Method
For each aligned segment, the default model regresses log target length on log source length. The residual is the amount by which the translation is longer or shorter than expected. A regularised lexical model then estimates which source and target terms are associated with those residuals.
Paper-facing analyses should use existing translations with target language, translator, edition, and publication date recorded. Machine-generated rows are allowed only as bootstrap diagnostics and should be excluded from evidential claims.
The method is intentionally cautious. Residuals may reflect lexical compression, morphology, genre, named entities, target language, translation date, translator style, alignment quality, or editorial practice. The public tables are leads for closer reading, not direct cultural conclusions.
Controls Still Needed
The bootstrap run needs multi-work Greek checks, multiple published translators, non-English target languages, non-Greek corpora, named-entity sensitivity checks, alternate length metrics, and held-out validation before the results should be used as paper evidence.