Repository path: workshop/experiments/E-20260813g-register-quadrants/design.md · rendered 2026-09-09

E-20260813g — the register quadrants: a formal census of six English «Flipperne»

ARM-ennoblement step 2, second half. RS-20260813c §10 wrote a five-item brief. Items 1–4 are repairs to a jury instrument and each of them costs API money the current UTC day does not have (§9). Item 5 is the question the arm was constituted for and it can be bought for almost nothing, because the thing it asks about is formal and formal properties can be counted:

Build the source-ward marked arm (R08 resistancy) on this same tale, so that "marked but fluent" and "marked and estranged" can be put to a source-relative question that a target-only question could not separate (RS-20260813b §7).

The arm is built (T-flipperne-R08-v1, frozen at b5ef9f7 before this design existed). This design is the question.

1. The question

With three constructed poles on the page — plain, raised, source-ward — and three published hands beside them, do the two kinds of marked English separate on countable formal properties; and which quadrant does published translation actually occupy?

RS-20260813b §7 asked a target-only style question of a jury and got 7 of 12 for ornate English matching a foreignizing arm and 0 for a plain one — i.e. on that instrument "marked but fluent" and "marked and estranged" are one thing. This design asks whether they are one thing in the text, or only one thing to that question.

What this is not. It is not a quality judgment, not a jury run, and not a repair of E-20260813c. Every number here is a count over stored text.

2. Materials

Source. Andersen, «Flipperne» (1848), Danish Wikisource copy-text, 756 words, 26 paragraphs, 74 sentence-final marks. Single-witness, uncollated, two known transcription slips (sprøge, seeet).

Six English arms of the whole tale.

(Token counts are [A-Za-z][A-Za-z'-]* matches and differ slightly from the whitespace word counts on the translation artifacts.)

2.1 The dependence table, re-run this session with the new arm in it

tools/dependence_check.py, all fifteen pairs, whole texts:

Four consequences, all binding on what follows:

1. A~BRAK remains the only clean published pair. PAULL is dependence-flagged against both other hands. No pooled three-hand published figure may be claimed, and P1 is stated per hand.
2. R08 is clean against all three published hands despite its declared priming — 7, 10 and 27 shared 7-grams, no 12-gram anywhere, longest run 11. This is reported, not assumed, and it is the strongest single fact in the table.
3. R08 and R06 are dependence-flagged against each other at 48 / 8 / 2 / 16 — two renderings by the same lead, three sessions apart, matching each other far above what either matches any published hand. This is note (bhb) on a third occasion and it constrains nothing in this design, because no prediction here treats R06 and R08 as independent observations.
4. R06 stays disqualified as an independent hand (declared session priming, 19-token run against A), and is used here only as the plain pole against which the two rule sets are read.

2.2 The BRAK reconstruction, enumerated, because a census is a count over characters

The only reachable Brækstad text is an Internet Archive OCR of a two-column page, read across the columns, so speech and tag are interleaved out of order in three blocks. The arm was reconstructed by the lead. Every intervention:

• Three column de-interleavings, at Danish paragraphs 8–10, 17–19 and 16–17. These reorder existing OCR tokens and add none.
• One inserted word: iron in "when the hot [iron] was passed over him" — the noun is absent from the OCR between the column fragments when the / hot / was passed / over him. This is the only invented token in 885.
• Character-level OCR repairs, exhaustively: Vou→You (×2), N\)w→Now, oft"→off, Hut→But, name r"→name?; and eleven hyphen-linebreak joins (waist- band, ornament- al, rail- way, car- riages, re- peated, pro- posal, boot- jack ×2, and three others).

F4 (§7) is registered against this: at more than three invented tokens the arm is dropped. It stands at one.

3. The two axes, and what they actually are

They are operationalisations of two constructed rule sets, not of Berman and Venuti in general, and the design says so before it measures anything. Axis E is the direction R25 was written in (R25 applies raised diction, expansion, clarification and decorous exclamation together — its own limitations section calls it a compound arm). Axis S is the direction R08 was written in (R08 R2 clause order, R6 calque, R1 discontinuity). The published hands signed up to neither. A floor score on Axis S is therefore a description of published practice and not a deficiency, and no sentence in the result page may read it as one.

Axis E — elevation, "marked but fluent"

• E1 Romance-origin content-word rate. Every one of the 758 union types across the six arms is classified G (Germanic/Old English/Norse), R (Latin/French/Romance) or O (other or undecidable) by a panel seat, blind to arm — the type list is alphabetical and carries no provenance. Rate = R-tokens / (R+G tokens), function words excluded by a frozen stoplist.
• E4 Speech-tag variation rate. Andersen uses «sagde» 20 times and «spurgte» once: variation rate 1/21 = 0.0476. For each arm, 1 − (plain said tags / all reporting-verb tags), over a frozen verb list.
• E3 Expansion ratio = arm tokens / 756. Reported descriptively and EXCLUDED from the composite — see §5.

