Эксперимент: Support-Sparse Koopman World Model (#1102)

{ "worked": true, "confidence": 8, "verdict": "Built a support-sparse Koopman autoencoder with reconstruction, latent prediction, multi-step rollout, L1 sparsity, and support-persistence losses, plus a two-basin double-well experiment. The three mathematical predictions were numerically confirmed: support derivative error 0.08%, rollout scaling error 0, and small-perturbation support scaling error 0.48%. Support persistence reduced adjacent support change by about 17x and basin-mask accuracy rose from 68.3% to 80.0%, but 20-step MSE worsened from 0.01398 to 0.01589, so the representation mechanism appeared without a forecasting win.", "metrics": { "baseline": "20-step MSE 0.013982; mask/basin accuracy 0.6833; support step change 3.3441e-05", "idea": "20-step MSE 0.015889; mask/basin accuracy 0.8000; support step change 1.9406e-06; active fraction 0.2458" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "Only one random seed and one small synthetic double-well system were tested. The dense and sparse models were not exhaustively parameter-matched or hyperparameter-tuned; no 40-minute super-check, longer horizons, multiple seeds, or rigorous held-out threshold selection was performed." }