Decentralized, physics-driven coupled learning in resistor/flow networks - does the substrate's own physics compute the weight update?

Hardware-transfer arm of coupled-learning-networks (01): toward a 3D-printed flow network that learns by directed aging. Verdict-gated, H5-first.

Coherent photonic MZI mesh + co-located adaptation: does the medium's own physics compute the weight update, or a CPU?

A detailed-balanced chemical reaction network realized as a gradient flow, where equilibrium propagation is rigorous — the chemistry computes its own weight update (verified: local update == true gradient). Audits 'who computes the update?' across a co-location spectrum vs an offline ceiling. Part of physical-learning-substrates.

Differential memristive crossbar with a hardware-friendly in-situ (Manhattan/sign-rule) learning rule, tested on parity-3 — the calibrated in-memory-compute baseline of the physical-learning-substrates portfolio. Verdict #1: PASS, learns parity-3 at SNR ~24.5 (half co-located: physics activations, off-array error sign, physical-pulse increment).

Energy-based (Ising/Boltzmann) learning substrate for parity-3 — can a local, physics-native contrastive rule both compute and learn the couplings? Part of the Physical Learning Substrates portfolio.

Aspirational floating-gate arm of memristive-crossbar (02): can in-fabric learning run on a real analog floating-gate array (tunneling/injection writes)? Verdict-gated, H5-first, provisional.

The meta-project / commons of the physical-learning-substrates portfolio: records, adapts, and synergizes the common lessons of projects 01-05 and the self-assembly parent line. A knowledge commons + a substrate-agnostic code commons + a harvest tool, joined by a human-triggered synergy loop.

Hardware-transfer arm of memristive-crossbar (02): does in-fabric learning survive a real floating-gate FPAA? Verdict-gated, H5-first.