Modular learning experiments

Experiments on combining independently trained model components and checking whether updates preserve earlier performance. Includes code, individual results, mathematical derivations and reproduction commands.

Findings

• Learned components beat random features on the main synthetic benchmark. Longer pooled training performed better than either.
• Comparing updates with an owner's joining checkpoint improved early learning. Strict retention checks still stopped all updates in the last half of every tested run.
• Simulated peers reconstructed the same combined predictor across 40 releases.
• An earlier real-data test found no advantage for the proposed learner.

These are separate small prototypes. They have not been tested together as a distributed language model. The peer simulation assumes honest participants and provides no privacy or security guarantee.

See results and supporting files and assumptions and derivations.

Run the checks

Python 3.12 and NumPy 2.3.5. No GPU, account or API key is needed.

python3 -m venv .venv
. .venv/bin/activate
python -m pip install -r requirements.txt
python -B verify.py

This checks artifact hashes, summaries, intervals, saved predictions and the peer simulation. To also retrain selected cases:

python -B verify.py --refit

Checks run sequentially with one numerical-library thread. Reproduction details describe the coverage and limits.

Files and reuse

experiments/ preserves source, protocols, development failures and results from three research rounds. ARTIFACTS.json records the public file hashes and metadata removals. Release notes explain the export.

Code and documentation use the MIT license. The included Yacht dataset retains its CC BY 4.0 attribution.