Repository path: workshop/experiments/E-20260814f-carriage-elevation-2/design.md · rendered 2026-09-09

E-20260814f — is source-carriage still what the seats track when the source-ward arm is the ORNATE one?

ARM-carriage-direction step 2, the arm's last declared step. Frozen before any cell is dispatched. The translation limb — T-miserere-R08-v1 and T-miserere-R07-v1 — was frozen at e3c14b9, before this design existed.

1. The question, and why it is not step 1 again

RS-20260813h (step 1) asked three seats which of two English renderings followed the original's own way of putting things, and got 15 of 15 for the source-ward arm with the source hidden, plus 15 of 15 with the Danish shown. It refuted the rival on the record — that such a question measures distance from plain English — at 0 of 30.

It left one rival standing, and said so in its own §7: R25, the elevated arm, was on the losing side of every condition, so the single heuristic "choose the less elevated version" reproduces all 45 choices without any reading for carriage at all.

The question here: when the source-ward rendering is the LONGER and MORE ELABORATE one, do the same three seats still choose it?

2. Why the pair can be built, and it is about language family rather than about this tale

On Danish, the source's marked word-formation is compounding, and carrying it into English yields short Germanic English: step 1's source-ward arm sat nearer the plain pole than its partner. On a Romance source the same rules invert. R08's R6 (calque the source's figure) and R3 (archaism licensed where the source word is dated or literary), applied to peristilo, pórfido, sepulcral, contrición, versículo, sillar, ojiva, produce peristyle, porphyry, sepulchral, contrition, versicle, ashlar, ogive — source-ward and register-raising at the same site, because a Latinate cognate is learned diction in English and ordinary diction in Spanish.

The rule set did not change; the language family did. That is the whole mechanism by which the confound becomes separable, and it is stated here before any number exists.

3. Materials

Source. Gustavo Adolfo Bécquer, «El Miserere» (1862), the §II vision: 668 words, ten paragraphs [P1]–[P10], copy-text Obras escogidas 1912 (PG #53552), collated token-for-token against the 1885 Obras witness — 668 tokens each, two substantive divergences. Copy-text gate, comparator location, and one declared exposure to the Bates 1909 English: provenance.md.

Arms, both by the lead, both frozen at e3c14b9 before this design existed.

The two arms are length-matched to 1.8% — 725 tokens against 712. This is not incidental. Step 1's pre-run critic raised as BLOCKING 4 that R25 was 20–47% longer than its partner in every segment and that a seat could be answering "which is longer"; step 1 accepted the finding and could not solve it. Here it is solved, and by the material rather than by an argument: elaboration on this pair is carried by word length and period length, not by word count.

Per segment, on the judged material only (amended on critic F-C, which was right that a whole-text figure including the two unjudged paragraphs is not the number a per-cell heuristic would act on). Declared tolerance, fixed before dispatch: ±10% per segment.

All five inside tolerance, and the sign is not constant: in S2 the fluent arm is 6% longer. A per-cell "choose the longer one" heuristic therefore predicts R07 in one of five segments and R08 in three, and cannot reproduce a 15-of-15 result in either direction.

The elaboration statistics, computed by analysis/stage1.py and committed with this design.

Stated plainly: the predictor was fixed after the two renderings' descriptive statistics were seen and before any jury cell existed. The four statistics were declared in device-census.md §" Registered, mechanical" before either rendering was written; which of them separate the arms could only be known afterwards, and is reported unselectively above, including the one that does not.

Contamination, measured after the freeze and before this design. (Table rewritten on critic F-D and F-L. tools/dependence_check.py's verdict rule, which the first draft left implicit: a pair is DEPENDENT? iff it shares one or more twelve-grams, and clean otherwise — a threshold, not an impression.)

The jury pair is clean, which is the condition the run needs, and it is clean despite the two renderings being written back to back by one hand in one session — the case note (bhb) says is the project's worst self-overlap risk.

