Repository path: workshop/experiments/E-20260810w-recurrence-census/design.md · rendered 2026-09-09

E-20260810w-recurrence-census — when a source repeats one designation, do published translators repeat one expression?

Frozen before dispatch. ARM-recurrence step 1, study limb.

1. Question

Gogol's «Шинель» (1842) never names the general who refuses Akaky his cloak. It calls him «значительное лицо» — a significant person — and does so 35 times across the story, 21 of them as the bare bigram, and it plays on the root openly: «одно значительное лицо недавно сделался значительным лицом, а до того времени он был незначительным лицом».

Do published translators carry that as one repeated expression, or do they vary it? And is the answer a property of the target language or of the translator?

Subject-rule sentence (continue-prompt.md §4.5): this unit teaches what published literary translators actually do with a source's repeated designation — the decision the project's consistency sense is about and has never observed a published hand making.

2. Why the question is owed

• S-berman-tendances §9. Berman's ninth deforming tendency is la destruction des réseaux signifiants sous-jacents. The project's only measurement of it (RS-20260806b) is on one Verga passage and reports itself as a lower bound with a named defect — the detector was vocabulary-bound and missed a network carried by do up.
• wiki/goodness-senses.md, consistency. S152 established that the jury barely sees terminology drift with the English alone (+0.417) and prices it heavily with the source on the page (+2.458). The sense's entry therefore turns on a decision no published translator in this project's evidence base has ever been observed making. The shelf holds no anchor for it.
• T4's growth condition (wiki/tracks.md §Deliverables): the shelf grows only where a deliverable needs it. Both of the above name the same missing anchor.

3. Materials, all frozen before dispatch

Two contiguous Russian regions, materials/ru_C.txt and materials/ru_AB.txt:

23 network sites. Region AB is also the source unit of T-shinel-znachitelnoe-R04-v1, whose translator's log was frozen and committed at f77b3ae before this design was written.

Four published hands, three of them independent of one another by measurement (§7), two target languages, each stored whole and each read only in the corresponding region:

A fifth published hand — Huugo Jalkanen's Finnish «Viitta» (Gutenberg 51747) — is reachable and is not censused here; recorded so the omission is a scope decision and not an availability claim.

4. Procedure

The alignment is not made by the lead. For each ⟨hand, region⟩ the seat receives the Russian region with every occurrence of the root numbered in place (⟦1⟧…⟦n⟧) and the hand's corresponding target region, unlabelled and unattributed, and returns for each numbered site the exact substring of the target that stands for it, or NONE.

• 3 seats, from config/models.md panel v1, non-Anthropic: P1, P3, P5.
• 16 items per seat = 4 hands × 2 regions × {network task, control task}; 48 calls in all.
• Blind to the hypothesis: the prompt names no theory, no hand, no date, and does not say that repetition is the object of interest. The control task is worded identically.
• Control task. The same alignment for a second recurring item in the same region — строгость in C (5 sites), шинель in AB (3 sites) — networks no hand can be supposed to have dropped.

5. Quantities

Per hand, over the 23 network sites, from the majority answer of the surviving seats (a site's expression is the string ≥ 2 seats quote, after case-folding and article-stripping):

• fix — share of sites rendered by that hand's single most frequent expression.
• var — number of distinct expressions.
• del — share of sites answered NONE by the majority.

6. Registered predictions and failure criteria

• P1. Carriage is a translator decision, not a language constraint: among the three English hands, max(fix) − min(fix) ≥ 0.30.
• P2. The play costs more than the label: at the six sites of region C where the root is not the character's designation (the pun's tail — «незначительным», «значительнейшими», «значительность», «значительно»), fix is lower than at the designation sites, for every hand.
• P3. The German hand is not the explanation: H4's fix lies inside the English range rather than outside it. Fails if German is at either extreme alone.

Failure criteria, registered:

• F1. A seat recovering < 0.80 of the control sites in a region has its cells for that region voided. If fewer than 2 seats survive for a ⟨hand, region⟩, no quantity is read for that hand and the cell is reported empty.
• F2. If three-seat agreement on the quoted expression at network sites is < 0.70, the alignment is not reproducible: fix / var / del are reported as descriptive counts with the disagreement rate attached, and P1–P3 are withheld.
• F3. If any seat's answers contain an expression not present as a substring of the target passage it was given, that seat's cell is voided as fabricated.

7. The contamination gate, run before this design existed

tools/dependence_check.py, immediately after the freeze commit of the lead's translation and before any of §4 was written:

Two consequences, both binding on this design.

1. The lead is not a hand in this census. Its rendering is the arm's translation limb and process evidence only. contamination: high was declared on the artifact before the gate ran.
2. H1 and H2 are not independent of each other in region B — 18 contiguous tokens, 4 shared 15-grams, above anything the lead shares with either. Any statement of the form two English hands independently did X must be read as one and a half hands in that region, and the result page says so wherever it matters.

8. Amendments after the pre-run critic, all made before dispatch

qwen/qwen3.7-max (reserve slug, outside the jury), NEEDS-REDESIGN, 5 findings, 2 BLOCKING — critic.md, $0.010654. Four accepted, one accepted as a written limitation, none overruled.

• A1 (from BLOCKING 1, accepted in full). §5's fix was defined over all 23 sites and P2 needs it split by site type: "you cannot compute a conditional metric from a global aggregate." The 23 sites are classified REF (the root designating the character) or ECHO (the pun's tail), frozen in sites.json before dispatch: 16 REF, 7 ECHO. All of region AB is REF; region C splits 9/7. Every quantity in §5 is now computed three times: over REF, over ECHO, and over all.
• A2 (from BLOCKING 2, accepted, the critic's own first remedy). P1 compares English hands, and §7 measures H1~H2 as dependent in region B. P1 is therefore computed on region C alone — 9 REF sites, where H1~H2 measures clean (3 / 0 / 0, longest run 8). Region AB's figures are reported descriptively and carry no prediction.
• A3 (from SERIOUS 3, accepted). F1's control gate was per region, and region C has 5 control sites, so one miss voided a cell. F1 now pools the 8 control sites of both regions and voids a seat at < 0.75 (fewer than 6 of 8). The bar is set here, before any control answer exists, and will not move afterwards.
• A4 (from SERIOUS 4, accepted as a limitation, not repaired). The blinding is structural, not semantic: sixteen marks on one root in a short passage tell a competent reader that repetition is the object. It cannot be repaired without changing the task. It is a common-mode influence — every hand is marked identically — so it cannot manufacture a difference between hands, which is what P1 and P3 are about. It could inflate every hand's fix together, and no absolute level is claimed on that account.
• A5 (from ADVISORY 5, accepted). var is confounded by the deletion rate. var is reported with its valid-site denominator and as ttr = var / valid.

9. What this cannot establish

• No jury, no scoring, no quality claim. This measures what is on the page. Whether a destroyed network costs a reader anything is not asked here and is not answered by it.
• Four hands, two languages, one story. Nothing here generalises to translation at large.
• Tier D is NOT PASSED; the page carries provisional: true. No recommendation may rest on it.