Repository path: workshop/experiments/E-20260904-french-in-russian/design-v2.md · rendered 2026-09-09

E-20260904-french-in-russian — design v2, rebuilt on the pre-run critics' findings, still before any English was fetched

Frozen 2026-09-04 (S243). Two independent pre-run critic seats (openai/gpt-5.6-terra, x-ai/grok-4.5), each shown v1 and the frozen locus list and nothing else, both returned NEEDS-REDESIGN: 31 findings, 11 BLOCKING, raw bodies at runs/RS-20260904-french-in-russian/raw/C1.json, C2.json, log at critic.log.

No volume had been fetched when v1 was written and none had been fetched when this was. The predictions below are registered against material still unopened. v1 is kept unedited; this file says what changed and why, finding by finding, in §1. The registered set below replaces v1's entirely — v1's P1–P6 are void as tests and are not reported as though they had been run.

1. What the critics found and what was done about it

Both seats converged on five defects. All five are granted.

(a) The typographic instrument is not one instrument (C1 A1,A2,A3; C2 A1,A2,A14, all BLOCKING). Three hands' italic would come from ABBYY's inference over a scan; the fourth's from Project Gutenberg's own _underscore_ markup. A pooled "distinguished share" over those four measures witness technology. Worse, v1's availability control — does the volume use italic anywhere — cannot separate a hand who marks nothing from a layer that recovered nothing at the loci.

Granted, and the whole typographic dimension leaves the registered set. Only MAU, whose type comes from markup rather than from OCR inference, carries a registered typographic prediction (P1t). The three ABBYY hands' marking is reported per hand, descriptively, with the instrument named on the figure and recall declared unverified, and a stated image spot-check; no rate crosses instruments and no pooled typographic figure is reported at all. The specific remedy both seats asked for — recover type from page images for all 152 cells — is overruled in writing: it is a session's work by itself, and the arm's question does not turn on type. §7.50.2 already settled the type question on Dante, on a corpus built for it.

(b) A non-blind prediction cannot sit in a registered set (C1 A4 BLOCKING; C2 A6). v1's P3 predicted that englishing would be the majority move, while declaring that the lead already believed this of two of the four hands.

Granted. P3 is removed from the registered set entirely and appears in §5 as a declared prior, reported descriptively. Nothing is scored against it and no other prediction shares a denominator with it.

(c) P4 was circular and length-confounded (C1 A5,A6; C2 A3; BLOCKING on both). The same reader classified WHOLE/INTO/TAG and then bet on that classification; and a whole French speech is longer as well as different in kind.

Granted twice. The classification rule is now written out (§3) rather than implied, and P4 is registered in a length-controlled form: the raw contrast is reported, and the prediction is scored on a length-matched band built from the word counts already frozen in loci-frozen.md. If either arm of the matched band holds fewer than 4 loci, P4 is reported as exploratory and is not counted as a test.

(d) P5 had no denominator and predicted zero (C1 A7,A8; C2 A4 BLOCKING).

Granted. P5 now has a denominator (retained mixed loci), an evaluability gate (a hand counts only with ≥ 2 retained mixed loci), and a written list of the marks that would count as preserving the second-order switch (§5). And the objection "is that outcome even physically available to a translator working into English?" is answered by material this session already holds: the lead's own frozen limb makes the move at all five mixed loci, rendering the Russian-inside-French as English roman inside italic French. The move is available. Whether published hands take it is the open question.

(e) P6 is not answerable from four dependent books (C1 A9 BLOCKING; C2 A7).

Granted. P6 is dropped as a prediction. Bell's provenance is reported as bibliography and her retention rate descriptively, with no causal claim and no rank test.

Three further findings changed the design rather than only its scope:

(f) gloss does not mean what v1 said it meant (C1 A11; C2 A8). Every class-A locus was selected because Tolstoy glosses it. A hand who prints no English rendering is not declining to explain French in general — it is declining an apparatus the author built.

Granted, and it is the best finding of the pass. gloss is re-coded against the author's footnote (§4), and the arm's new primary prediction P2 is about exactly this, which v1 did not have.

(g) Pooled rates over four non-independent books and one scene (C1 A13 BLOCKING; C2 A9).

Granted. Every registered prediction below is per hand. Pooled figures appear on the result page only as corpus description, explicitly non-inferential, never as a threshold.

(h) The set of hands was not actually frozen (C1 A10; C2 A13) — v1 listed DOL as "to be confirmed from the title page".

Granted. The hand is identified and frozen from the volume's own front matter before any locus is located in it; if it cannot be identified, or is not an independent translation, the volume is dropped and the drop is reported. §2.

(i) One coder, who is also the predictor and one of the translators (C1 A15; C2 A5 BLOCKING).

Granted in the form the material allows. A blind second coder is bought on the keep / english / delete dimension — the dimension every registered prediction turns on and the one a reader can settle from a quoted passage: seats see the French locus and the hand's English around it, with the hand's identity masked and no prediction shown, and code one of five values. Agreement and every disagreement are reported, and a registered prediction whose outcome flips under the second coder's calls is reported as unresolved, not as passed. Typography is not sent out: it is not recoverable from a quoted passage.

Accepted and applied without further comment: C1 A12 (all claims scoped to French passages Tolstoy glosses), C1 A14, C1 A16 / C2 A10 (the merge rule is written out in §3 and the full Russian context of all five merged loci is added to loci-frozen.md), C1 A17 (boundary rules, §4), C2 A11 (classification rule written out), C2 A12 (class B split by kind).

