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 6: normalized physical 4, bias, pole-mass deviation
Two ReLU hidden layers: transition/encoder width32; observation width16
Parameter counts phi=2508 theta=1708 psi=452
Validation loss initial -> best: transition -5.283914274401522 -> -8.109421749140566; observation -4.482077054698502 -> -4.489808337359221; encoder 4.004796861462309 -> 0.865231973645045
Decoder fixed physical projection; mean/covariance networks physically anchored
Weights projected into [-2,2]; input normalized using training statistics and clipped [-6,6]
