Эксперимент: ESS-Aware Byzantine Gradient Fusion (#1130)

{ "worked": false, "confidence": 9, "verdict": "Built a reproducible ESS-aware Byzantine fusion MVP with bounded-rational p_m dynamics, q estimation, MAP fusion, algebraic checks, and a Monte Carlo sweep over attack rate, beta, and honest error. The core boundedness and q endpoint formulas passed, but the paper/idea's claimed universal P_a=1/2 threshold is not implied by the supplied q formula: at eps=0.12 and beta=0.35, q(P_a=0.5) is 0.2185 when p_m=0 and 0.7815 when p_m=1, with implied q=0.5 thresholds 1.586 and -0.586. Therefore the central claim was not reproduced, and no justified adaptive improvement was established.", "metrics": { "baseline": "Majority accuracy in the toy sweep remained 1.0 for beta=0.25 across the tested attack rates; for beta=2.0, eps=0.05 it fell from 0.9958 at P_a=0.5 to 0.0008 at P_a=0.9.", "idea": "MAP matched majority in low/moderate attack regimes and improved over majority for beta=2.0, eps=0.05 at P_a=0.8 (0.8875 vs 0.1442) and P_a=0.9 (0.9983 vs 0.0008), but this used the known global q/attack prior and is not evidence that the proposed adaptive estimator works. Math check: q(P_a=0.5,p_m=0)=0.2185 and q(P_a=0.5,p_m=1)=0.7815." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json", "run.log" ], "limitations": "This is a binary-report toy experiment, not CIFAR-10 or neural federated averaging; no coordinate gradients, median/trimmed mean/Krum, client holdout estimation, realistic payoff estimation, or learned per-client q_i were tested. The MAP sweep uses the known attack rate and beta, so its favorable high-attack results are an oracle-informed diagnostic rather than a deployable adaptive comparison." }