Эксперимент: Gumbel escape-time controller (#1126)

{ "worked": true, "confidence": 9, "verdict": "Built and numerically verified the Gumbel escape-time mechanism in a delayed linear toy model. The predicted delay threshold matched to 2.3e-13 relative error, beta·r stayed within about 5% of its predicted constant, Gumbel fits strongly beat exponential fits, and logarithmic boundary scaling matched within 1.2%. This validates the mechanism, but not yet an optimizer or MNIST training win.", "metrics": { "baseline": "Exponential escape fit: consistently much worse AIC; no optimizer/MNIST baseline was run.", "idea": "Threshold tau_c=1.5707963 observed numerically; beta*r={0.996,1.005,0.964,1.049}; log-boundary slope=2.248 versus predicted 2.222; all Gumbel KS p-values exceeded 0.05." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json", "REPORT.md" ], "limitations": "The experiment used explicit extreme-value seeds rather than arbitrary minibatch noise. No true noisy double-well integration, stale-gradient optimizer, MNIST MLP, GPU benchmark, or measured validation-loss improvement was tested; the controller's validation signal is externally supplied." }