Repository path: workshop/experiments/E-20260805b-kinship-line/design.md · rendered 2026-09-09

E-20260805b — does an English reader recover the line Finnish draws with sinä and te?

Frozen 2026-08-05, S110, AFTER the translation limb it tests was frozen (T-koyhaa-kansaa-R05-v1 span 6, commit 9c41878) and before any seat was dispatched. The materials are that frozen English and nothing else; no comparator in any language has been opened.

1. Where the question comes from

The translating of span 6 produced the second-person grid of the whole novella (D103), and the grid has a within-speaker control in it. Tiina Katri, in one household on one day, says sinä to Mari and te to Holpainen, and both pairs return the form she uses. Across the work the single-addressee te pairs are Mari ↔ the landlord, Mari ↔ a servant girl, Mari ↔ a beggar-woman, Tiina Katri ↔ Holpainen, the doctor and pastor downward to the poor and between themselves, and a stranger to Tiina Katri; the single-addressee sinä pairs are husband ↔ wife, mother ↔ children, Mari ↔ Tiina Katri, and Mari to God. So the line is kinship, not intimacy — register.md V25, which corrects V23's gloss without changing a rendering.

English has one "you" and carries none of it. Five spans of this register have written the words declared loss about this system (V13, V14a, V18a, V23, V25) without ever measuring one. The question this run asks is the only one that decides whether the phrase is doing any work:

Given the English alone, does a reader separate the two pairs that the Finnish separates — Tiina Katri to Mari, and Tiina Katri to Holpainen?

What was checked first — and the check was mis-scoped, which the pre-run critic caught. The obvious free compensation would be the personal-name vocative: if Canth supplied a name exactly where she supplied sinä, English would get the map without the translator doing anything. Over the whole novella she does not — ¶383 has Holpainen address Tiina Katri by name and use te in one sentence («Kiitoksia, hyvä Tiina Katri, ja tulkaa pian…»), and ¶425 has Tiina Katri use sinä to Mari with no vocative at all. But inside this extract the alignment is close on the side that matters: Tiina Katri names Mari three times (¶409, ¶412, ¶423) and never names Holpainen. That was a claim about the book stated as though it were a claim about the scene, and it is the critic's F4. The vocative is therefore a live candidate carrier here, and arm 3 exists to test it (critic.md A3).

2. Materials

Arm 1 — the frozen English, unaltered. extract-arm1.txt: paragraphs ¶376–386 and ¶407–429 of the frozen span-6 English, 34 paragraphs, 1,147 words, with an explicit […] marking the cut. The extract contains, in direct address:

Mari and Holpainen are fixed as one household by the extract's own narration, not by the pronoun: ¶386 reads "knew that that was her man and that those children were her children." That is what makes them usable as the instrument check in §4.

Arm 2 — one carrier inserted, the positive control. extract-arm2.txt is byte-identical to arm 1 except that Tiina Katri's two speeches to Holpainen carry an English formality marker: "Go in now, Mr Holpainen, …" and "Come away, Mr Holpainen, if you please," she whispered. Seven words changed in 1,147. This arm is not a candidate rendering and is not proposed as one — it is a fabrication, forbidden by the register's own V18a reasoning, and exists only to establish that the seats can read an address distinction out of English when English supplies one.

Arm 3 — the vocative stripped, the subtle control (amendment A3, critic F4/F3). extract-arm3.txt is arm 1 with Tiina Katri's three name-vocatives to Mari removed and nothing else touched — four words out of 1,147. Holpainen's vocative to Tiina Katri at ¶383 stays, and so do Mari's to Holpainen, so only the asymmetry on Tiina Katri's side is gone. This is the comparable-subtlety control the critic's F3 asked for, and unlike arm 2 the string it removes is one the source actually supplies.

3. Procedure

Three seats, one call each, per arm — nine dispatches: P1 openai/gpt-5.6-terra, P3 x-ai/grok-4.5, P5 deepseek/deepseek-v4-pro (config/models.md). P2 and P4 are excluded on notes (bhf), (bhq) and (bit) — both have returned zero-content bodies on short-answer shapes. max_tokens 12,000 on every dispatch, sized from measured reasoning appetite and not from the answer, per (bhq); a seat probe at that ceiling runs first, per (bit), and its per-seat minimum is recorded.

Each seat is shown one extract and nothing else — no title, no author, no language of origin, no statement that a translation is involved, and no indication that the two Tiina Katri pairs are what is being compared. Each is asked, for each of the three ordered pairs, to pick one of

• A — of one household or family
• B — not family, but as close as family
• C — neighbours who help one another
• D — acquaintances, no more

and to quote, verbatim and in ≤ 25 words, the words in the passage that decided it; then to rank the three pairs from the one the passage makes closest to the one it makes least close. The pair order in the prompt is fixed and identical across arms and seats.

Judgment is not parallelised: the six dispatches are independent single calls, and no seat sees another's answer.

