A twenty-questions game where Jev's calibrated confidence decides how many questions to ask. It commits above 55%, hedges between 28-45% ("can't say for sure, but probably..."), and below 28% it refuses to answer rather than guess: "I don't know. Who even are you?" Confidence is the game mechanic here, not a debug field. In a guessing game, "how many more questions do I need" is the entire product — so if the confidence is real the game works, and if it is decorative you can watch it fall apart. One request per turn carries up to 18 independent judgments: Choice over 110 personas, 47 prefectures, 8 regions and 14 closing voices, plus 8 Scores, 8 Nouls and a speculative next-question Choice. Measured at 230-440ms, so there is no loading state anywhere — every candidate bar moves the instant you click. The UI narrates each turn, separating what the code decided from what Jev decided. Four patterns did the heavy lifting: a coarse 8-way region Choice used as a prior over the 47-way prefecture Choice; scoped evidence per question via structured instructions, so one request feeds geography answers to one judgment and personality answers to another; eliminating low-probability candidates between turns so confidence can climb over a 110-option set; and selecting a hand-written closing line instead of generating one. The repo ships build notes with measurements for five failures, all of them my mistakes rather than the model's. Passing 47 prefectures as `criteria: null` collapsed geography into two sinks — "rural" became Hokkaido, "no evidence" became Tokyo. Region accuracy went 6/16 to 12/16 once the rubrics were written and a core question set was forced. Counter-intuitively, enlarging that core made it worse again, because only five questions actually get asked. And 97% of every request turned out to be the same criteria resent, so skipping judgments whose evidence had not changed cut input tokens by 52%.