节点 n10 在终选来历上

1. shape_scale 对目标 RMS 敏感且方向明确：factor 1.2→54.84、1.05→60.16、1.0→62.83、0.95→66.27，越接近参考尺度 335 越好，超过 max(输入 radii) 一定变差。
2. α 在 0.02–0.05 区间的伪批量位移对 expression_change/cell_state 的影响都在噪声内（≤0.4 分）；机制关闭（α=0，factor 1.0）曾得 53.21，不低于父节点 52.98，说明该位移机制本身没有产生超出噪声的收益，本节点的 +0.96 总分几乎全部来自尺度调整（全局重缩放），不是来自声称的表达匹配位移机制。
3. 榜分 +0.96 处于 T2 噪声（约 1 分）边缘，只有 shape_scale 分组 +3.44 是可信信号；评价 tune 节点时应看分组增量而不是只看总分。

1. 针对 shape_scale：继续沿 factor 方向细扫（0.90、0.925、0.975），RMS≈318–345，找 66.27 之上的峰值；factor 是最强的单一杠杆。
2. 针对 expression_change/cell_state：既然 α∈[0.02,0.05] 与 α=0 均在噪声内，直接固定 α=0（去掉位移机制）以简化方法，把 tune 预算全部用于尺度和形状参数；或改用非全局的形变（如按基因表达梯度做局部位移）再验证是否有分组收益。
3. 若继续保留位移机制，先做机制开关对照（α=0 vs α=0.02，同 factor），确认 expression_change 是否有 >1 分的真实差异，再决定是否值得调 α。

ANALYSIS.json written

Let me analyze the results:

- Parent baseline: 52.928 (shape_scale=62.83, with target_rms = max(radii) = 354, α=0.05)
- Round 0: α=0.02, factor=1.2 → 51.077 (shape_scale=54.84, overshot)
- Round 1: α=0.02, factor=1.05 → 52.408 (shape_scale=60.16, still below parent's 62.83)

The shape_scale scores tell the story:
- factor=1.0 (parent, α=0.05): 62.83
- factor=1.05 (α=0.02): 60.16
- factor=1.2 (α=0.02): 54.84

So increasing RMS beyond max(radii) *hurts* shape_scale. The optimal is at or below factor=1.0. The parent's factor=1.0 gives the best shape_scale observed (62.83).

The α=0.02 helped expression_change (50.00 vs 49.67) and cell_state (49.41 vs 49.06) compared to α=0.05, but the shape_scale loss from factor>1.0 dominates.

Best strategy for round 2: keep factor=1.0 (parent's scale rule) but reduce α to 0.02 to recover the expression/cell_state gains without the shape penalty. This should give approximately: expression_change ≈ 50.0, cell_state ≈ 49.41, shape_scale ≈ 62.83, local_spatial ≈ 50.06. That's roughly 53.0+ which would beat the parent's 52.928, but likely not by more than 1 point.

Alternatively, try α=0 (mechanism off) with factor=1.0. From the table: "RMS = max(输入) = 354（机制关闭）" scored 53.21. That's only 0.28 above parent. Still not >1.

What about trying factor slightly below 1.0? The parent's scale_log_ratio was +0.0528 (354/335). Maybe the optimal is slightly less. Let me think: if factor=1.0 gives 62.83 and factor=1.05 gives 60.16, the shape score is decreasing as we go up. What about factor=0.95? That would give RMS ≈ 336, very close to the target's reference RMS of 335. The scale_log_ratio would be ≈ 0, which might score closer to 50 on shape_scale... no, that's worse.

Actually, the scoring likely rewards being close to the true target scale. The true E9.5 RMS is unknown but the score of 62.83 at RMS=354 vs reference 335 means we're already above the reference and still scoring well. Going higher (425) dropped to 54.84, and 372 dropped to 60.16. So the optimum is near 354 or slightly below.

Let me try factor=0.95: RMS = 354*0.95 ≈ 336, scale_log_ratio ≈ 0.003. This might score even better on shape if the true value is near 335.

Combined with α=0.02 for the expression gains:

CHANGE: α from 0.05 to 0.02 (recover expression_change and cell_state gains seen in rounds 0-1), and target RMS scale factor from 1.0 to 0.95 (round 1 showed factor>1.0 hurts shape_scale; try slightly below max to approach the reference scale of 335).

<<<<<<< SEARCH
ALPHA = 0.05
=======
ALPHA = 0.02
>>>>>>> REPLACE

<<<<<<< SEARCH
    target_rms = float(np.max(radii))
=======
    target_rms = float(np.max(radii)) * 0.95
>>>>>>> REPLACE