Both lanes see byte-identical questions. The modelless lane is ns–µs tier and bit-deterministic on a given build; the open-weights laya lane is the accuracy heavyweight. No single "we win" number exists or will exist — audit the tables. The modelless lane's near-chance rows on out-of-domain text are structural (it is corpus-bound by design), and they are published here rather than trimmed.
loading data/bench.json …
Not affiliated with, or endorsed by, TypeSafe AI; Jev is their product, named only to compare against. The laya lane loads Apache-2.0 checkpoints from the brainfunctioncollapse laya Hugging Face repositories at runtime — unmodified, SHA-256-verified, never redistributed.