WeirdBench tests modern LLMs on the weird corners other evals skip.

Unconventional tasks, clear score direction, and a dead-simple score table. The benchmark definitions and code are written locally and published openly.