Two corrections to what a first draft of this section said, both on critic F-D. (i) It said the Bates figures "are equal for both arms": they are not. The longest runs are equal at 13 and 13; the 7-gram and 12-gram counts are 29 / 3 against 16 / 2, so R08 shares nearly twice as much with the 1909 translator as R07 does. (ii) It read as though "one declared exposure to Bates" and "neither lead arm had Bates open" were both being asserted of the same text. They are not in conflict and the wording was loose: the declared exposure (provenance.md §5) is nineteen grep-hit lines, of which exactly one falls inside the span and contains no English at all (the Latin verse), and the Bates English of the span was opened only after both renderings were committed at e3c14b9.

The asymmetry is registered as a hypothesis, not explained away — see H-PERIOD in §5.

Segments, frozen in stage1.py before dispatch: S1 = P1+P2 · S2 = P3 · S3 = P4+P5 · S4 = P6 · S5 = P8. [P7] is the Latin verse and is in no segment; [P9]/[P10] are translated but not judged.

Where the two substantive collation divergences fall (critic F-H). Deus/Domine is in [P7], the Latin verse, which is in no segment. capiteles/chapiteles is in [P3], inside judged segment S2 — and both arms render the copy-text reading with the identical English word capitals, so the variant cannot differentiate them in any cell.

4. Conditions

3 conditions × 5 segments × 3 seats = 45 cells, one call each.

Seats: P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, P3 x-ai/grok-4.5 — the same three as step 1, so the instrument is imported unchanged. P5 excluded on note (bne); P4 is the pre-run critic and takes no cell.

Presentation order fixed per cell by sha256("E-20260814f|" + condition + "|" + segment + "|" + seat) mod 2.

The carriage question is E-20260813h's as dispatched, verbatim — not as first frozen, but as amended on that run's critic BLOCKING 1, which struck a gloss naming the measured features:

Two English translations of the same passage from a nineteenth-century short story are printed below. They translate the same original, which is in another European language. Neither is the original.

One of these two translators set out to follow the original's own way of putting things as closely as English allows. The other set out to do something else.

Which of the two followed the original's own way of putting things?

Answer on ONE line, in exactly this form, and print nothing else: A or B, then a semicolon, then at most fifteen words naming the one feature you used.

The elevation question (EL only), built parallel to it in form:

Two English translations of the same passage from a nineteenth-century short story are printed below. They translate the same original, which is in another European language. Neither is the original.

One of these two is written in a higher and more elaborate English than the other.

Which of the two is written in the higher and more elaborate English?

Answer on ONE line, in exactly this form, and print nothing else: A or B, then a semicolon, then at most fifteen words naming the one feature you used.

In AS the Spanish is prepended under the heading ORIGINAL. The word Spanish appears in no dispatched prompt in any condition, which is stricter than step 1, where the language was named in the gate.

5. The three hypotheses, and what each predicts here

• H-CARRIAGE — seats track which way a translation is marked. Predicts R08 on AB and AS. On EL it makes no prediction until G2 has returned (critic F-F: the first draft predicted R08 on EL "since R08 is in fact the more elevated arm", which takes as a premise the outcome of a registered gate). Conditional form: iff G2 passes, H-CARRIAGE predicts R08 on EL; if G2 fails, all EL predictions are void.
• H-ELEVATION — seats reject the more elevated version, the rival RS-20260813h §7 could not exclude. Predicts R07 on AB and AS.
• H-DISTANCE — seats name whichever version is further from plain English. Predicts R08 here, the same arm as H-CARRIAGE. This design does not separate those two and does not need to: step 1 separated them and H-DISTANCE lost at 0 of 30.
• H-LITERALITY — registered on critic F-A, and it is the finding that most changes what this run may conclude. Seats choose whichever text reads more like a translation — less fluent, more awkward, more calque-bearing — without reading for direction at all. R08 is by construction the less fluent arm here and was the less fluent arm in step 1, so H-LITERALITY predicts the same arm as H-CARRIAGE in both runs and neither run separates them. This is not a speculative rival: this project has three measurements of it firing. Manufactured oddity out-scored real carriage on perceived-source-carriage at +1.861 against +1.750 (RS-20260807f) and at +2.286 against +1.810 (RS-20260808d), and a word-scramble out-scored the text it permutes by +1.4583 (RS-20260809h). Separating it needs a fourth arm that is fluent and source-ward, which this design does not have. Consequence, registered: see the truncation of P1's licensed conclusion in §6.
• H-PERIOD — registered on critic F-D. Seats choose whichever text sounds more like a genuine nineteenth-century translation. R08 shares 29 seven-grams and 3 twelve-grams with Bates 1909 against R07's 16 and 2, so this heuristic also predicts R08, and this design does not separate it either. It is stated so that it appears in the limits rather than being discovered afterwards.

