Repository path: workshop/experiments/E-20260815c-footing-direction/design.md · rendered 2026-09-09

E-20260815c — when a Japanese source says who is above whom, does the English say the same person?

ARM-footing-price step 2. Frozen before any call is dispatched. Every number below is recomputed by verify.py on an independent path.

1. Question

RS-20260814h §6 found that a maximally footing-marked rendering delivers a relative grading whose direction is not the source's: the same rendering of the same scene pointed at the woman in one paragraph and at the man in the next, following the local density of her ladyship and his lordship rather than anything in the Japanese. That was one artificial regime written by the hand that then wanted it to fail. framework/v0.2 §10.5 item 6 states it as a claim about what a translator is doing to a text, and §10.6 says the remedy would have to be "a rule about whose footing a paragraph is allowed to mark."

Does the direction problem hold for hands that were never asked to carry anything — two published human translators of the same Japanese? And can a rule fix it?

The source chosen is one whose grading is known site by site, because this project already censused it: 『源氏物語』 ch. 15 「蓬生」 §§2‑3–2‑4, 67 sites with glosses, frozen 2026-08-14 in E-20260814b, audited OK at 64 of 67 by an independent seat. Twenty of its thirty paragraphs are graded by their own marks, and seventeen of the twenty put the same person on top. Three do not, and those three are where this design does most of its work.

2. The subject-rule sentence (continue-prompt.md §4.5)

This unit teaches whether the social relation a source states grammatically arrives in English at all, whether what arrives instead is set by where the English happens to put its rank words, and whether a translator can write to a rule that changes the answer. That is a sentence about translating literature. The three model seats are the instrument, not the subject: nothing here asks whether the panel is reproducible or whether a past figure holds. A null is a finding about the translations.

3. Materials

3.1 The source's grading, S

From sites.tsv, by a rule fixed here: each of the 67 sites is assigned (party raised, party lowered) from its frozen gloss (predictors.py, table printed in full). Sites that grade a character against the narrator are excluded — the item asks the judge about people in the passage, and the narrator is not one of them. A segment's S is the party raised by the most inter-character sites in it; a tie between two elevated parties is broken by a site that explicitly places one above the other (site 35 puts the Princess above Genji), and by nothing else.

3.2 The segments

Ten, defined on the Japanese and identical for every hand; each hand's English is cut by literal anchor strings committed in extract.py and printed. The segmentation is exhaustive and non-overlapping — every hand's segment word counts sum to its whole-span count, which verify.py re-checks.

A cell below 25 words is NO-RENDERING and is excluded, declared here before dispatch. Exactly one cell falls: SUEM/G10 at 13 words — Suematsu cuts the old women's scene entirely. 39 live cells.

3.3 The four hands

R33 is excluded from the registered primary by construction — it was built to make the two predictors agree. R06 is excluded from the primary too, and enters one pre-registered secondary, reported unconditionally: it is the lead's hand, and the lead also wrote the S coding.

3.4 The competing predictor, E

For each cell, the party carrying the most raising rank tokens in that hand's English of that segment. A rank token names or implies a station — princess, prince, ladyship, excellency, highness, chancellor's wife, madam, mistress — or is deferential periphrasis from the vocabulary RS-20260814b §4 already published (deign, vouchsafe, be pleased to, see fit, bestow). A token that names a kinship or a job without station — aunt, visitor, mother, nurse, maid — is not a rank token. Lowering tokens are recorded and printed but do not score: a speaker's self-humbling raises her hearer, and which hearer is a judgment rather than a count.

Three assignment rules, stated on the critic's F4 because the first draft applied them without saying so: (i) an elevating verb raises its subject (bestow, vouchsafe); (ii) a humbling verb raises its object (be presented to, come asking after); (iii) a formula of deference addressed to somebody raises the addressee (deign, be pleased to, see fit, much obliged, Madam). Narratorial marking and in-speech marking both count, because the item asks what the wording marks and the wording marks through both channels. Sensitivity: dropping rule (iii) entirely — the class the critic challenged hardest — moves no cell's E except by removing AU from R33/G4, which stays PR either way.