Axis S — source-ward markedness, "marked and estranged"

• S1 Compound-calque retention, over 16 Danish compounds frozen in sites.json from the Danish alone: Bohave, Støvleknægt, Redekam, Strømpebaand, Livbaand, Strygebrædt, Strygejern, Dampkjædel, Papirs-Sax, Klude-Selskab, Pralhans, Kjærestesorg, Vandballe, Enkefrue, mandfolkeagtig, Grevskab. Each site carries the two Danish elements with neutral dictionary glosses. Coded by two panel seats, blind to arm (§4): does the English at this site consist of two elements corresponding to the two Danish elements?
• S2 Fronting-inversion carriage, over 6 frozen Danish clauses where a fronted element puts the finite verb before its subject: «nu har jeg aldrig seeet», «der kom det varme Jern», «saa blev de tagne af Vadsken», «saa kom Flipperne i Kasse», «Aldrig glemmer jeg min første Kjæreste», «Og det blev de». Coded by the lead, every cell printed in the result page.
• S3 Number carriage, 0–2: (a) is the garment first named with a plural noun, as «Flipperne» is; (b) is it later carried by a masculine singular pronoun, as Andersen's «han» is. Coded by the lead, all twelve cells printed.
• S4 Period carriage = |arm sentence-final marks − 74|. Reported descriptively and EXCLUDED from the composite — see §5.

4. Procedure

Stage 1 — the mechanical census, $0. analysis/census.py computes E3, E4, S2, S3, S4 and the token/type inventory from the six stored texts. No judgment enters except S2 and S3, whose every cell is printed.

Stage 2 — the independent classifications, 3 calls.

Blinding. The six arms are presented to CMP-1/CMP-2 as V1…V6, scrambled by sha256("E-20260813g" | seat) mod 720, so no arm sits behind a fixed label and the two seats see different label maps. No provenance, no translator name, no date, no regime name, and none of the words Berman, Venuti, ennoble, foreignizing, resistancy, regime, published, Gutenberg, Andersen, lead appears in any prompt. Verified by analysis/verify.py.

Re-dispatch guard: a body that fails to parse is re-dispatched once, irrespective of finish_reason (note (bmz)); costs accumulate across attempts, not overwrite (the RS-20260813c §9 repair).

Stage 3 — the lead's audit of LAT. The lead independently classifies a 60-type sample fixed by sha256("E-20260813g-audit" | type) rank order, written before LAT's body is opened, and agreement is G3.

5. What was already seen when these predictions were registered — declared, not glossed

E3 and S4 were computed before this design was written, because token counts and sentence counts were needed to file T-flipperne-R08-v1 and to check the BRAK reconstruction. They are therefore excluded from both composites and reported descriptively only. E1, E4, S1, S2 and S3 had not been computed for any arm when the predictions below were registered.

Composite definition. Each measure is ranked across the six arms (1 = most elevated on Axis E, 1 = most source-ward on Axis S); the composite is the mean rank. E = mean rank over E1, E4. S = mean rank over S1, S2, S3.

6. Gates, run before any prediction is read

• G1 — Axis E construction check. R25 must have the best (lowest) Axis E composite of all six arms. R25 was written to raise the register; if the axis cannot see that, it cannot see elevation.
• G2 — Axis S construction check. R08 must have the best Axis S composite of all six arms. Same logic, other pole.
• G3 — E1 reproducibility. Lead-vs-LAT agreement on the 60-type audit sample ≥ 0.85.
• G4 — S1 inter-coder agreement. CMP-1 vs CMP-2 agreement over the 96 cells ≥ 0.80.

7. Failure criteria, registered

• F1 — G1 fails → every Axis E prediction is withheld, the axis is reported as a failed instrument, and P1's E-half goes with it.
• F2 — G2 fails → every Axis S prediction is withheld.
• F3 — G3 < 0.85 → E1 is struck from the composite; Axis E then rests on E4 alone and P1's E-half is demoted to exploratory, since a one-measure axis is not a composite.
• F4 — more than three invented tokens in the BRAK reconstruction → BRAK is dropped; since A~BRAK is the only clean published pair, P1 is then demoted to exploratory and cannot be primary on A alone.
• F5 — G4 < 0.80 → S1 is struck from the Axis S composite, which then rests on S2, S3, and P2 is demoted to exploratory.
• F6 — any of the three stage-2 bodies dead after one re-dispatch → the measure that body supplies is struck by the same rule as its gate, and the run reports which.
• No override of F1–F6 is available to this session. RS-20260813b §6 and RS-20260813c §5 both refused overrides that would have rescued their own headlines; that is the precedent.

