Not all research succeeds. We document our failures to save time—ours and others'.
Each failed attempt contains information: boundaries of approach validity, hidden assumptions, and often seeds for future success. We publish detailed lab notes for select failures.
Attempted to build a single geometric framework unifying different causal inference approaches. Failed due to incompatible independence assumptions across methods.
Read detailed lab note →Tried to encode all coordination logic at the type level for compile-time guarantees. Became intractable for real-world agent populations; abandoned in favor of hybrid approach.
Built neural physics engine trained entirely on real-world data. Achieved impressive interpolation but catastrophic extrapolation failures. Hybrid symbolic-neural approach proved necessary.
If you attempted similar approaches or encountered related issues, we'd value hearing about it. Shared negative results accelerate progress for everyone.