8通道生理狀態路由 Shallow Experts v2

Global-init specialists 使用局部凍結／差別學習率，baseline routing 特徵與 soft-routing 溫度只由 validation 選擇。Outer test 在設定鎖定後才讀取。

鎖定設定

Expert adaptation	Baseline feature set	τ	γ
head_only	corrected	2.0	0.25

設定	模型	Accuracy mean ± STD	Balanced accuracy	Macro-F1
All trials	Global Shallow	71.93% ± 17.61%	71.92%	0.711
All trials	Optimized hard K=2	72.01% ± 17.38%	72.01%	0.717
All trials	Confidence-aware soft K=2	72.03% ± 17.48%	72.02%	0.713
20 Left + 20 Right	Global Shallow	59.75% ± 10.58%	59.75%	0.562
20 Left + 20 Right	Optimized hard K=2	59.95% ± 11.35%	59.95%	0.578
20 Left + 20 Right	Confidence-aware soft K=2	59.70% ± 10.58%	59.70%	0.562

判斷：優化後 routing 仍未對 Global Shallow 產生足夠穩定的優勢。

設定	比較	平均差	Bootstrap 95% CI	Holm p	改善受試者
All trials	optimized_hard_k2 - global_shallow	+0.08 pp	[-0.77, +0.94] pp	1.0000	5/10
All trials	optimized_soft_k2 - global_shallow	+0.10 pp	[-0.05, +0.28] pp	1.0000	3/10
20 Left + 20 Right	optimized_hard_k2 - global_shallow	+0.20 pp	[-1.65, +1.40] pp	0.6395	7/10
20 Left + 20 Right	optimized_soft_k2 - global_shallow	-0.05 pp	[-0.15, +0.00] pp	0.9519	0/10

設定	Silhouette median	Bootstrap ARI median	Expert disagreement	Oracle accuracy	Router hit rate	Fallback clusters
All trials	0.139	0.161	9.52%	76.74%	52.07%	0
20 Left + 20 Right	0.099	0.065	16.90%	68.25%	52.55%	70

所有 scaler、PCA、KMeans 與距離尺度只 fit outer-train baseline。現有頻帶特徵本來已取 log；新版另外比較 Fisher-z correlation 與較集中於 Mu/Beta 的特徵。低頻 slope 只視為探索性準備狀態，不宣稱為 MRCP。