Repository path: workshop/translations/kongyiji/contamination.md · rendered 2026-09-09

〈孔乙己〉 — contamination, measured

Run after all three renderings and all three logs were committed and frozen, so no measurement could prime a translation, and before E-20260807e-sense-tradeoff was designed, which is the order CLAUDE.md §Contamination requires. Tool: tools/dependence_check.py (whitespace tokenisation, standalone digits deleted). Cells and raw output: runs/ below.

Comparators, and what could not be reached

Logged in wiki/base/consulted.md. Neither comparator was read. Identification of the YANG file was itself mechanical: its <title> and a probe count of six strings.

Length asymmetry, stated because it matters. YANG is the whole story, 1,913 words; EAS is 1,061 words of extracted English lines covering roughly the same first half as the selection; the lead's arms are 1,012 / 949 / 972 words over the six segments only. A comparator longer than the subject can only raise the shared-n-gram count, so the subject cells below are, if anything, generous. The EAS extraction retains some page furniture, which cannot share n-grams with a translation and so cannot raise anything.

The measurement

Declaration: contamination: none, and the basis

The lead's overlap with the published translator sits at the independent-hand floor measured on the same materials in the same run. 6 shared 7-grams and a longest run of 11 tokens against the reference pair's 4 and 10 — the same order, on cells whose lengths are comparable, with zero twelve-grams anywhere in either. The standing rule requires the 7-gram count and the run (CLAUDE.md; run length alone has failed as a proxy three times), and both agree here.

Scope of the declaration, stated rather than implied: none means not distinguishable from independence against the two comparators reachable on 2026-08-07. Lyell and Lovell were not measured and this says nothing about them.

The finding that is not about contamination, and is the more useful one

The lead matches itself far above where it matches anyone else. R06 ~ R07 returns 51 shared 7-grams, 9 twelve-grams, 3 fifteen-grams and a 16-token run — the tool's own verdict line reads DEPENDENT? — against 6 / 0 / 0 / 11 for the same translator's R06 against a published hand. This is note (bhb)'s finding (the lead matches itself at up to 37 contiguous tokens, above its record against any published human translation) reproduced within a single session and across two opposed programmes, which is new: (bhb)'s cases were the same regime eleven days apart.

The 16-token run is in S6 — spread his five fingers over the dish to cover it and stoop down and say there — a stretch of narration where no fluency rule bore at all, so the fluency arm had no reason to write anything else.

What this does to E-20260807e, in one sentence: the three arms are not three independent renderings, they are one rendering and two rule-driven departures from it, and shared wording can only shrink a measured difference between arms, never manufacture one. So the dependence makes every contrast in the experiment conservative, and it is carried into the design as a declared limit rather than a discovered one. The R08 arm departs about three times as far as the R07 arm (17 vs 51 shared 7-grams), which is itself a fact about how much of a translation each rule set actually touches.