ALA Experimental Architecture

This repository contains experimental implementations of the Adaptive Learning Architecture.

Layout

• model/ - model implementation files
• model/legacy/ - older v7-style demo implementation files
• docs/ - architecture specifications
• tests/ - experiments and older compression tests
• runtime/ - generated states, checkpoints, and caches
• archives/ - backup zip files

Current Trainable Target

Use model/ALA_v8.py for the v8.1 trainable prototype. It implements a prediction-error-driven loop, bounded primitive memory, and an adaptive strategy layer. The default configuration is intended to be roughly a 5M-parameter model.

Colab Smoke Train

pip install -r requirements.txt
python train_v8_colab.py --steps 1000 --batch-size 64

Checkpoints and metrics are written to runtime/v8_runs/.

The older files are kept in model/legacy/ for reference and comparison, but they are closer to the v7.1 demo architecture than the v8.1 training target.