ARC-AGI Reproduction Project

Project goal
Recreate the system described in A Pre-Emergence Mechanics Framework for Solving the Abstraction and Reasoning Corpus ARC-AGI by Achilles. The codebase implements a boundary constrained field solver (Intent Tensor Theory substrate) and a beam search driver to propose transforms that reduce L1 residue against ARC tasks.

Current status

Core driver itt_solver/experiment_driver.py runs experiments and saves artifacts to experiments/.
Runs complete and save *_phi_best.npy, *_logs.json, *_result.json.
Immediate blockers discovered and addressed: W&B recursion removed; logging lacked candidate arrays; some transforms were no-ops. Diagnostics and small patches are in experiments/ and itt_solver/ to inspect and fix behavior.
Smoke tests show the beam accepts candidates but residue has not decreased for the example task. See experiments/ for logs.

Quick start minimal run

Install dependencies (if not present): numpy, matplotlib.
From project root run a single job in a notebook cell:

from itt_solver.experiment_driver import default_atomic_factory, run_single
# build task and params (example provided in repo)
res = run_single(task, default_atomic_factory(params, task), params, out_dir="experiments")
print(res)