Author here.

I've been working on this inside another project called mouse.dev. It started to take on its own life, so alongside claude and few other models i distilled my thought process into something hopefully more scientific and research worthy.

The idea for Reinforcement Dreaming came from a single thought of "What if every night while I sleep an agent could imagine all the possible futures of my codebase?"

Each dream run gets its own sandbox. The system keeps the actual execution history outside the model’s context, and a separate process has to reproduce the result before anything survives.

I don’t have benchmark results yet. Section 11 is basically the experiment I think needs to be run.

Would especially love to hear from anyone who has insight into once an agent knows what the evaluator rewards, how do you keep it optimizing the actual work instead of optimizing the evidence?