PILOT-Q benchmark: photon-budget-aware training and closed-loop operation of delocalized photonic neural inference at the SQL

# PILOT-Q benchmark: photon-budget-aware training and closed-loop# operation of delocalized photonic neural inference at the SQL Complete data underlying every figure and number in the article"PILOT-Q: physics-in-the-loop training and photon-budget-awareclosed-loop operation of delocalized photonic neural inference at thestandard quantum limit"(T. M. Mahim, M. N. Islam, M. M. Rahman, and A. S. M. Mohsin). ## Contents- `results.json` -- all computed results: twin validation against the analytic shot-noise (SQL) law (N = 4096 and 32768), and for each dataset (MNIST, Fashion-MNIST, FSDD) and protocol (conventional, PILOT-Q, PILOT-Q-full, fixed-budget ablation): clean ceilings, accuracy vs photon budget on the shot-limited and compound-impaired chains (mean +/- s.d. over five seeds), dark/trim/ADC sweeps, full closed-loop controller sweeps, and oracle escalation bounds.- `checkpoints/` -- trained model weights (PyTorch state_dicts).- `csv/` -- flat exports of the headline curves. ## ProvenanceModels are real-valued fully connected networks on three public datasets:MNIST, Fashion-MNIST, and the Free Spoken Digit Dataset(https://doi.org/10.5281/zenodo.1342401). The stochastic twin models thedelocalized photonic architecture of Sludds et al., Science 378, 270(2022), https://doi.org/10.1126/science.abq8271, with exact standard-quantum-limit shot-noise statistics; electronics constants follow Gao etal., Sci. Adv. 12, eadz0817 (2026). ## LicenseCreative Commons Attribution 4.0 International (CC BY 4.0).