Repository path: workshop/experiments/E-20260725-slatef-verification/verification.md · rendered 2026-09-09

Second-reader verification of the three Slate F source readings

Why this exists. The D-20260725-05 ratification vote refused to let a changelog caveat carry the single-reader problem and made application of the typology changes conditional on one independent, original-language verification — of Yan Fu's 譯例言 item 3 and its context (for Q1), and the cited Schleiermacher and Futabatei passages (for Q2) — with the verifier "not told the desired typology outcome", and with instructions to reopen rather than silently apply if the check materially disputed either reading.

This is not an experiment testing a hypothesis and was not run under the full experiment discipline (frozen design → independent critic → run → verification). It is a mandated check. What was fixed before dispatch: the verifier, the materials, and the exact question set below.

Method

Verifier: google/gemini-3.6-flash (panel P2). Chosen because it is neither the reviewer nor the voter on D-20260725-05 (P5 and P1 respectively) — in particular not the model that ordered the check. One verifier across all three items, so the three readings are comparable.

Materials given, per call: the primary text in its original language (classical Chinese / German / Japanese) and the lead's English translation with the translator's log stripped, so the lead's own reasoning about term choices could not steer the answer. Nothing else. No sense id, no proposal, no typology, no mention of the project's questions.

Task A — check the translation against the original: wrong or unwarranted English; any key term rendered so as to foreclose a reading the original leaves open (name the term, the rendering, the foreclosed alternative); additions or omissions of substance. Explicitly instructed not to manufacture faults.

Task B — three neutral questions per source about what the original says, each requiring a quotation. The questions were written to be answerable in either direction; the wording is the point, so it is reproduced verbatim.

Yan Fu. B1. The author defends his choice of a particular kind of written language for his translation. On what grounds does he defend it — aesthetic grounds (that this language is more beautiful, refined, or dignified), instrumental grounds (that this language does some job better), both, or neither? Quote the sentence or sentences that decide it, and say whether the text lets you rule either ground out. B2. The author names three things a translation must achieve. State each in your own words. Are they presented as three independent goods, or are some subordinate to or instrumental for others? Quote what decides it. B3. Is the third of the three named as a property of the translation's language, a property of the reader's experience, a property of the source, or something else?

Schleiermacher. B1. The author distinguishes texts that require translation proper from texts that require only interpreting/transacting. What property decides which side a text falls on — and where does that property reside: in the source text or author, in the target language, in the reader, in the translator, or elsewhere? Quote what decides it. B2. The author states conditions on which one of his two methods depends. State each condition in your own words and say, for each, whose capability or disposition it is a condition on. B3. Does the text treat the receiving language's capacity as fixed, or as something that can be changed — and if the latter, by whom? Quote.

Futabatei. B1. The author states what he considers the fundamental and necessary condition of translation. State it in your own words. Where does the thing named reside — in the source text or its author, in the target language, in the reader, in the translator, or elsewhere? Quote. B2. The author describes a method he considers better than his own and then does not adopt it. What reason does he give? Is the reason a claim about the method, about himself, or about the languages? Quote. B3. Does the author treat style as something to be matched directly, or as something that follows from something else? Quote.

B1 on Yan Fu names the aesthetic option first and invites ruling either way. B1 on Schleiermacher and Futabatei lists five possible locations, of which "source text or author" is one of five. Neither question tells the verifier which answer is wanted.

Results

All three readings were supported. Nothing was disputed. Two translations returned SOUND; one returned SOUND-WITH-QUALIFICATIONS on a sentence no reading turns on.

Yan Fu 〈譯例言〉 — bears on Q1 (雅)

This is TH-20260725-capability-conditions K3, reached by a reader that did not know K3 existed. The instrumental reading of item 3 is not an artefact of the lead's gloss.

Defect found (Task A), the only one across three sources: 「頗貽艱深文陋之譏」 → "drawn no little mockery for papering over poverty with difficulty". 文陋 means crude or unpolished prose style; the sense is mocked for being abstruse and unpolished. The verifier's statement of the damage: a reader of the English "might incorrectly assume Yan Fu was accused of using obscure language specifically to conceal a lack of substance, rather than simply being criticized for an unpolished and overly difficult writing style." Recorded on T-yanfu-yili-yan-R04-v1 and S-yanfu-yili-yan; the frozen artifact is not silently edited.

Schleiermacher, Ueber die verschiedenen Methoden des Uebersezens — bears on Q2 (voice)

TRANSLATION-ASSESSMENT: SOUND. Biegsamkeit → "pliancy" was specifically approved as preserving the metaphor "without prematurely tying it to modern linguistic terms like 'plasticity' or 'malleability'" — i.e. the term the reading leans on was checked for exactly the foreclosure the question asked about.

Futabatei Shimei 「余が翻訳の標準」 — bears on Q2 (voice)

TRANSLATION-ASSESSMENT: SOUND, no material errors. 詩想 → "poetic conception" approved.

What this does and does not establish

• Does: the two readings the typology change rests on are not artefacts of the lead's rendering. An independent reader with the originals in front of it, told nothing about what was wanted, located the deciding property in the source for both European and Japanese sources, and read Yan Fu's defence as instrumental. The condition the ratification vote imposed is discharged.
• Does: lead translations, like published ones, carry errors. Six lead translations existed before this session; the first one independently checked returned a defect. NEXT.md note (s) said "published is not checked" — this extends it: the project's own translations were not checked either, and one of the three checked here was not clean.
• Does not: establish anything about translation quality, any goodness sense, or the jury. This is factual and interpretive adjudication, the direction config/models.md already records the panel as usable in.
• Does not: constitute independent verification of Berman, which was read by S020 and is not covered here. Q2's weak fourth corroborant remains single-reader.
• Single verifier. One model, one call per source. Agreement of one model with the lead is weak evidence in the sense of charter §4; what makes it worth something here is that the questions were non-leading and the verifier was blind to the proposals, so the failure mode being guarded against — the lead finding what it went looking for — is the one this design addresses.

Cost, and a substantial recorded waste

Usable spend: $0.115349. Three calls: verify__futabatei $0.043929, verify2__yanfu $0.024828, verify2__schleiermacher $0.022737 (the two verify2__ runs supersede earlier attempts and are the cited outputs).

Wasted spend: $0.190665, four calls, none of it recoverable. Recorded in full because the project's ledger discipline is worth more than the appearance of a clean run.

The fix that worked, and it is cheap. Run 3 added one instruction — "Be brief. Answer Task B FIRST, then Task A. Quote only the decisive words; do not restate or summarise the text; do not walk through the translation line by line" — and reasoning fell from ~5,800 tokens to ~1,800, cost per call from ~$0.051 to ~$0.024, and the answers got sharper, not thinner. Putting the decisive task first is a second, independent protection: if a reasoning model truncates, it truncates the part you can afford to lose.