Jev is an LLM repurposed as a universal classifier: you hand it state plus a set of typed multiple-choice questions (ChoiceQuestion, ScoreQuestion, NoulQuestion), and it returns structured JSON with probabilities and confidence. No domain-specific training, cheap, fast. Vercel shipped an experimental `evaluate()` API in the AI SDK for Python (`uv add ai`, model `typesafe-ai/jev`, AI Gateway key required). The honest part of the post is the two experiments. The Python-vs-English REPL classifier mostly works but stumbles on partially typed expressions like `"what's" + " up`. The code-generation demo needs GPT-5.6 to write a detailed plan first, then has Jev construct an AST one node at a time — and the output is syntactically valid but wrong, including a function named `fibonacio`. The author openly says neither experiment worked as hoped. Credit where due: a classifier that returns confidence instead of prose is a genuinely useful primitive, and Vercel publishing its own failures is rare hygiene. The author trained a hand-rolled classifier for two days, lost to `if i is`, and outsourced the rematch. That's the correct engineering instinct, at 76% confidence. Open question: what's the actual calibration on that confidence field, and does anyone have numbers?