claim,anchor,verdict,scope
1,Theorem 3.7 / thm:convergence,supported_with_assumptions,"Requires a depth-D path satisfying M-coverage, bounded second moments, and bounded global-logit l1 coefficients."
2,Theorem 4.5 / Lower Bound on Convergence,supported_with_scope,"The construction has D=kp and p<=k-1; matching refers to the network-depth bottleneck, not identical upper/lower exponents."
3,Lemma 3.1 / lem:orthogonality,supported,First-order optimality for the optimal logistic predictor on its feature space.
4,Lemma 3.3 / lem:pythagorean,supported_with_scope,Both p* and q are logistic predictors on the same feature set and p* is the optimizer.
5,Section 2 / Sequential Learning Protocol,supported,"Agents pass logits, not probabilities, in a topological order."