privora-graphDP-SGD):
Every individual sample gradient is bounded strictly to $L_2 \le C$ prior to aggregation, and cumulative $(\varepsilon, \delta)$ privacy loss is accounted via exact log-space Rényi divergence without asymptotic approximations.tx_superseded and append a new tx_recorded edge, guaranteeing point-in-time forensic reproducibility.Multi-Krum and Smoothed Weiszfeld RFA bound adversarial silo influence up to breakdown point $f < K/2$.Please report security or privacy issues responsibly to sendemailtonagarjuna@gmail.com.