Comment by arjmandi

I wrapped a human-in-the-loop RSI loop around ASSAY (my open-source reasoning harness) and ran it on the 25 public ARC-AGI-3 games with Haiku 5.5. Assay also introduces 5 levels of abstraction for reasoning and has equivalent plugs for each layer. This design was used to guide the analyst model in this experiment to recommend fixes at the layer where the reasoning broke, and the harness runs again.