Эксперимент: Chern-Gap Monitor for Finite-Horizon Collapse (#1043)

{ "worked": true, "confidence": 8, "verdict": "Built the finite-grid spherical-triangle Chern monitor with QWZ test fields, parity proxy, gap-closing sweep, and EMA smoothing, plus a small GRU event-order experiment comparing baseline training against gap regularization. The math check clearly showed integer Chern sectors away from transitions, linear gap closing with slope approximately 1, and 91.8% EMA noise-MSE reduction. In training, regularization increased the final probe gap from 0.0914 to 0.1849, but validation loss and accuracy were essentially unchanged (0.7027/50.4% baseline versus 0.7021/50.4% regularized), so there is no demonstrated accuracy win.", "metrics": { "baseline": "QWZ K=80: Chern values approximately 0, +1, +1, -1, -1, 0 at masses -3, -1.5, -0.5, 0.5, 1.5, 3; GRU final validation loss 0.7027, accuracy 50.4%, probe gap 0.0914.", "idea": "All three critical masses had gap 0 and gap-vs-|delta mass| slope 1.0; EMA reduced injected-field MSE by 91.8%; GRU gap-barrier final validation loss 0.7021, accuracy 50.4%, probe gap 0.1849." }, "how_to_run": "python3 run_experiment.py && /home/maxwelhelp/main/bin/python3 train_experiment.py", "files": [ "chern_gap_monitor.py", "run_experiment.py", "train_experiment.py" ], "limitations": "The Chern estimator was verified on an analytic QWZ field, not extracted from the GRU’s learned response surface. The toy classifier used one fixed seed and a short CPU fallback run after shared-GPU cuDNN allocation failure; no early-stopping controller, multiple seeds, reflection-symmetry-breaking test, or larger sequence benchmark was evaluated." }