Covering-Relation Optimizer Corridors

Integrate this at the optimizer level rather than inside individual network layers. Start with SGD or SGD with momentum on a small multilayer perceptron. Represent the optimizer state by two or three coordinates such as minibatch loss, gradient norm, momentum norm, or two principal components of the parameter displacement plus one transverse residual coordinate. Define boxes N_i around an empirically observed training trajectory, orient u along progress toward lower loss, and orient s along transverse variation. For every box, sample its corners and several interior points, apply one optimizer update using a fixed minibatch or deterministic full-batch gradient, and compute coordinatewise output intervals. Accept a link N_i => N_(i+1) only when the output interval stays away from the transverse boundary, crosses the required progress faces, and overlaps the next box. Use automatic differentiation to estimate local Jacobians and calculate the cone margins. Pseudocode is: construct boxes; evaluate optimizer updates at corners; enlarge bounds by a safety factor; test boundary avoidance and face crossing; reject or shrink the learning rate if any link fails; then train normally while periodically checking the certificate. The paper supplies the covering and cone logic; interval widths and Jacobian bounds are estimated numerically. The first cheap experiment should use a two-layer MLP on MNIST and compare ordinary SGD with corridor-constrained SGD over a logarithmic learning-rate sweep. The predicted signature is a sharp certification boundary eta_c where either the transverse enclosure first touches ||s||_infinity = 1 or the progress crossing margin becomes zero. Near certified settings, measured transverse amplification should stay below kappa_s and progress amplification should exceed kappa_u. Test whether eta_c is reproducible within 20 percent across seeds and whether certified runs have fewer corridor exits.