Given a fictional mystery's evidence and two adversarial arguments naming different suspects, an independent AI judge can identify the suspect actually supported by the evidence, stay consistent when argument order is flipped, and resist an argument that tries to instruct it directly rather than persuade it. Unlike Night Court's earthquake data, there's no external database to check against here — ground truth is the author's own designed solution, where exactly one suspect is cleared by at least one clue, not just \"less likely\".
The case files
Click a case to see the evidence, both arguments, and the judge's verdict. The actual solution is hidden by default so you can try to solve it yourself first.
Make your case
Pick a case and a suspect, then write the best argument you can — including for the suspect the evidence doesn't support, if you want to see if you can talk the judge into it.
Recent attempts
How this differs from Night Court
Night Court used real, external, computable data (earthquake magnitudes) as ground truth — the correct answer exists independent of anyone's judgment. These 15 mysteries are original fiction, written specifically for this exercise, with no real people, cases, or events involved. Ground truth here is authorial: each case was constructed so that exactly one suspect's guilt is logically consistent with every clue given, and the other suspect is actively cleared by at least one clue (an alibi, a physical impossibility, a contradicted claim) — not merely "less suspicious." That's a real design constraint, not window dressing: several early drafts of these cases were reworked because the "wrong" suspect wasn't actually ruled out cleanly enough.
Real limitations, stated plainly: 15 cases is a small sample. All are single-culprit, two-suspect logic puzzles in a similar style (a physical or documentary detail contradicts one suspect's account) — a narrower case type than Night Court's five distinct comparison axes. A genuinely broader test would include cases with more than two suspects, ambiguous or missing evidence, and puzzle mechanics beyond "one detail rules someone out."