Crucible a neurosymbolic stack pairing mamba2 with z3 verification, JEPA, and HDC memory

Hi everyone! Ive been working on crucible, an experimental transformer free stack built around mamba/mamba2 and an explicit symbolic verification loop. it uses a z3 solver so the model can emit specs, verify them, and retry on failure. verified patterns get saved into a hyperdimensional computing associative memory, while jepa handles hidden space concept prediction and micro lora handles test time adaptation. the whole pipeline is configured to train and adapt within a free kaggle t4 session. its still an experimental prototype, but feel free to check it out or give feedback on the code. also i am currently training a model using this and when ill be finished i probably will post that here! https://github.com/JustRoccat/Crucible/