Эксперимент: Log-Hölder Lyapunov Trust Region (#694)

{ "worked": false, "confidence": 9, "verdict": "Built a two-mode conformal Markov-cocycle verifier and a small CUDA RNN trust-region comparison. The exact math checks confirmed zero Lyapunov gap and the expected local perturbation slope (observed 0.999999 versus predicted 1.0), while an inverse-log envelope held only when calibrated over the full tested range; calibration on tiny perturbations had 0/7 held-out coverage, showing it is a loose local bound rather than a reliable practical controller. In the RNN test the controller backtracked all 90 updates in every seed and reached 0.539 accuracy versus 0.983 for baseline, so the proposed adaptation did not produce a training win.", "metrics": { "baseline": "3-seed mean final loss 0.0453, accuracy 0.9831, 0 backtracks per seed", "idea": "3-seed mean final loss 0.6864, accuracy 0.5391, 90 backtracks per seed; exact cocycle gap 0.0 and local log-log slope 0.999999" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 verify_math.py && /home/maxwelhelp/main/bin/python3 run_experiment.py", "files": [ "run_experiment.py", "verify_math.py", "results.json", "results_stdout.txt", "verification.json", "verification_stdout.txt" ], "limitations": "Only a tiny synthetic sequence task and three seeds were tested, not permuted MNIST or a realistic learned multi-mode recurrent model. The controller used a scalar spectral-growth proxy for the RNN matrix rather than full QR exponent estimates at every optimizer proposal, and no careful constant-estimation or FLOP-matched hyperparameter retuning was performed." }