A defect found while building the controls and fixed before dispatch: the first token table matched the Princess case-sensitively, so a sentence-initial The Princess matched nothing. Four cells' E moved (SUEM G4, WALEY G5, R06 G1, R06 G4). The table is now matched case-insensitively and every match is printed in predictors.log.

E = NONE where a cell has no raising token; E = TIE where two or more parties tie. TIE cells make no prediction and are excluded from the head-to-head, declared here.

This table is the lead's, written with the texts in front of it, and that is a limit (§8.4). It is frozen before dispatch and every match is printed, so what it cannot be is tuned to an outcome: the outcome does not exist yet.

4. Procedure

1. python3 extract.py — cuts the 40 cells and checks the segmentation is exhaustive.
2. python3 predictors.py — computes and prints S and E and every site and token behind them.
3. python3 build.py — writes cells.json; assigns each cell an opaque pid from sha256("E-20260815c|<seg>|<hand>"), and sorts by it so hands are interleaved and no seat meets one grouped. Adds the three gate duplicates (§5).
4. python3 run_critic.py — one non-Anthropic adversarial pre-run critic, P4 moonshotai/kimi-k3, given the design and every passage. P4 grades nothing, so no comparison in this run has to be reported with and without its author. Note (bny): an explicit reasoning: {"max_tokens": 2000} inside max_tokens: 9000, never an effort pin. The run does not proceed on a critic that returned no verdict line. Findings and what was done with each go into §10 before dispatch.
5. python3 run.py --stage gate — the gate, §5, 18 bodies. The main dispatch does not run if the gate fails.
6. python3 run.py --stage main — 39 cells × 3 seats = 117 bodies, one cell per request, temperature: 0, dispatched one at a time in a fixed recorded order. Judgment is not parallelised. Every body is appended to bodies.jsonl as it returns and a re-invocation resumes — note (bnx). Note (boe) is fixed in this runner: a re-dispatched body's discarded attempt keeps its usage and is ledgered.
7. python3 analyse.py → python3 verify.py, which imports nothing from analyse.py, recomputes every reported number, and runs mutation tests against the per-cell, per-seat answers, note (bio).

4.1 Seats

P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, P3 x-ai/grok-4.5, resolved from config/models.md at run time and logged as provenance. P5 is excluded on any task shape, note (bne).

4.2 The prompt, byte-identical across every cell but for the passage and its label

You are reading a passage from the middle of an English translation of a work of
fiction. You are not told what the original language is, who translated it, or
where the passage comes from.

Three women figure in the story: the lady who lives in the house, an older female
relative of hers who is visiting and is about to move away to the provinces, and a
younger woman who has served in the house for many years. Other people may be
mentioned.

PASSAGE {pid}
------
{passage}
------

Answer with exactly four lines and nothing else.

REL: an integer 1-7. Setting aside what the passage tells you about who these
people are and what happens to them, how strongly does the WORDING ITSELF mark one
person in the passage as socially above another - deference, rank, station, one
person raised or lowered relative to another by the choice of words rather than by
the events described? 7 = the wording marks it insistently; 1 = the wording marks
it not at all.

WHO: one of A, B, C, D, E, F. Which person does the WORDING mark as being above the
others?
A = the lady who lives in the house
B = the visiting relative who is moving away
C = the younger woman who has served in the house
D = somebody else mentioned in the passage
E = the wording marks no one as above another
F = cannot tell

NAME: if you answered D, the words the passage uses for that person; otherwise
write NONE.

WHY: one sentence of at most 25 words, quoting the words that decided WHO, or
saying that there were none.

The cast line is constant across all 42 passages, so it cannot explain any difference between cells; it is there because a judge shown one segment of a long story cannot otherwise tell three women apart. It contains no rank word, deliberately.

REL is not a sense on wiki/goodness-senses.md and is not proposed as one. It is the descriptive item E-20260814h §4.2 used, carried over verbatim so the two runs are comparable, and it is deliberately not perceived-source-carriage, whose own entry records a 0.40–0.60 false-positive rate on manufactured oddity. Nothing here is a quality judgment and none is licensed. Tier D is NOT PASSED; every score is internal-judgment-only and provisional.

naturalness is not scored in this run. RS-20260814h priced that already and this design has nothing to add to it; asking would invite the reading that R33 is being recommended.

