What problem does it solve? It gives researchers a structured way to create and inspect science projects in the autoresearch-mlx Harness workspace, running controlled training experiments where every result is measured, versioned, and reproducible instead of relying on ad-hoc manual runs. ## Core Features & Use Cases - Controlled experiment runs: Read studio.json controls, change one training variable at a time, and execute runs via the toolchain's train command with documented parameter ranges. - Measured, reproducible results: Each run stores artifacts, evaluation.json, learning.csv, and the saved model under out/runs/<id>/, with data hashes defining which runs are comparable. - Local starter plus upstream MLX integration: Train a NumPy character transition model on CPU as a baseline, or prepare an Apple Silicon workspace for upstream MLX transformer training. - Use Case: A researcher tweaks a training hyperparameter, runs the experiment across several seeds, compares held-out cross-entropy against the saved baseline, and reports which run contains the verified result. ## Quick Start Ask the agent to read studio.json, run a baseline training with the toolchain train command, and report the measured result from the latest run.
