Glitch Token

How Glitch Token Works

Somewhere in this vocabulary lurks a SolidGoldMagikarp — a token so cursed it makes the model speak in tongues. It was never trained on, only tokenized, and now it sits in the embedding matrix like a landmine. Your job: find it before it ships to a billion users.

1. The grid is the model's token vocabulary. Exactly one token is glitched.
2. Tap a token to run inference on it. You'll see the output it produces.
3. Clean tokens decode cleanly. Tokens near the glitch come out increasingly garbled, looping, and unhinged.
4. You have a limited probe budget — every probe costs compute, so triangulate wisely.
5. When you've found the strongest corruption, tap the cursed token to select it, then Accuse it.

Reading the Corruption

Corruption strength is a heat signal: the closer a probed token sits to the glitch on the grid, the more broken its output. A token that returns pure noise is a direct neighbor. A token that decodes almost-cleanly is far away. The glitch itself is silent — it just breaks everything around it.