Control-aware physical neural-network regulation research software

Project-maintained research software for the manuscript “Control-aware training of physical neural networks for closed-loop regulation.”

Maintained physics-constrained estimators

• InterferometricOracle: coherent multi-path phase encoding, interference and intensity readout.
• NonlinearOscillatorOracle: 24-node damped coupled Duffing-type dynamical system.

Confirmatory record

• E-010C1: matched-static interferometric behavioral separation.
• E-020C1: interferometric robustness success.
• E-030C1: oscillator matched-static threshold failure retained.
• E-031C1: oscillator robustness success.
• E-050C1: task-matched Pareto success.
• E-051C1: exact decision-regret mechanism success on both substrates.
• E-060C1: canonical Cart-Pole task-generalization success; 54.63% reduction in the predefined moderate+strong failure-padded stabilization-loss endpoint.
• E-040: descriptive external measurement-scale grounding only, not hardware validation.

Reproducibility

python -m venv .venv
# activate environment
pip install -e '.[test]'
pytest -q

Source-tree invocation:

PYTHONPATH=src python -m pytest -q

Frozen confirmatory runners verify their checkpoint/seed/configuration manifests. Project-internal freezing is not claimed as externally time-stamped preregistration.

Public archive

Target GitHub release: v1.1.0.
Zenodo v2 DOI for public release v1.1.0: 10.5281/zenodo.22069483.
Persistent all-version concept DOI: 10.5281/zenodo.22068582.

No third-party research source code is copied or vendored. Generative-AI assistance is documented in the manuscript and repository documentation; the author remains responsible for the scientific work.