On AB, the primary, H-CARRIAGE and H-ELEVATION predict opposite arms. That opposition is the whole design, and it is the opposition step 1 could not construct. H-LITERALITY, H-PERIOD and H-DISTANCE all predict the same arm as H-CARRIAGE and are not separated from it here.

The triangulation, registered now — and narrowed on critic F-A. Across the two runs the same three seats answer the same question on two pairs whose elevation ordering is reversed relative to carriage. What that licenses is exactly one exclusion: H-ELEVATION explains step 1's 45 choices and predicts the opposite arm here, so a P1 that fires kills H-ELEVATION as an explanation of step 1. The first draft of this paragraph claimed more — that no single heuristic except carriage predicts both outcomes — and that claim is struck, because H-LITERALITY predicts both. The triangulation is one rival narrower, not a proof of carriage.

6. Registered predictions, gates, failure criteria

G1 (gate, condition AS). With the original on the page, seats choose R08 in ≥ 12 of 15 cells.

G2 (gate, condition EL, the manipulation check). Seats choose R08 as the higher and more elaborate English in ≥ 12 of 15 cells. The design asserts that the source-ward arm is here the elevated one, and until seats say so that assertion rests on the lead's own word-length arithmetic.

What G2 is worth, corrected on critic F-E. A first draft called this gate "the reason the run can mean anything", which inverts the epistemics. G2 asks three literate models which of two texts has the longer words and the longer sentences, at gaps of 9% and 83%, engineered by the hand that wrote both arms. It is expected to pass near-certainly, and its evidentiary value is limited to detecting seat malfunction and to putting the elevation ordering on the record as something other than the lead's arithmetic. It is not independent confirmation of anything and will not be leaned on as such. (Near-certain is not certain: on 2026-08-14 a comparable manipulation check in the sibling arm failed at +0.0833 with a registered +0.25 — which is why the gate is here at all.)

F1 — if G1 OR G2 fails, P1 and P2 are WITHHELD and the run is reported as numbers licensed for nothing. No override is available. (Not a formality: the same clause withheld the sibling arm's headline earlier today.)

P1 (PRIMARY, condition AB). Seats choose R08 in ≥ 12 of 15 cells (exact binomial against p = 0.5: P = 0.0176).

What P1 firing licenses, truncated on critic F-A and registered in this form before dispatch. It licenses exactly one sentence: H-ELEVATION — "choose the less elevated version" — is refuted as an explanation of RS-20260813h's 45 choices. It does not license "carriage is directional on a second language family", because H-LITERALITY and H-PERIOD (§5) predict the same arm and are not separated here. The stronger sentence would need a fourth arm, fluent and source-ward, and that is a new design, not an amendment. The result page may not print the stronger sentence.

P2 (registered opposite tail, condition AB). Seats choose R07 in ≥ 12 of 15 cells. Fires → H-ELEVATION survives and H-CARRIAGE is confined to sources where carriage and plainness coincide, which would mean step 1's headline is an artifact of Danish.

Neither firing is a real and informative outcome, registered here so it cannot afterwards be reported as a boring null: it would mean the judgment is available with the source present (G1) and unavailable without it on Romance material — perceived source direction being unreachable rather than mis-assigned.

G3 (registered decision rule, on critic F-B — whether AB and EL are the same judgment wearing two hats). The design's own mechanism (§2) makes source-ward and elevated the same sites on this material, so a seat could answer AB by reading for elevation and calling it carriage. That would be the confound re-asserting itself, not being reversed, and it must not be adjudicated after the reasons are read.

Mechanical, fixed now. For each of the 15 matched (segment, seat) pairs, the AB stated feature and the EL stated feature are compared. SAME iff, after lower-casing, dropping tokens shorter than four characters and stripping a trailing s/ed/ing, the two ≤ 15-word strings share at least one token. Let AB≡EL be the count of SAME pairs, of 15.

