claim,source,status,scope
1,"arXiv v1 Proposition 3.1, Eq. 16",supported,A1 finite learned-path energy and A2 true Girsanov martingale; KL direction is learned endpoint to training target
2,"arXiv v1 Definition 3.2 and Proposition 3.3, Eqs. 17/20",supported,"A1-A4, i>=i0, epsilon_star^2<=1; C depends on regularity/observability constants"
3,arXiv v1 Theorem 3.4,supported,"A1-A4, positive observability and tail effects; equivalence constants are not one"
4,arXiv v1 Proposition 4.1,literal_conflation_unsupported,non-summability yields non-summable drift/cannot vanish; the displayed positive floor additionally requires a uniform per-generation error floor
5,"arXiv v1 Theorem 4.2, Eq. 23",supported,"A1-A5, positive observability, small summable errors, and C_bias; asymp hides constants"
6,"arXiv v1 Figures 1,5-10 and Appendix G",supported,"10D GMM is quantitative; Fashion/CIFAR divergence uses learned-feature Gaussian proxies and is source evidence, not rerun here"