Measured on documents,
published with the limits.

How TranslateX evaluates

We test the engine the way its users judge it: on whole documents, in both directions between Arabic and English, with the translation rated for meaning, fluency and completeness and the file checked as a file, for direction, alignment, fonts, images, headers, footers and margins. The two scores are kept apart, because a translation can be word-perfect and still arrive as a broken page.

Three rules hold. The figures come from the saved evaluation records, and anything computed on top of them is marked as added. Comparisons show the competitor's output as rendered and say what was inspected. And the limitations are on the page, not in a footnote: what a model judge can and cannot tell you, which groups are too small to compare, and the fact that Lisan Research runs its own evaluations and does not call them independently verified.

The canonical copy of the white paper, with its citation, is on lisan.com and listed on the Lisan research hub. To test the engine on your own material, TranslateX is free to download, and support@lisan.com is the address for methodology questions.