2. Hands — frozen before any locus is located

A volume whose front matter shows it to be a reprint of another hand in the set, or an abridgement missing more than 6 of the 38 loci, is dropped, and the drop is reported.

3. Rules written out, because v1 only implied them

Merge rule. Five of Tolstoy's footnotes govern French that his own text splits in two. The rule is: a locus is all the French that one footnote's gloss renders. The narrator's tag between the halves («— сказала она, —») is not part of the locus and is not the speaker's Russian. The five merged loci now carry their full Russian context verbatim in loci-frozen.md, so the call is auditable without the copy-text.

Switch rule. WHOLE — the whole of the speaker's utterance, or the whole quoted document, is French. INTO — the utterance contains the speaker's own Russian as well as French, and the French forms one or more complete clauses. TAG — the French is a phrase serving as a constituent inside a Russian clause. The narrator's tag never counts as the speaker's Russian, which is why LII.07 is WHOLE and LI.16 is INTO.

Length. Every locus's French word count is already in loci-frozen.md and is used unchanged.

4. Coding scheme

Per cell — the retention code (this is the dimension the blind second coder also codes):

Per cell — the fate of Tolstoy's own footnote (new; critic finding (f)):

Per cell — verbal: does the hand add a signal the Russian has not ("in French", "she said in French")? cyr (the five mixed loci only): does the hand distinguish the Russian standing inside the French from the French around it, by any of the marks listed in P5?

Typography, per hand, outside the registered set except for MAU: ITAL / ROMAN / UNRECOVERED at each retained run, with the instrument named on every figure.

Boundary rules, frozen now (C1 A17): a hand who keeps one French word of a nine-word locus is KEPT-PART, not KEPT; a hand who prints the French only in a note and English in the text is ENG+NOTE, never KEPT; NA requires the sentence to be absent, not the French; a locus englished in a way that drops its propositional content entirely is DELETED, not ENGLISHED.

5. Registered predictions — v2, all per hand

• P1t — the signal is supplied, on the one hand whose type is not inferred (§7.45.1 as repaired by §7.50.2). For MAU alone: among retained French runs, the share distinguished from the surrounding type is ≥ 0.80. The three ABBYY hands are described, not tested.

• P2 — a hand who keeps the French inherits the author's own apparatus. For each hand, among loci coded KEPT or KEPT-PART, the share coded BOTH or NOTE is ≥ 0.80. This is the arm's primary. It can fail loudly: a hand may keep the French and gloss none of it, leaving the English reader with less than Tolstoy gave the Russian one. Registered because the project has no idea which way it falls, and because §7.50.4 found glossing — not marking — to be the decision that tracks what an embedded passage is doing.

• P3 — nobody deletes (§7.50.1, per hand). No hand codes DELETED at more than one of the 38 loci.

• P4 — the switch type governs, and length does not explain it. For each hand, ENGLISHED+ENG+NOTE share at WHOLE loci ≥ that share at TAG loci + 0.15, in at least 3 of 4 hands; and the same direction holds within the length-matched band (loci of ≤ 4 French words, of which the frozen list holds WHOLE: LI.04, LI.25, LI.27, LII.04; TAG: LI.11, LI.12, LI.15, LI.18, LII.02, LII.03). Both arms hold ≥ 4 loci, so the matched test is evaluable; if a hand's NA cells reduce either arm below 4, that hand is dropped from the matched test and the drop is reported.

• P5 — the second-order switch is lost. Denominator: loci coded KEPT or KEPT-PART among the five mixed loci. A hand is evaluable only with ≥ 2 such loci. For each evaluable hand, the count of loci coded cyr = YES is 0. The marks that would count as YES, named in advance: roman inside italic (or italic inside roman); quotation marks around the Russian words alone; transliteration or Cyrillic left standing where the rest is English; a note saying those words are Russian; any verbal signal at that point. The lead's own frozen limb takes the first of these at all five loci, so the move is demonstrably available in English typography. If fewer than 2 hands are evaluable, P5 is reported exploratory.

Declared prior, not a prediction, and scored against nothing (critic finding (b)): the lead believes, from training and not from this session, that MAU englishes most of the French and that GAR retains it with footnotes. The observed englishing rates are reported per hand as description.

Dropped from v1 and not reported as tests: v1 P1 (pooled typographic share — instrument asymmetry), v1 P3 (non-blind), v1 P6 (Bell's provenance — not answerable from four dependent books).

6. Failure criteria

1. Alignment. A hand aligning at fewer than 32 of 38 loci is reported for what it has and is dropped from P4's matched test.
2. Second-coder disagreement. If the blind second coder disagrees with the lead on more than 15% of the cells it codes, every retention-based prediction is reported as unresolved, not as passed or failed, and the disagreements are printed.
3. Prediction flip. Any registered prediction whose outcome changes when the second coder's calls replace the lead's is reported unresolved, whichever way the lead's own calls fell.
4. P1t void if the Gutenberg text's italic markup is absent from the sampled chapters.
5. Hand identity. DOL unidentifiable → volume dropped.

7. Cost

Critic pass, spent: $0.0896509 (C1 $0.0672525, C2 $0.0223984). Blind second coder: 152 cells in batches, one seat, max_tokens 1200 per call — worst case built from the cap, 20 calls × (≈2,000 in + 1,200 out) at google/gemini-3.6-flash $0.75/$3.75 per M read from the API this session (note (bsw)): $0.12. A second independent seat at qwen/qwen3.7-max $1.48/$4.42: $0.17. Declared ceiling for the session: $0.50. Everything else — loci, collation, translation, all lead coding, all analysis — costs $0.