Эксперимент: Coarsening-Aware Global-Consensus Scheduler (#1420)

{ "worked": false, "confidence": 9, "verdict": "Built the coarsening-aware scheduler with empirical Q/C diagnostics, gated learning-rate slowdown, formula sanity checks, and a paired fixed-seed MLP experiment. The constructed orthogonal-sector case produced high Q≈1 and low C≈0.334 and triggered the controller, verifying the gate mathematically, but the slowdown was only about 2.4e-6 in that setup. On two-moons, baseline and controller were identical because the observed high-Q/low-C regime never occurred; both reached 96.11% accuracy and validation loss 0.1061, so no promised training win was observed.", "metrics": { "baseline": "final accuracy 0.961111, validation loss 0.106121, mean Q -0.02443, mean C 0.35659, high-Q/low-C fraction 0.0, minimum LR 0.0018", "idea": "final accuracy 0.961111, validation loss 0.106121, mean Q -0.02443, mean C 0.35659, high-Q/low-C fraction 0.0, minimum LR 0.0018; synthetic check Q 0.999997, C 0.333641, eta ratio 0.9999976" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "Only a small two-moons MLP was tested, not MNIST/CIFAR, CNNs, transformers, longer coarsening dynamics, multiple seeds, or block-count scaling. The empirical block-update projection and adjacency construction are lightweight approximations rather than a production optimizer integration." }