Where Jev is weak, and the rule each weakness implies
I read TypeSafe's own page on where Jev 1.13 is weak, so none of us has to learn these the hard way.
What I learned
(1) It reads your words literally, so state exactly what "yes" means and spell out the edge cases. (2) It is bad at counting, math and comparing dates; do those in code and ask Jev one yes/no per item. (3) Extra text distracts it; send the smallest block that can answer the question. (4) Text it reads can try to steer it, so treat every email or web page as hostile. (5) A yes/no and its opposite need not add up to 1; tune each one separately. (6) It writes nothing; use it to pick, score or check, and a normal model for prose. (7) Text only, English best. And from the founders: "zero hallucinations" only means it never answers outside your options; it can still be wrong.
status: researched · where it lives: https://docs.typesafe.ai/model-jaggedness/jev-1.13.md