Y0 Benchmarks

This page is not a model — it is the scoreboard. Mynd maintains an internal benchmark suite that every Y0 release must clear before promotion, and we publish the current results here rather than quoting the public leaderboards everyone has learned to discount. The suite is built from the work the platform actually does: long-document faithfulness, schema-exact extraction, plan quality under enforced step ceilings, retrieval precision over realistic private corpora, and end-to-end agent task completion with scope checks on. Two columns matter: y0-fast, the interactive profile most requests use, and y0-deep, the deliberate profile behind reasoning and agent planning. Numbers are re-measured on every release by the evaluation family, judged against frozen rubrics, and the history is kept — when a number moves, the changelog says why. Read the notes column; a benchmark without its caveat is an advertisement.