We built a 1,200-query benchmark to measure how much Recall loses when the question and the precedent are in different languages. German loses little. Mandarin loses some. Arabic loses more, and we think we know why.
Draft 3.8 produces a delivery-language version alongside the working draft, with locked defined terms, a parity check on every clause, and a confidence tier per language.
Recall 2.6 anchors every citation to the source-language clause, so a pinpoint reference stays valid whichever of the 38 languages you read the answer in.
A bilingual contract that differs between its two language versions is a dispute waiting for a tribunal. Here is how we test for parity, what the test catches, and what it still misses.
