Эксперимент: Capitalization-Efficiency Monitor (#877)

{ "worked": true, "confidence": 8, "verdict": "Built a capitalization-efficiency monitor MVP using posterior-to-prior KL as the information ledger and held-out value gain as the numerator. The exact toy sweep confirmed the quantitative predictions: KL slope 2.000 vs 2 predicted, small-step value slope 0.999 vs 1, efficiency slope -1.001 vs -1, and nuisance efficiency -1.600 vs -1.6; the value transition was bracketed at [1.518, 1.800] around the predicted 1.7. In the neural sanity experiment, monitored Adam achieved lower test MSE and slightly lower total KL than baseline, but this is only a small supporting signal rather than proof of general benefit.", "metrics": { "baseline": "Adam: test MSE 0.18153, total KL 1.72803, late efficiency -0.0583, 180 updates", "idea": "Monitor Adam: test MSE 0.16839, total KL 1.67706, late efficiency -0.5277, 180 updates; toy predictions observed: KL slope 2.000, value slope 0.999, efficiency slope -1.001, nuisance eta -1.600" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json", "README.md" ], "limitations": "The MVP omits optimizer-noise dissipation Sigma, uses fixed posterior variance and squared parameter displacement as the KL approximation, and uses direct held-out regression value rather than paired deletion-counterfactual task evaluations. Only one tiny seeded dataset and one neural architecture were tested; the monitor's trust-region behavior was not evaluated across multiple seeds or compute-matched benchmarks." }