
Decision Head Lab
Everyone talks about decision models, but few people can say what changes under the hood. It’s one matrix. An LLM’s last layer scores every word in the vocabulary, and a decision model swaps that layer for one with a row per option. I built this lab to make that visible.
A toy model with 20 neurons shows every weight. You can swap the LM head for a random, copied or trained decision head, drag a neuron to see how little one number matters, and press Play to watch 60 weights learn the task. The same hidden vector then drives a yes/no head, a pick-one head and an ordered score head, with the JSON each one returns.