5. The gate — validity first, then repeatability

RS-20260814h §5 was withheld because one seat scored the same text REL 4 under one label and REL 1 under another, at temperature: 0. That is why this design's primary is a categorical item. But the pre-run critic's BLOCKING F3 is right that a duplicate-label check cannot fail for the failure mode that matters: at temperature: 0 on byte-identical text, a seat that always answers A passes it 9 of 9 while measuring nothing. So the gate has three conditions and the first two are about validity.

Two manufactured instrument controls, written for this purpose and not translations of anything: the same 70-word scene twice, matched sentence for sentence, differing only in whether the visiting relative carries rank material. CTL.POS marks her her ladyship four times and has the lady of the house say she is much obliged; CTL.NEG says the same things with no rank word in it. Both are printed in cells.json.

Three duplicated cells, one per hand for SUEM, WALEY and R33, chosen mechanically as the live cell whose sha256 is lowest for that hand, presented under a second pid.

C1 — CTL.POS: at least 2 of 3 seats answer B, the visiting relative. This is the condition that tests whether the item can name a party other than the salient default, which is what P1 requires of it. C2 — CTL.NEG: at least 2 of 3 seats answer E or F. This tests that the item is not simply naming whoever is most prominent in the scene. C3 — label stability: the WHO answer is identical under both labels in at least 7 of the 9 (text, seat) pairs.

All three must hold. If any fails, every registered claim below is WITHHELD and reported as withheld, not as a null, exactly as RS-20260814h withheld its middle rung. REL's label stability is reported either way, as a number, because that is the figure RS-20260814h could only report anecdotally.

6. Predictions, registered before dispatch

Rewritten after the pre-run critic returned NEEDS-AMENDMENT, before any grading call was dispatched. The first draft made the horse race the primary on seven cells, four of which had E = NONE; the critic's BLOCKING F1 showed that "the reader follows E" on a NONE cell means only "the judges said nobody is marked", which is not a claim about direction and is easiest to get in the hand that compresses. Its BLOCKING F2 showed the 0.5 binomial null was asserted and never derived. The primary is therefore the reversals, tested against a base rate the run measures for itself, and the horse race is split in two and demoted to description.

Scoring. Per cell the modal WHO over the three seats: a strict plurality. Three different answers is SPLIT, and a SPLIT cell matches no predictor. Options map to parties as A→PR, B→AU, C→JJ, D→whoever NAME gives, E→NONE, F→UNCLEAR. For the three duplicated cells the original pid's answers are the ones used everywhere below; the duplicate is gate material only.

P1 — PRIMARY. Where the source reverses direction, does the English go with it?

Twenty of the thirty source paragraphs are graded by their own marks and seventeen put the Princess on top. The three that do not fall in G4 (the visiting relative), G8 (Jijū's dead mother) and G10 (Jijū). Over the three hands not written to this rule — SUEM, WALEY, R06 — that is 8 reversal cells (SUEM/G10 is excluded at 13 words) against 21 non-reversal cells.

P1: the modal WHO names the source's party at a lower rate on the reversal cells than on the non-reversal cells. One-sided Fisher exact on the 2×2, enumerated in full.

The null is the run's own non-reversal rate, which is why no coin-flip has to be assumed. The comparison is not clean of salience — on non-reversal segments the source's party is also the most prominent character — and that is exactly the alternative C1 and C2 exist to exclude: a seat that cannot name anybody but the lady of the house fails CTL.POS and nothing here is reported. A registered floor: if the modal WHO is A in more than 32 of the 39 live cells, the item is reported as measuring salience and P1 is withheld whatever its P.

P2 — the horse race, named against named. Descriptive, no P value.

The cells where E names a party and that party differs from S's: SUEM G3 (E=GJ), SUEM G4 (E=PR, S=AU), SUEM G8 (E=PR, S=NU), WALEY G10 (E=PR, S=JJ) — four published cells, and R06 G4 as a fifth reported separately.

P2: the modal WHO follows E rather than S. Reported as a count out of 4 (and 5) with no significance test, on the critic's F2: at this n a binomial adds nothing an n=4 pilot can support. Cells matching neither are counted as matching neither.

