Repository path: workshop/experiments/E-20260803g-address-axis/design.md · rendered 2026-09-09

E-20260803g — an adversarial check on the address-axis claim

Frozen before dispatch, 2026-08-03 (S100). This is not an experiment on a translation and not a jury run. It is a single adversarial pass over a claim about a source text, run in the one direction charter §4 and the S015 instrument note permit: models are used to find errors, never as validation. Whatever the seats refute is acted on; whatever they confirm is not cited as support anywhere, and the result page carries no number from this run.

1. What is being attacked

RS-20260803g §3 claims, of Minna Canth's «Köyhää kansaa» (1886):

C1. The te/sinä contrast in this novella is not status-asymmetric. te is used mutually across the sharpest status gaps in the translated text — landlord ↔ tenant (¶80, ¶83, ¶84, ¶97), servant ↔ pauper (¶122, ¶123, ¶126), pauper ↔ beggar-woman (8 sites, ¶135–159) — and therefore does not mark who is above whom.

C2. The only asymmetric te inside intimacy is a child to her mother (¶67, ¶103, ¶175, ¶389), and that is why those four sites, and only those, read as deference.

C3. Six te events in span 4 and one in span 1 are plural by number, not politeness: ¶29, ¶151, ¶222, ¶240, ¶279, ¶283, ¶284.

2. Materials

materials/sites.json — 29 paragraphs of the Finnish, verbatim from workshop/translations/koyhaa-kansaa/source-fi-full.txt, with paragraph indices and nothing else. No English, no translation, no classification table, no argument from the result page beyond C1–C3 as stated above.

3. Seats and procedure

Two non-Anthropic seats (config/models.md), each dispatched once, statelessly, blind to the other, both instructed to refute:

Prompt, identical for both: You are a Finnish philologist. Here are 29 paragraphs of an 1886 Finnish novella, verbatim. Three claims are made about the second-person system. Try to REFUTE each. Quote the Finnish you rely on. Default to REFUTED where you are uncertain; a claim you cannot break, say so in one line and move on. Also list any second-person form in these paragraphs that the claims fail to account for.

effort: low on the first dispatch (note (b)); max_tokens: 3000; "usage": {"include": true}.

4. Registered predictions

1. C1 survives — the mutual landlord/tenant pair is on the page in both directions and is not arguable.
2. C3 is the vulnerable one. ¶240, ¶279, ¶283, ¶284 and ¶29 have no same-scene minimal pair, and a competent Finnish reader may call one of them a polite singular. This is registered as the expected failure, and §7 of the result page already names it.
3. The seats will find at least one form the classification missed. ¶122's «näettekös» and ¶249's «miehesi» were added to the census only after a second pass, so the census is known to be the kind of object that grows.

5. Failure criteria

• FC1. If a seat produces no Finnish quotation in support of a refutation, that refutation is discarded — an assertion about a text without the text is not a finding.
• FC2. If either seat refutes C1 with a quotation showing asymmetric te at a landlord ↔ tenant or stranger ↔ stranger site, RS-20260803g §3 is withdrawn before the session lands, not amended.
• FC3. If a seat refutes C3 at a site, that site moves to the result page's limits section with the seat's reading named. It does not silently disappear.
• FC4. A seat that returns finish_reason: length is re-dispatched once at max_tokens: 4000; a second failure withdraws that seat and the run reports with one.

6. Pre-flight cost estimate (note (abc): built from max_tokens, not from an expected length)

Declared worst case including one retry each: $0.42. Today's headroom at dispatch: $1.737138691 of the $5.00 UTC cap, six prior sessions. Opening key snapshot runs/snap/session-open-S100.json, usage 54.776232642.