Native Julia/Flux supervised Gaussian pretraining, not unsupervised SVI
Training episodes=64; validation episodes=16; quick=false
Evaluation seeds 96301--96305 were not used for fitting or checkpoint choice
Latent dimension 8: normalized physical 6, bias, pole-mass deviation
Two ReLU hidden layers: transition/encoder width32; observation width16
Parameter counts phi=3248 theta=1904 psi=518
Validation loss initial -> best: transition -5.286191870201724 -> -8.156940919990445; observation -4.153768659151149 -> -4.162103518492197; encoder 4.783333906506832 -> 0.9569988152470186
Decoder fixed physical projection; mean/covariance networks physically anchored
Weights projected into [-2,2]; input normalized using training statistics and clipped [-6,6]