P3 — the other half of the split: does the grading simply vanish?

The cells where E = NONE and S names a party: SUEM G2, WALEY G4, WALEY G8 published, and R06 G2, G6, G7, G8, G9, G10 — 9 cells.

P3: the modal WHO is E ("the wording marks no one as above another") rather than the source's party. Descriptive, a rate out of 9, published and lead reported separately. This is a different claim from P2 and the two are never pooled.

P4 — the rule, and the sharpest thing this design asks

R33 marks per paragraph. Its three reversal paragraphs sit differently: ¶19 (G8) has unmarked poem lines on one side and nothing on the other; ¶7 (G4) and ¶29 (G10) are flanked on both sides by paragraphs the source marks for somebody else, and R33 marks them accordingly. The regime's own page predicts, before any call, that a per-paragraph rule cannot control what a longer stretch reads like.

P4: R33's modal WHO is D-naming-the-dead-mother at G8, and is NOT B at G4 and NOT C at G10. Three cells, each a separate registered call; the interesting outcome is the pattern, and a mixed result is reported as mixed.

P5 — the manipulation check, registered as near-guaranteed and never reported as a finding

P5: R33's modal WHO matches S at more segments than R06's does. R33 was built to make E = S; a check, not a result, and RS-20260814h §4's treatment of M1 governs the write-up.

P6 — descriptive, not a test

REL per hand; the share of cells in which each hand's seats name any individual at all; the distribution of modal WHO over all 39 live cells.

7. Failure criteria, fixed before the run

1. The gate (§5) fails → P1, P1′, P3, P4 withheld; P2 and P5 reported with the gate's number attached to every sentence.
2. More than 4 of the 39 live cells return SPLIT → P1 and P1′ are reported but the design declares itself underpowered on the item and says so in the result's one-sentence answer.
3. More than 3 bodies die unrecoverably → the run is reported as incomplete and no primary is claimed.
4. A seat answers D and gives a NAME that is not in the passage → that cell's seat answer is UNCLEAR, decided here rather than after seeing it.
5. The modal WHO is A in more than 32 of the 39 live cells → the item is reported as measuring salience and P1 is withheld, whatever its P (§6 P1).

8. What this design cannot establish, written before it is run

1. Three model seats are not readers. Charter §4. The word reader appears in no claim this run licenses, and the framework edit it licenses carries the same constraint.
2. One work, one chapter, one language pair, two published hands, both pre-1930. Period is not separated from anything, exactly as RS-20260814b §9 records.
3. The hands differ enormously in length — Suematsu renders the span in 651 words and Waley in 1,441 — so a cell's chance of containing any rank material at all is not matched across hands. The primary is a within-cell contest between two predictors and does not compare hands, but NONE cells are more likely in the compressing hand and the result must say so.
4. The E table and the S coding are both the lead's, written with the texts open. Both are frozen before dispatch and printed in full, so neither can have been fitted to an outcome that does not yet exist; neither is thereby independent.
5. R33 is one hypothesis-aware hand and is untested in the sense of D-20260724-04. P3 and P4 license nothing about what English does.
6. n is small. Seven cells for the primary; the smallest attainable one-sided P is 0.0078 and a unanimous outcome is the only route below 0.05. This is a provisional internal pilot.
7. Each of the three reversals rests on exactly one site — 25 at G4, 58 at G8, 66 at G10 — and the result's one-sentence answer must say so. RS-20260814b §6's independent gloss audit passed all three, and none of them is among the two flagged cruxes; but a single gloss is a single gloss. G5's S is a 6–6 tie between the Princess and Genji broken by site 35 alone, and G5 is a non-reversal cell contributing to P1's base rate.
8. WHO is a forced choice among six options and the cast line names three of them. A seat that defaults to the most salient character will look like it is following S, since S is PR at seven of ten segments — which biases the run against P1, and is the conservative direction.

9. Pre-flight cost estimate

Panel prices from config/models.md, re-read 2026-08-14. Per body: prompt ≈ 200 (instructions) + up to 350 (passage) ≈ 550 tokens in; the answer is four short lines, but every seat may spend hidden reasoning, so the worst case is built from the cap the request permits — note (abc) — at max_tokens: 900 with reasoning: {"max_tokens": 400}.

