Test a stand-up set, speech or pitch on a virtual audience of 200 people. Returns each line's crowd reaction (laugh, chuckle, moved, groan, gasp, applause or silence), whether the line is a joke or sincere, a laugh score and a warmth score, and a 0-100 landed score that judges jokes on laughs and heartfelt lines on warmth. Weakest/strongest are ranked by landed score; use them to suggest rewrites. Pass audioPath to judge a recording: it is transcribed, split into timed lines, and pace and pauses are taken into account.

Judge every sentence of a draft against its goal. Flags sentences that are off-goal, filler, repetitive or likely invented facts (or unsupported by the given sources). Rewrite only the flagged sentences, then call again to confirm.

Find the commit that introduced a problem a normal test can't detect. Runs command at commits between good and bad in a temporary git worktree (the user's checkout is untouched) and asks Jev whether each output shows symptom. Runs the user's shell command, so confirm it with the user first.

Resolve what happens when any two things meet (a rubber duck and a laser, a toddler and a vending machine) using common sense instead of hand-written rules. Returns the outcome, intensity 0-10 and likely effects with probabilities. Narrate the result vividly.

Drop 2-8 things into one setting; every pair interacts in parallel and events come back sorted by intensity. Use it to narrate a chain of chaos.

Find surprising, useful combinations in a pile of notes by scoring every pair with Jev. Give either a folder/file path or a list of notes. Returns the top collisions; develop each into a concrete idea. maxPairs caps cost (2,000 pairs is typically well under $0.05).

Low-level access: evaluate a state against your own typed questions. Question types: {"type":"noul","instructions":"..."} (probability 0-1), {"type":"choice","instructions":"...","criteria":{"key":"description"}} and {"type":"score","instructions":"...","criteria":["level 0","level 1"]}.