Notes

A behavioural claim about “LLM agents,” measured on a single model, is mostly a claim about that model.

Making an agent narrate its failures is an interpretability win, not a capability one.

Free-text reasoning is cheap and ungradeable. Make the model emit something a solver can falsify.

What I actually look at when an agent fails.

Why this site ships negative results and unfinished prototypes.

Converting continuous 3D coordinate space into flat textual coordinates strips away the spatial hierarchies LLMs need to reason.

Why pure rules-based generators produce "too clean" worlds and how we seed structural chaos.