4. Registered predictions and failure criteria — frozen before dispatch

H1 — the loss is real. In arm 1, ≥ 2 of 3 seats give the SAME letter to Tiina Katri–Mari and to Tiina Katri–Holpainen.

H2 — the ranking agrees with H1. In arm 1, ≤ 1 of 3 seats ranks Tiina Katri–Mari strictly above Tiina Katri–Holpainen.

H3 — the vocative test (registered under amendment A3, before any seat ran). The number of seats separating Tiina Katri–Mari from Tiina Katri–Holpainen is lower in arm 3 than in arm 1. If arm 1 separates and arm 3 does not, the carrier is Canth's own vocative, supplied for free.

FC1 — the reading check (renamed under A5; the critic's F6 is right that it is not an address-sensitivity check). If fewer than 2 of 3 seats in arm 1 place Mari–Holpainen at A, the seats are not reading the passage, and nothing from this run is reported as a result. The pair is named as one household in the extract's own narration. It certifies reading and nothing more; sensitivity to address is certified by FC2 at the loud end and tested by H3 at the subtle end.

FC2 — the positive control, and the licence for H1. If fewer than 2 of 3 seats in arm 2 separate Tiina Katri–Mari from Tiina Katri–Holpainen (different letters, or Mari ranked strictly above Holpainen), then the seats cannot read an address distinction out of English even when English supplies one, and H1 is withheld whatever arm 1 shows — a null on an instrument with no demonstrated ceiling is not a measurement. This is the criterion note (bis) was written about.

FC3 — blind. If a seat names the author or the work, that body is dropped and re-dispatched once to qwen/qwen3.7-max; if two or more seats do so in one arm, contamination is declared and that arm is withheld. Naming the language or region is expressly NOT contamination and is not counted — the characters are called Tiina Katri, Holpainen and Mari, and inferring a Nordic setting from the names is reading the passage, not recognising the book. Contamination on this work is declared none and argued in ../../translations/koyhaa-kansaa/contamination.md: no English rendering of any Canth prose was found on three documented search routes.

FC4 — quote integrity. Every quoted string must appear verbatim in the extract that seat was shown, checked mechanically by verify.py. A seat with any non-verbatim quote is dropped from the counts and the drop is reported.

FC6 — the floor check (amendment A2, critic F1/F2). If fewer than 2 of 3 seats in arm 1 place Tiina Katri–Mari at A or B, then a same-letter result is a floor effect — the seats did not read that pair as close at all — and H1 is withheld. B and C are not mutually exclusive for a neighbour who washes the corpse, and a same-letter answer must be shown to be a conflation and not a default.

FC5 — empty or truncated bodies. Any body with finish_reason: length or zero content characters is re-dispatched once at max_tokens 20,000; a second failure drops the seat and the cost is ledgered. Notes (b), (bhq), (bit).

5. What each outcome licenses, and what none of them licenses

• H1 and H2 both hold, FC1, FC2 and FC6 all pass → an English reader of this translation does not recover the kinship line the Finnish draws, and the register's five "declared loss" entries become one measured loss.
• H1 holds and H2 fails → partial recovery, reported as such and never as a confirmed loss (amendment A4, critic F5): the seats perceive a difference below the granularity of four letters.
• H1 fails → the English recovers the map by other means, the seats' quotes name the means, and the declared loss is smaller than five spans of this register have assumed.

• FC1 or FC2 fires → nothing is reported and the run is a cost.

• H3 holds → the carrier is the vocative, and the distinction reaches an English reader without the translator adding anything, because the source supplied it.

The narrowing the critic's F3 forced, and it binds every line above (amendment A6). No outcome of this run may be reported as showing what English can or cannot carry. Arm 2's marker is a sledgehammer next to an unmarked Finnish pronoun; its passing shows the seats read titles, not that they read register. What is reportable is what this translation delivers to a reader of this extract, and which carrier decides it.

The limit that no outcome removes, stated before the run. This design cannot separate English deleted the carrier from the content of this extract does not differentiate the pairs either. A Finnish reader given the same twenty-three paragraphs with the pronouns masked would face the same extract. What is measured is whether the distinction survives into a reader of the English, which is the question a translation is answerable for; what is not measured is whether the distinction is recoverable from the scene in principle.

Every sentence of the result is provisional and internal-judgment-only. Tier D is NOT PASSED (config/models.md); the seats here are readers reporting what a passage presents, not jurors rating quality, and no translation is judged in this run.

6. Pre-flight cost

Seat probe (3 calls, ~30 tokens out) + pre-run critic (1 call, z-ai/glm-5.2, max_tokens 24,000) + 9 seat calls at max_tokens 12,000, with FC5 allowing 2 re-dispatches at 20,000.

Worst case, built from the caps and not from expected output (note (abc)), with a 4× routing margin (note in config/models.md §Pricing caution): probe $0.01 · critic $0.36 · seats 9 × 12,000 out at the panel's worst plausible billed rate $0.82 · re-dispatches $0.18. Declared worst case, revised after amendment A3 added a third arm: $1.40. Today's headroom at design time: $4.070065689.