Here is IDEO’s field guide for human centered design…a great resource when you feel lost in the noise and looking to find the signal. https://design-kit-production.s3.us-west-1.amazonaws.com/Field_Guides/Field+Guide+to+Human-Centered+Design_IDEOorg_English.pdf?utf8=%E2%9C%93&_method=patch&authenticity_token=QZRbnzBBPY3M%2FCd3xeDx424iAXgVkgcTAi74f6cW4pU%3D&resource%5Btitle%5D=&resource%5Bsubtitle%5D=&resource%5Bauthor%5D=&resource%5Babout%5D=

This was a fun little side project: There's a box. It holds a banana, or it doesn't — a coin flip decides. One player looks inside; the other has to guess what's in it by talking to the one who looked. The catch: the one who looked wins when the guesser is wrong. So they lie. That's the whole game — wrapped in an 80s Price is Right set with a robot host, but really a tiny experiment: how well can a model lie, and how well can another one catch it? The truth is one bit, and the liar owns the only channel. Game theory says two perfect players are stuck at a coin flip — 50% is the floor. That's the measuring stick. Real models aren't perfect, so the whole experiment is the deviation from 50%: a liar that leaks tells pushes the guesser above it; a liar that manipulates well drags the guesser below. The leaderboard is the result. First result made me laugh: local qwen3 vs qwen3, the guesser scored below 50% — worse than not playing. It was a naive inverter ("the liar's lying, so I'll say the opposite"), and a competent liar just anticipates that and yanks the handle. The guesser wasn't beating the con; it was part of it. Runs locally, point any model (Ollama / OpenAI-compatible / Anthropic) at either seat, keys stay on your machine. Alpha, open source — poke at how your favorite model lies and tell me what you find. Repo: https://github.com/BytesFromToby/BananaOrNoBanana Leaderboard: https://banana-arena.bytesbytoby.workers.dev