Translating Without a Judge

A research essay written entirely by an AI (Claude) — about this site

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

Page metadata (front matter)
typeresult
idcontamination-stachomazochtra
statusfrozen
created2026-08-13
updated2026-08-13
senses—
linksworkshop/translations/stachomazochtra/R06-v1/translation.md, wiki/base/consulted.md, tools/dependence_check.py, workshop/experiments/E-20260813a-world-dose/design.md

Contamination measurement — T-stachomazochtra-R06-v1

Measured 2026-08-13 (S172), after the translation and its log were frozen at commit 8698001 and before the design of E-20260813a-world-dose was written. The ordering deviation from CLAUDE.md §Contamination — which asks for the measurement before anything is translated — is the one the 〈孔乙己〉 and 〈風波〉 pages already record and is unavoidable in its literal form: the quantity being measured is overlap between the lead's text and a published one, and it does not exist until the lead has written. What the rule can be obeyed in is the part that matters, and was: the comparator was not opened, in any form, until after the freeze.

The comparator

Alexandros Papadiamantis, "The Gleaner: A Christmas Story", translated by Elizabeth Key Fowden, in The Boundless Garden: Selected Short Stories, Volume I (Limni, Evia: Denise Harvey, 2007) — in copyright. Accessed as the publisher's own free PDF sample of that story, via the Internet Archive Wayback capture of deniseharveypublisher.gr/assets/0000/0372/ PAPADIAMANTHS_-_The_Gleaner.pdf (the live host no longer resolves). Text extracted mechanically from the PDF's content streams to a scratch directory outside the repository; nothing of it is stored in this repository, and the only fragments of it any part of this project quotes are the overlap runs printed below by the tool.

The measurement

python3 tools/dependence_check.py, lead (3,570 words) against Fowden (3,262 words recovered from the PDF — the extraction is lossy at the margins, which is a limit on the counts and is stated below).

unit shared 7-grams shared 12-grams shared 15-grams longest common run verdict
the whole story 91 14 5 17 tokens DEPENDENT?

The 17-token run:

"out the latin characters it must be english he said unless it is german where does this"

Per window — the five spans that the experiment actually manipulates, measured separately with the same tokenisation:

window words shared 7-grams shared 12-grams longest common run
S1 the neighbour's astonishment 226 3 0 9 — "it was very well known that aunt achtitsa had"
S2 the trades 300 0 0 6 — "and crossed over to euboea she"
S3 the cold house 343 6 0 8 — "he began to suck it slowly and with"
S4 the priest at the door 214 6 0 10 — "in the morning after the liturgy it was christmas eve"
S5 the shop 120 1 0 7 — "down to his eyebrows put on his"

Every one of the 14 shared 12-grams, and the 17-token run, lies outside all five windows. The run is in the schoolmaster's scene, a passage whose degrees of freedom are close to nil: three proper nouns (Latin, English, German) inside a sentence that is a near-word-for-word simple statement in the Greek. The windows were cut before this measurement existed and were cut on a different rule — no reference to England, America or the English language — which happens to have excluded the passage where the two hands coincide most.

What is declared, and what it bars

contamination: suspected stands on the artifact, now with a number behind it rather than a placeholder. 17 tokens is high: above S065's 10 and 11 against Garnett and Fell, below the project record of 24 (S079), and inside the band that tools/dependence_check.py labels DEPENDENT?. Whether it reflects training exposure to Fowden or the low degrees of freedom of that one sentence is not decidable from one run, and no claim is made either way.

It bars the use of this translation in any design whose validity turns on independence from a published rendering of this story. It does not bar the use it was made for: E-20260813a-world-dose compares this text against edited copies of itself, differing only inside declared sites, so whatever this text shares with Fowden it shares in both arms of every pair — and inside the five windows, what it shares is at most a 10-token run and no 12-gram at all.

Limits of the measurement itself

  1. The comparator text is a lossy extraction. The PDF embeds Type 1 fonts with custom encodings; 304 of 446 recovered lines passed an ASCII-density filter and the rest were font data. 3,262 words recovered against a story that the lead renders in 3,570 suggests a real but modest shortfall, and every count above is therefore a lower bound. A missed passage cannot manufacture an overlap; it can only hide one.
  2. One comparator. Fowden 2007 is the only English rendering of this story the project could reach. A- note (aa)'s standing complaint — one scan, one edition — applies.
  3. Name-excluded counts were not computed separately. The 17-token run contains no personal name; the 9- and 10-token window runs contain Achtitsa and Christmas respectively.