Kinetic-AG: rebuilding the foundation

I started Kinetic-AG as a project about using symbolic operations and deterministic grid tools to help an AI reason through ARC-style visual puzzles. Then I left it alone for a while.

Coming back to it, my first priority was to re-ground myself in what the code actually does before adding anything else on top. Some earlier progress reports made the project sound further along than it really was: scanning ARC files and extracting objects is not the same as solving ARC tasks.

Over the past few hours, I’ve been reviewing and hardening K‑VL, the library that handles the grid operations. I fixed a bug where one same-color shape could leak into another shape’s mask, added stricter checks for malformed grids and objects, replaced an expensive distance calculation, and removed Python code execution from the training verifier. I also wrote down the library’s rules and added tests for the behaviors I’m relying on.

This is still a prototype. I haven’t tested it on the ARC‑AGI‑2 tasks; I’m saving those as a later benchmark once the library is more dependable. My next goal is to keep tightening the contracts and tests before I build the model-facing DSL.

It feels good to return to this project with a clearer idea of what “working” should actually mean.