Publishing a per-version failure list is unusual, and it helps because the entries are specific. Date and time comparison is one of them: “Jev reads dates as text, not as ordered quantities”, so asking whether one timestamp comes before another is unreliable. Math and numbers is another, with arithmetic and counting called out directly. Adversarial content is a third worth planning around, because “State is data, and jev-1.13 does not treat it as hostile by default”. The fourth is structural: there is no guarantee that P(A) and P(not A) sum to 1 across related questions, so two questions that logically constrain each other can both come back confident and mutually inconsistent.
Each of those has a fix on your side. Parse dates and compare them in code, which is deterministic anyway. Do the arithmetic yourself, and let the model supply only the per-dimension judgments that feed it. Screen untrusted state with a separate injection-detection question before the answers reach anything that acts.
One entry on the list is a design boundary rather than a defect: “Jev-1.13 is not trained to generate text.”
See the prompt injection screening recipe and the hallucination entry for what jaggedness replaces as the thing to watch.