Эксперимент: Feasibility-Margin Training and Intervention Control (#843)

{ "worked": false, "confidence": 9, "verdict": "Built an exact scalar interval/projection toy and a closed-loop MLP comparison with width sweeps. The mathematical checks confirmed the predicted feasibility boundary x*=2.8, hinge-loss derivative, and exact projection-correction scaling; the single feasibility-count discrepancy is floating-point sign noise at M=0. However, margin training produced no measurable change in fifth-percentile margin or intervention statistics because the saturated projected action made the rollout insensitive to the policy, so the promised training effect was not demonstrated.", "metrics": { "baseline": "At widths 0.55/0.70/0.85: p05 margins 0.1398/0.3572/0.5747; intervention rate 1.0 throughout; correction RMS 0.6958/0.5689/0.4432.", "idea": "Margin and margin+baseline matched baseline on p05 margin, intervention rate, and correction RMS to numerical precision; no observed training win. Math sweeps: boundary error 4.4e-16, hinge derivative RMSE 4.2e-7, projection scaling error 0." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "This is a small deterministic toy rather than cart-pole or robot dynamics; no disturbance-observer uncertainty, policy-gradient training, multi-seed confidence intervals, or physical safety evaluation was tested. The toy’s action saturation also prevents the margin loss from influencing future states, exposing an unfavorable experimental design rather than validating the proposed intervention mechanism." }