• AB≡EL ≥ 12 → P1's carriage reading is reported as INDISTINGUISHABLE from an elevation reading, and the headline is downgraded to "the seats ordered these two arms the same way under two different questions and gave the same reason for both."
• AB≡EL ≤ 11 → the two questions are reported as separable on this material, and P1 reads as §6 licenses.

Both the count and the 15 pairs of strings are printed whatever the outcome.

P3 (exploratory, no bar). Per-condition agreement between the seats' choices and each of the five hypotheses' predictions, and the ≤ 15-word stated features, printed in full for EL as well as for AB/AS. The EL reasons are the only handle on what the seats read as elevation.

Further failure criteria, fixed now:

• F2 — a seat whose reply cannot be parsed to A/B in > 20% of its 15 cells has all its cells struck and every statistic recomputed on the remaining seats, with the strike reported.
• F3 — a cell returning finish_reason: length is re-dispatched once at double the cap; the cost of both attempts is accumulated. Dead after two attempts → struck.
• F4 — more than 10% of cells struck in any condition → that condition is withheld entirely. Order of operations, on critic F-I: F2 strikes first, and F4 is then evaluated on the post-strike denominator. Without this a single malformed seat (5 of 15 cells, 33%) would void all three conditions, which was not the intent.
• F5 (position lock) — a seat choosing the first-presented version in ≥ 14 of its 15 cells is reported as position-locked and excluded from P1, P2, G1, G2, which are recomputed.
• F6 — the verifier finding any dispatched prompt containing a banned string (§7) voids the affected condition.
• F7 (language-leak robustness, on critic F-G) — see §7. If the post-exclusion subsample has fewer than 12 cells, P1 is reported UNDERPOWERED and no verdict is read either way. If it has 12 or more, P1 must clear its bar on the post-exclusion subsample AND on the full sample; failing either, P1 does not fire. (The critic's own fix was to take the verdict from the post-exclusion recompute alone. Modified, with the reason: the exclusion filter is a post-hoc read of self-reported strings and is noisy, so making it the sole arbiter lets a seat's phrasing decide the headline. The conjunction registered here is strictly harder to satisfy than either figure alone.)

7. Blinding and banned strings

Seats see only the two rendered segments, and in AS the original under the heading ORIGINAL. The verifier asserts that no dispatched prompt contains any of: R06, R07, R08, R25, resistancy, fluency, ennoblement, Venuti, foreignizing, domesticating, Bécquer, Becquer, Miserere, Fitero, Spanish, Castilian, Romance, Latinate, calque, lit-trans, or the name of any regime rule. The AB and EL prompts must be byte-identical outside their two question lines; the AS prompt must equal the AB prompt plus exactly one prepended ORIGINAL block.

A is the first-presented text in every cell — the labels travel with position, not with arm (critic F-J; this is what stage2.py already does, and it is now assertable by the verifier).

Known leak, from step 1 and expected to be worse here. Seats named the source language unprompted in 4 of 15 source-blind cells there. Here the signal is stronger, and the critic (F-G) is right about why: R08's Latinate diction points at the language family, the prompt itself volunteers "nineteenth-century", and the scene — skeleton monks climbing out of a gorge to chant a psalm — is identifiable to a seat with literary training. Every stated feature is scanned for a language name, an author name or a title; the primary is recomputed excluding those cells; both figures are reported; and F7 in §6 fixes in advance what happens when the exclusion bites.

The prompt keeps "nineteenth-century" anyway, because it is E-20260813h's dispatched wording and importing the instrument unchanged is the design's premise. Changing it would make this a different question asked of the same seats, which is the one thing the run must not do.

