Quantum Error Correction Meets Graph Neural Networks: Decoding Graphs, Ramanujan Expanders, and the Code-Agnostic Decoder
QEC is becoming a graph-learning problem. We analyze the October 2026 GNN-decoder wave (POLYMECHANON's code-agnostic decoder, modular neural decoders, L-NBP) and add our own spectral analysis: the Z-syndrome graph of the toric code Z_d² IS the Cayley graph C(Z_d², ±e1, ±e2), whose second eigenvalue 2+2cos(2π/d) crosses the Ramanujan bound 2√3 at d≈7-10 — exactly where BP fails and learned decoders take over. Includes a 5-code sum-product BP benchmark showing spectral gap alone does not predict decoding success.


