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

«Köyhää kansaa» — the contamination selection gate

Run before the work was selected and before a word was translated, as CLAUDE.md's standing rule requires. Outcome: contamination: none, and the measurement was NOT RUN, because its comparator does not exist.

What the gate measures, and why it could not run

tools/dependence_check.py measures the longest common contiguous run and the shared 7-/12-/15-gram counts between the lead's English and a published English rendering of the same source. It needs a published English rendering. No English rendering of «Köyhää kansaa» — or of any Minna Canth prose — was located.

Three routes, each a real tool result and not an assertion:

Why the declaration is none and not suspected

This is a different position from S084's, and the difference is the point. RS-20260801g declared «Gudsfreden» suspected UNMEASURED, and said plainly that the work had been chosen partly because its gate could not be run — with two measured siblings that had come back at 17 and 16 tokens against an 1899 English comparator. There, the comparator existed for the neighbours and the risk was demonstrated.

Here there is no comparator for this work, for its siblings, or for the author. The gate measures overlap with a published English rendering; where no such rendering exists in English at all, there is no English text of this work for the lead to have absorbed. That is a claim about the world rather than about obscurity, and it is the one form in which "unmeasured" can honestly support none.

The project's standing rule is not being contradicted. CLAUDE.md holds that obscurity is not protective and that the boundary is measured overlap — established three times, most sharply at S079 (24 tokens against an obscure 1923 anthology). Every one of those measurements was against a work that had an English translation the lead could have seen. The rule says an obscure translation is still a translation; it does not say a translation that does not exist can contaminate.

What remains unmeasured, stated so it is not lost

1. The Swedish channel is open and is not measured. Rafaël Hertzberg's authorized 1886 Swedish «Bland fattigt folk» (Gutenberg 20518) is freely available and plausibly reachable in training data. It cannot put English strings into the lead's head, but it could shape construals — which reading of an ambiguous clause looks obvious. It was not opened before span 1 and its log were frozen (R05, charter A4); it is opened afterwards, deliberately, as this session's study limb, and every use of it is after the freeze.
2. No Finnish→English dependence tooling has ever been exercised in this project. dependence_check.py handles Latin-script tokenisation and would work; it has simply had nothing to run on here.
3. R05's own known limitation applies: the gate is a selection gate, run once. A later span could sit on more heavily reproduced material without anything re-firing. For this work that risk is unusually low — there is no English «Köyhää kansaa» for any span to collide with — but the statement is recorded rather than assumed.
4. If an English rendering is ever located, the gate must be run retrospectively on the frozen spans, which is possible precisely because every span is frozen in git before any comparator is read.