Repository path: workshop/experiments/E-20260731c-futabatei-alignment/amendments.md · rendered 2026-09-09

E-20260731c — amendments from the independent pre-run critic

Seat qwen/qwen3.7-max, provider Alibaba, in 5,253 / out 9,585, stop, 174 s, $0.0501618. Probed-but-not-selected: not an alignment seat and not the alignment seats' declared reserve, so the S053 role-collision fix holds — fifteenth session running. Verdict NEEDS-AMENDMENT, six findings, two BLOCKING. Raw at runs/critic.raw, text at runs/critic.txt.

All six accepted, none declined. Amendments frozen before the translation limb was begun and before the aligner was run.

A1 — finding 1 (BLOCKING, §3): the JA type rule misclassifies the commonest dialogue shape in Japanese

The defect. The exception list contained the quotative particles と and ト. 「ありがとう」と彼は言った — a short quote, no internal punctuation, followed by a quotative — is the standard Japanese dialogue paragraph, and the frozen rule would have classified it N. The rule was written from five observed instances of the name-quote pattern and generalised in the wrong direction.

Adopted, taking both halves of the critic's repair. The exception now requires all four of:

1. t ∈ {は, が, も, の, を, に, へ, で, や} — と and ト removed;
2. s contains no character in 。！？、;
3. len(s) ≤ 8;
4. s is entirely katakana (including ー and ・) — the critic's own alternative, adopted as a conjunct rather than an alternative.

The direction of the residual error is stated. The exception is now strictly narrower, so a name-quote that is not all-katakana is classified D and a dialogue-tag paragraph is no longer classified N. Over-classifying as D is the conservative direction here because the RU side's D marker (the em dash) is unambiguous, so a spurious D on the JA side produces a visible type mismatch that hand-verification will catch, whereas a spurious N would quietly pull a dialogue paragraph out of its run.

A2 — finding 2 (BLOCKING, §7 and PL4): the causal claim is withdrawn, and the control that would have saved it does not exist

The defect, and the critic is right. PL4 said that if Futabatei's compliance exceeds the lead's unforced compliance, "the 1906 rule left a measurable trace." It does not follow. Futabatei was a native writer of 1888 Japanese; the lead is a non-native writer of modern Japanese. A higher figure for Futabatei is exactly what "native/period Japanese preserves more source punctuation" also predicts, and §7 of the design concedes the premise of that alternative in its own words.

The repair the critic prescribed was to recruit a native unforced baseline. It was looked for and it is not reachable. Checked from actual tool results this session:

So PL4 is withdrawn and replaced. The lead's Japanese limb is demoted: it is no longer the baseline, and it now bounds one thing only — what the rule costs a modern non-native translator who is trying to keep it. It is a ceiling-cost measurement, not a comparison class for Futabatei.

PL4′ (replacing PL4). Descriptive only: Futabatei's block-level J1 and J2 compliance are reported beside the lead's unforced and forced Japanese figures, with the confound named in the same sentence, and no inference from the ordering to the presence or absence of the 1906 rule is made from that comparison.

The inferential weight moves to two controls that are internal to Futabatei's own text, and are therefore native, period-matched and same-translator by construction:

• A3's length-rate null (below), which asks whether his counts match the Russian's better than a translator who punctuates Japanese at its own natural rate for a paragraph of that length;
• the discourse-type split — J1 and J2 compliance computed separately over blocks whose RU paragraph is dialogue and blocks whose RU paragraph is narration. Both are Futabatei, 1888, native. Registered prediction PB4: if the compliance figure is an artifact of Japanese rather than of a rule, it should not differ materially between the two; a large split is evidence that something is being done deliberately in one of them.

A3 — finding 3 (SERIOUS, §6): the block sum hides redistribution, and the null was too weak

Two changes, one from the finding and one it exposed.

(i) The primary statistic is now the STRICT 1:1 figure, and the block-sum figure is reported beside it as the secondary. The critic is right that a block sum conserves total count while permitting a translator to move three commas from one paragraph to the next. The arm's instruction — measure at the aligned-block level rather than on the 1:1 subset — is honoured by reporting both and by reporting their difference as the redistribution quantity, which is more than either number alone says.

(ii) A second null is added, and it is the one that matters. The permutation null shuffles JA counts across blocks, which holds the work's own distribution fixed but ignores length: a translator who merely preserves paragraph length will match counts above chance with no rule at all. The length-rate null: fit the work's own JA marks-per-character and commas-per-character rates, draw each block's JA count from a Poisson at that block's own JA character length, 10,000 draws, seed 20260731, and report the expected exact-match rate. A compliance figure that does not clear the length-rate null is not evidence of a rule.

A4 — finding 4 (SERIOUS, §5): the candidate list is dynamic

Adopted verbatim. The independent alignment check presents the aligned Japanese paragraph(s) plus the two immediately preceding and the two immediately succeeding, so a 1:1 block offers 5 candidates, a 1:2 block 6, a 1:3 block 7. The hard "five" is removed, and the prompt states the count for each item.

A5 — finding 5 (SERIOUS, §3): the tie-break did not exist

Adopted, taking the critic's second option because it needs no tie-break at all. The type penalty is now:

TYPE penalty  60   if ANY RU paragraph's type in the block differs from
                   ANY JA paragraph's type in the block; else 0

Deterministic on every block shape including even ones. The consequence is stated: a genuine split of one Russian dialogue paragraph into a Japanese speech paragraph plus a Japanese narrative-tag paragraph now pays the penalty. The penalty is finite, so the length evidence can still buy that block, and every such block is a non-1:1 block and is therefore hand-verified.

A6 — finding 6 (SERIOUS, §10): F3's unit is now defined

Adopted. L06 and L09 are 1:1 with the Russian by construction — R09 J1 and J5 forbid splitting or joining, and the unforced limb was written paragraph-for-paragraph as well — so for the translation limb block and paragraph are the same object. F3 now reads: if L09 cannot reach exact equality on J1 and J2 at ≥ 90% of its aligned blocks, PL3 is failed and the forced arm is reported as a partial ceiling. The design's §7 statistics are block-level throughout, matching §6.

What the critic did not find, recorded because a critic's silence is not evidence

It did not challenge the exposure record in §1, the choice of translated span, the inherited counting definitions, or the claim that the 1969 anthology's punctuation may not be Futabatei's — which §11.2 already names as the largest threat to condition B and which no amendment here touches.