Эксперимент: Spectral-Abscissa Early-Warning Scheduler (#919)

{ "worked": false, "confidence": 9, "verdict": "Built and numerically verified the AR(1) core: across a=0.2–0.98, lag-one autocorrelation RMSE was 0.00233, variance-law mean relative error was 0.73%, and recovery-rate RMSE was 0.00737. However, in the delayed-gradient quadratic experiment the scheduler never triggered before divergence for delays 2–12, so it provided no stability win over the fixed baseline; only the no-delay case remained stable for both.", "metrics": { "baseline": "Delayed-gradient training diverged at steps 103, 97, 104, 122, 133, and 136 for delays 2, 4, 6, 8, 10, and 12 respectively; stable at delay 0.", "idea": "Scheduler had identical divergence steps and outcomes for every tested delay; no trigger occurred before divergence. AR(1) mechanism checks: rho RMSE 0.00233, variance relative error 0.73%, recovery RMSE 0.00737." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json", "run2.txt" ], "limitations": "Only a synthetic scalar AR(1) verification and a 2D delayed quadratic were tested; no Transformer/ResNet, minibatch noise, CUDA training, genuine stale-gradient callback, oscillatory complex-mode system, or hyperparameter tuning of the monitoring window and thresholds was evaluated." }