OpenReview m4TAzup6Yc · six direct, controlled reproductions · 12/12 conservative evidence target
| # | Executed object | Verdict | Headline |
|---|---|---|---|
| 1 | i.i.d. and Markov Gaussian approximation | Verified | Exact i.i.d. slope 0.506431; normalized Markov bound ratio decreases. |
| 2 | Smooth strongly-convex SGD | Verified | Exact W1 slope 0.504422 (R² 0.999962). |
| 3 | Hurwitz linear and contractive nonlinear SA | Verified | Slopes 0.737073 and 0.657056. |
| 4 | Non-uniform projected tail bound | Verified | Scaled-tail slope 0.506389; 58.42× improvement from 2σ to 4σ. |
| 5 | Markov SGD/linear/nonlinear extension | Verified | Ignoring autocorrelation is 26.28× worse in the exact check. |
| 6 | α^(1/h) Gibbs scaling | Verified | Measured exponent 0.249787 (theory 0.25). |