Jax / numpy simulation of a stochastic turing tumble like game

Hopefully can be made traiable =)

Screen.Recording.2025-05-12.at.5.01.24.PM.mov

TODO

ML

• perplexity? / zip compression for randomness? run_sim but dont print choices, print perplexity etc instead
• grad and train
• force prefix

ML / physical? problems:

• low perplexity, BOO state is 50% L/R, and looks like ball will almost never fall far from init pos

Bugs

• some weird step bug: python pboard.py --force-np-random -r -n2 -i 0.0
• -vvv breaks render

Refactor

• logical sim, disentangle from debug/render
• -rr render should show "sampling mask" and "physically achievale states"

Physical board

• boundaries
• multilayer board
  • switch layers from off boundaries
  • or just "drop down on next layer" essentially "holes". How to make them trainable?
  • some "jump forward" holes too?