexically identical between R06 and R26 and S20a differs only in terminal punctuation. Half the cells of the registered resolution gate would carry no manipulation at all.

The amendment, and what it deliberately does not do.

• All 7 majority-marked sites are dispatched, not 4 — 12 sites, 36 calls instead of 27.
• No registered bar moves and no registered denominator changes. G1, G2, Q1, Q2, Q3 and Q5 are computed on the frozen 4-site marked sample, exactly as §5 and §6 say. The 7-site figure is reported as a stated sensitivity and may not replace the primary, whichever direction it goes.
• This is registered here, in writing, before dispatch, because the failure mode it courts — computing two figures and choosing between them afterwards — is the one this project's verification discipline exists to stop. Adding cells cannot rescue a bar that is computed over a fixed set; what it can do is make a G1 failure interpretable rather than ambiguous, which is the whole reason G1 was built to be failable.
• The three sensitivity sites are S18a, S14a and S12a, named now so the set cannot be chosen later.

Two stage-1 facts recorded here rather than in the result, because they bear on the dispatch. P3 returned 6,516 reasoning tokens against a max_tokens of 4,000 and still delivered a complete body — so on that provider the cap did not bound the reasoning at all, and the call cost $0.045616 against a per-call worst case of $0.0272 built from the cap. Note (abc)'s worst case is not a bound on a reasoning seat, which is note (bnk)'s point arriving from the other side. And the three annotators' base rates for below over the same 27 spans are 7, 20 and 3, which is the measurement §5's G4 was written to catch and the reason the stage-2 probe is worth its three calls.