Scale_Screen

Checkpoint collection for the Neural Interaction Law shape scans. The experiments/ hierarchy contains the original screen and independent-selection layout, fitted checkpoint weights, result tables, learning curves, convergence audits, and available input-AGOP measurements. Failed preliminary checkpoints are deliberately retained under attempts/ when available.

The source corpora (FineWeb and WikiText) are not redistributed. The relevant model, data-preparation, training, and AGOP scripts are preserved under code/language_model/. FineWeb runs use base D = 60 * active_parameter_count bytes; explicit convergence extensions remain visible in each result table.

MODEL_MANIFEST.csv enumerates every uploaded checkpoint and byte size.