English for the passages with none

How they were made and checked

Claude (Anthropic) translated each passage from the original text as the library holds it, one English line for each line of the original, without consulting a published translation, and flagged doubtful readings in notes. A second AI, Codex (OpenAI), then checked every passage against the original and listed what it took for errors. Each objection was weighed against the original and corrected where it held up, about four in five; the rest, where the translation follows a standard reading, were kept, some with a note.

Before any of this, a blind test set 187 AI translations beside the published translations of the same passages. The independent reviewer judged the AI translations of Greek and Latin prose at least as accurate as the published ones, with no major error in 75 prose passages. Verse, inscriptions and the less-studied languages had occasional errors, and Akkadian and Coptic were weaker than their published translations, so they are not translated here. Across the 718 Greek and Latin passages in these pages, the reviewer found a major error in about one in twenty before correction; in the Old Norse, about one in twelve, most of them in skaldic verse; in the Middle High German, about one in seventeen; in the Egyptian, about one in sixteen; in Dante, one in 69; in the Old French, one in 34; in the Hebrew and Aramaic, one in 23.

The Egyptian texts are translated from transliteration: the Pyramid Texts from the UD Egyptian-PC treebank, the rest from the Thesaurus Linguae Aegyptiae. The Thesaurus has its own German translations; they were not used, so each English version is an independent reading of the Egyptian, and where the reading is uncertain the notes say so.

The Old Norse texts are transcriptions of medieval manuscripts from the Medieval Nordic Text Archive. Each reference is a leaf or page and a line of the manuscript, and the English keeps to those lines, so a sentence breaks where the scribe's line did. A few medieval letters that Unicode has no standard place for are shown by their nearest standard letter and accent.

No specialist has reviewed any of it. Every page names its source edition and its licence; most are CC BY-SA 4.0, as their sources require, and a few are CC BY 4.0.