I'm building an agent architecture called Nofae. The core idea is combining three components that are usually studied separately: world models, recursive transformers, and multimodal encoders and adding a Z3 SMT solver as a formal verification layer on top.

The system combines: 1. World models where learning a compressed, predictive simulation of the environment 2. Looped transformer where depth through iteration rather than parameter expansion, shared weights 3. Multimodal encoders where there's grounding language, vision, and audio in a shared representation 4. Z3 verification where neural policy proposes plans, Z3 checks logical consistency before execution 5. Metaheuristics where Occam, uncertainty-gather, contradiction-backtrack governing loop termination

Early result on a sorting task:

Loops: 1 | Final Loss: 1.0187 Loops: 2 | Final Loss: 0.9818 Loops: 4 | Final Loss: 0.9607 Loops: 8 | Final Loss: 1.0031

1→2→4 loops improves performance, 8 degrades.…