8. Predictions, registered before any of E1, E4, S1, S2, S3 was computed

• P1 — the quadrant claim. PRIMARY, on the clean published pair A and BRAK only. Each of A and BRAK has an Axis E composite better than R06's and an Axis S composite worse than R06's. In words: published translation of this tale is raised above the plain rendering and is at the same time less source-ward than a plain rendering made under no rule at all. PAULL is reported alongside as a dependence-flagged sensitivity, never as part of the primary.
• P2 — the separation claim. SECONDARY, and it is RS-20260813b §7's question. R25 beats R08 on Axis E and R08 beats R25 on Axis S, each gap ≥ 1.5 rank points. The outcome of interest is symmetrical: if it holds, the two kinds of marked English are formally distinct objects and the target-only style question was the wrong instrument; if it fails, the two are not separable even by counting, which is a much stronger result than a jury null.
• P3 — where the constructed poles sit relative to published practice. EXPLORATORY. The gap R25 − (mean of A, BRAK) on Axis E, and R08 − (mean of A, BRAK) on Axis S, reported as rank distances with the note that R25 and R08 are single executions of single rule sets by one translator.
• P4 — the register-carriage obstacle. EXPLORATORY, and it is the arm's completion criterion. If P1 holds, framework/v0.2 §8 Q-e gains a stated obstacle: there is no published-practice precedent to borrow for a source-ward register recommendation, because published practice does not occupy that quadrant. This is written as an obstacle, not as a recommendation.

9. Money

Pre-flight, priced from config/models.md at the caps actually sent (note (abc): the worst case is built from the cap, never from an expected length), with this project's standing 4× routing surcharge.

Day headroom at design time: $0.950602566 of $5.00 (2026-08-13 UTC, six prior sessions at $4.049397434). The ceiling fits with $0.25 to spare. The stage boundary is stage 2: if the critic's cost leaves less than $0.60, CMP-2 is dropped, F5 fires by construction, and the run says so.

T-flipperne-R08-v1 — 756 Danish words rendered whole — is $0 and is not ledgered (charter §3, A4), as are stage 1, stage 3, the BRAK reconstruction, the dependence table and the verifier.

10. Verification

analysis/verify.py, importing nothing from tools/, recomputes every reported number from the stored texts and stored bodies rather than from the runner's bookkeeping, and asserts: every prompt is free of the eleven banned provenance strings; the V1…V6 scramble is a bijection for each seat and the two maps differ; every per-cell code matches its stored body (RS-20260813b §8's repair); every rank, composite and rate recomputes; and the six stored arm texts hash-match the committed artifacts for the three lead arms.

11. What this run cannot establish

• Nothing about quality. Position on a formal axis is not a verdict. Tier D remains NOT PASSED; every evaluative word is internal-judgment-only and provisional.
• Nothing about three independent published hands — one clean pair, one flagged third.
• Nothing about Danish, or about Andersen, generally. One tale, one author, one language pair.
• Nothing that attributes R25's or R08's position to any single rule. Both are compound rule sets executed once by one translator who also designed this instrument.
• Nothing about what readers perceive. A formal census and a jury measure different objects; RS-20260813b §7's question was about perception and this is not an answer to it in its own terms. What it can show is whether there is a formal difference available to be perceived.
• Nothing that escapes the lead's authorship of three arms. The mitigations are: the two judgment-laden measures (E1, S1) are coded by seats blind to arm; every lead-coded cell (S2, S3) is printed; and the primary rests on the two published hands, which were written by people who had never heard of this instrument.

12. Amendment before dispatch: the pre-run critic's findings

Critic: moonshotai/kimi-k3 (P4), no role in this run's coding seats (P1, P3). Verdict PROCEED-WITH-AMENDMENT, 8 findings, 2 BLOCKING. Seven accepted, one overruled with a written reason. Nothing had been dispatched. Cost $0.081607200; raw body run/critic.json.

Amended definitions, in force from here:

• Ranks are midranks, on every measure, both axes.
• S1, S2, S3 are blind-coded by P1 and P3. A cell scores 1 iff both seats code it 1 (conjunctive, the conservative direction for Axis S); the disjunctive variant is reported as a sensitivity, and the lead's coding as a never-primary third column.
• G4 is inter-seat agreement over all 144 cells, bar unchanged at 0.80.
• P2 is the strict-separation form above.
• F4 is no longer a failure criterion; it is an assertion in verify.py.
• Caps revised for the added cells: CMP-1/CMP-2 cap 4,000 with compact array output ({"C01": ["V2","V5"], …} — only the labels scoring 1), LAT cap 6,000. Re-priced worst case: LAT $0.038 + CMP-1 $0.033 + CMP-2 $0.042 = $0.113, × 4 routing surcharge = $0.452, plus the $0.081607200 already spent on this critic. Revised declared ceiling $0.55, against $0.950602566 of day headroom.