8. Procedure

1. Stage 1 (analysis/stage1.py), already run and committed with this design: corpora, segments, elaboration predictor. Dependence: analysis/dependence.json. Census: analysis/census.json.
2. Pre-run critic, P4 moonshotai/kimi-k3, given this design whole, cap 8,000 — note (bng): a critic is an open-ended enumeration and S183 truncated one at 6,000 and lost its verdict line for want of $0.11. Findings applied before dispatch; each recorded accepted or overruled, with a reason.
3. Pilot: three cells — one per seat, condition AB, segment S1 — dispatched first, their completion tokens read, and the remaining 42 sized from the measurement (note (bnl)). Pilot cells are kept as real cells.
4. Dispatch the remainder; raw bodies stored per cell under run/.
5. analysis/analyse.py computes every reported number; analysis/verify.py recomputes them from the stored bodies, importing nothing from analyse.py, stage1.py or tools/.

9. Budget

Declared ceiling: $0.75. UTC day 2026-08-14 stands at $2.656570 of $5.00 across four sessions, leaving $2.343430. Worst case built from max_tokens, per note (abc):

Worst case including F3 retries, on critic F-K: a re-dispatch at double the cap on every one of the 45 cells would add ≈ $0.746, for ≈ $1.376 — above the declared ceiling and inside the day's $2.343 headroom. The ceiling stays at $0.75 because a 45-of-45 truncation is not a worst case any observed run approaches (step 1: 0 of 45; S183: 5 of 36 at a tighter answer allowance). If retries carry the run past $0.75 the run stops there and the partial state is reported, which is the rule the ceiling exists to state.

Effort pinned low on all cells. A cap is not a bound on a reasoning seat — S183 billed P3 6,516 reasoning tokens against a max_tokens of 4,000 — so the ceiling carries headroom above the worst case rather than equalling it. Lead translation of 668 Spanish words twice over is $0 and is not ledgered (charter §3, A4), as are the collation, the census, the dependence table and the verifier.

10. Known limits, written before the numbers exist

• One tale, one author, one language pair, one execution of each rule set by one translator, who also chose the segments and wrote both questions. What the run can add to step 1 is a second language family with the confound reversed, not generality.
• Three models sharing a training distribution are not three readers. Convergence among them is what a shared projection looks like, and no design here can tell that apart from perception.
• The device census is translator-coded and it is 24–0. A perfect split is what two opposed rule sets executed by one hand look like; it is an audit column, never a primary, and it is not evidence independent of the translator's intention. Three of R07's sites are partial and are named in census.json.
• The two arms were written back to back by the same hand. The dependence table says the pair is clean at 10 tokens, which is the strongest form the reassurance can take, and it is not the same as independence.
• EL and AB are the same two texts asked two questions. If a seat answers both from one undifferentiated impression, G2 passing does not prove the seats have two separable readings — it proves they order the arms the same way twice. The EL stated features are the only evidence about that and are printed in full.
• Nothing about quality. Tier D remains NOT PASSED; every evaluative word is internal-judgment-only and provisional.
• R07 here is contamination: suspected against the lead's own R08, written minutes earlier. The direction of that bias is toward the arms resembling each other, i.e. against every prediction registered above.

11. Amendments on the pre-run critic

Critic: P4 moonshotai/kimi-k3, effort pinned low, given this design whole plus both arms whole and the Spanish span — a wider brief than step 1's critic had, which is why several findings are about the texts rather than about the prose of the design.

It truncated at a 6,000-token cap in S183 and it truncated again here at 8,000, mid-finding F-G, with 6,085 of the 8,000 spent on hidden reasoning. Note (bng) says a critic is an open-ended enumeration whose cap cannot be guessed; NEXT.md recorded of S183 that "a second $0.11 would have bought a verdict word" and it was not spent. It was spent here: a continuation call, given the design and the truncated critique and asked for F-G's completion, any further findings, and the verdict line only, returned F-G complete, five further findings, and the verdict for $0.035. Both calls are stored (run/critic.json, run/critic_continuation.json) and both costs accumulate.

Verdict: NEEDS-AMENDMENT. Twelve findings, four BLOCKING. Eleven accepted, one accepted with a stated modification, none overruled.

What the critic cost the run, stated plainly: its headline. The design was frozen intending to conclude that source-carriage is directional on a second language family. F-A shows that conclusion is not available from this pair, and the amended P1 may license only the narrower one. The run is dispatched anyway because the narrower sentence is ARM-carriage-direction step 2's registered question — the arm page asks it to separate tracks source-carriage from rejects the more elevated version, and that is exactly what survives.