Day headroom at the time of writing: $5.00 − $2.538366 = $2.461634. The ceiling is 37% of it. A run that overruns the ceiling stops and reports what it has.

10. Pre-run critic findings

P4 moonshotai/kimi-k3, one call, 14,801 in / 2,968 out, finish_reason: stop, $0.050216. Given this design and all 42 passages before a single grading call was dispatched. Verdict NEEDS-AMENDMENT, seven findings, three BLOCKING. All seven accepted; none overruled. The amendments below were made before dispatch and the design was re-frozen.

One defect the critic did not find and the amendment work did: the token table matched the Princess case-sensitively. Fixed before dispatch; §3.4 records which four cells moved.

11. Addendum — written after the gate fired, before a single main body was bought

The gate failed on C2 and every registered claim in §6 is WITHHELD. Nothing below un-withholds anything. This section is dated, is declared post-gate, and exists because the failure told the run something it did not know.

11.1 What the gate returned

C3 passed at 9 of 9; REL was exactly equal in 5 of 9 label pairs, mean |Δ| 0.556, max 2.

Every seat named the same thing on CTL.NEG, and it was not a word: "The household stood back and the hostess came forward and thanked her"; "the phrase the household stood back as she came out marks deference"; "household stood back … thanked her for the visit". The two control passages differ in rank words and they do not differ ONLY in rank words: both of them narrate a woman being deferred to. §5 claimed otherwise and §5 was wrong. The control is a defective control, and the defect is visible in the seats' own reasons rather than inferred from the outcome.

11.2 The diagnostic, registered before it was dispatched

A second matched pair, same three women, no behavioural deference anywhere in either member — they sit over a fire and talk about the roads — differing only in whether the visitor is called her ladyship. Registered in advance:

• D1 — CTL2.POS draws B from ≥ 2 of 3 seats and CTL2.NEG draws E or F from ≥ 2 of 3 → the item reads wording when events are flat, and C2's failure is the control's defect.
• D2 — otherwise → the item does not separate wording from event at all.

D1 HOLDS, unanimously and at the ends of the scale. CTL2.POS: B, B, B at REL 7, 6, 6, every reason quoting her ladyship. CTL2.NEG: E, E, E at REL 1, 1, 1 — "No honorifics, titles, or deferential wording mark anyone as socially above another."

11.3 What the two pairs jointly say, and it is the run's first real finding

With the events held flat, four rank words take the item from REL 1 to REL 6.3 and from "nobody" to a named person, in three seats out of three. With deferential events narrated underneath, the same four rank words move REL by +2.33 and move the named person NOT AT ALL.

Narrated behaviour dominates wording on the direction of a perceived social grading, and wording survives only on its intensity. That bears directly on framework/v0.2 §10.5 item 6, whose direction claim rests on this item run without any such control, and the framework edit this run licenses says so.

11.4 The one statistic the demonstrated confound cannot touch, registered now as post-hoc

P7 — POST-HOC, registered after the gate failed and before the main dispatch, and reported under that label wherever it appears. At a fixed segment the four hands narrate the same events of the same story. Any disagreement between two hands' modal WHO at the same segment is therefore attributable to the wording, since the events are held constant by construction.

P7: at one or more of the ten segments the four hands' modal WHO answers are not all the same. Reported as a count of segments and a per-segment table, descriptively, with no significance test — the n is ten and the claim is existential.

Its weak point, stated now: Suematsu compresses 1,841 characters into 651 English words and paraphrases rather than cuts, so his cells hold the story constant but not always the narrated incident. P7 is therefore reported for all four hands and for the three that render the span at length (WALEY, R06, R33), where the events-constant premise is strong.

11.5 The main dispatch is bought anyway, and why

108 bodies at an estimated $0.55, inside the declared $0.90 ceiling. It buys P6 and P7, and it buys the descriptive table a successor needs in order to build the design this one could not: the withheld primaries are withheld, and the data behind them is preserved and printed so that nobody has to re-buy it.