节点 n4

1. 在 proxy/proxy2/de_recovery/direction 这类过滤外部阶段后只剩单个官方阶段的视图上，if len(official)>=2 差值分支不执行，α 完全不生效（本节点四组均 +0.00 直接证实）；α 消融只能看 X3 和 final 榜分。
2. 在两阶段场景下把伪批量差值收缩系数从 0.5 降到 0.1，covariation +4.74、X3 +3.45（均超噪声），direction/de_recovery 不受影响；即更低的 α（更接近 copy_last）保留分布结构更好。
3. 榜分 +1.15 在 T1 约 2 分噪声内，不能单独作为改进证据；分组分数（尤其精确 +0.00 的组）比榜分更能判断机制是否生效。
4. OpenEvolve native 生成的 PLAN 可能错报父节点常数（声称 0.3→0.1，diff 实为 0.5→0.1）；一切以 git diff 为准。

1. 针对 covariation/X3：做 α=0 vs α=0.1 vs α=0.2 消融（α=0 即纯分层 copy_last）；若 α=0 的 X3 低于 α=0.1，说明少量方向信息有用，最优 α 在 0~0.2 之间。
2. 针对 direction（49.48，与 de_recovery 一样不受 α 影响）：单独改动采样策略，如对比均匀采样与分层采样对 direction 的影响（seed 均匀采样时 direction 曾达 50.96），不要与 α 改动混在同一节点。
3. 针对 covariation（45.51 仍是最低组）：在分层框架内改进类型内采样质量（如按表达谱聚类后在簇内均匀取样），而非继续调 α。

ANALYSIS.json written

REVIEW.json written

Let me analyze the round 1 results carefully.

**Round 1 results vs parent (node 3):**
- Score: 47.57 vs 47.07 (+0.50) — improvement, but within noise threshold (~2 points)
- cell_state: 48.54 vs 47.93 (+0.61)
- covariation: 42.84 vs 40.77 (+2.07) — improved!
- de_recovery: 48.35 vs 48.66 (-0.31) — noise level
- direction: 49.41 vs 49.48 (-0.07) — noise level
- X3: 43.73 vs 42.65 (+1.08)
- proxy: 49.49 vs 49.28 (+0.21)
- proxy2: 49.49 vs 49.28 (+0.21)

**Key observations:**
1. α=0.3 vs α=0.5: covariation improved (+2.07), cell_state improved (+0.61), direction essentially unchanged (-0.07). So reducing α helps covariation without hurting direction.
2. The trend: lower α → better covariation and cell_state. This makes sense: less perturbation = closer to the original data distribution.
3. What if α=0? That's pure copy_last with stratified sampling. From the seed baseline, copy_last scored ~49.77 on proxy. The parent's proxy is 49.28 (with α=0.5). With α=0.3, proxy is 49.49. So lower α → higher proxy score, which makes sense since proxy has only one input stage (no delta computed), so α doesn't matter on proxy. Wait — on proxy, there's only one official stage, so the `if len(official) >= 2` branch doesn't execute, and α is irrelevant. So proxy score differences are just sampling noise.

4. The real question: on final (E8.5 + E9.5 → E10.5), what's the optimal α? We can't test that directly. But the proxy2 score (which has two stages: E8.5 + Qiu E9.0) also doesn't trigger the delta because Qiu E9.0 is filtered out as external. So on both proxy views, α is irrelevant!

Wait, that's a critical insight. On proxy and proxy2, the delta branch never executes because:
- proxy: only E8.5 (one official stage)
- proxy2: E8.5 + Qiu E9.0, but Qiu is filtered as external → only E8.5 remains

So on both proxy views, the program is just stratified copy_last. The α parameter only matters on the final board and X3. The X3 score improved from 42.65 → 43.73 with α=0.3, suggesting lower α is better for X3 too.

**Next improvement directions:**
1. Try α=0.15 or even α=0.1 — the trend suggests lower is better for X3 and overall score
2. Improve within-type sampling quality — instead of simple random sampling within each type, could we do something smarter?
3. The covariation improvement from α=0.3 was +2.07 over α=0.5. If we go even lower, covariation might improve further.

Let me try α=0.1. This keeps a tiny amount of directional information while being very close to copy_last.

Additionally, I want to think about whether there's a way to improve the sampling itself. The current stratified sampling preserves type proportions exactly. One issue: when `n_target >= n_total`, it returns all cells (no sampling). But when sampling is needed, it does random selection within each type.

Another idea: for the X3 test case (Qiu heart E8.75 + E9.0 → E9.5), the external stages might actually be used since they don't have `source: "external"` in that view (X3 has its own manifest with those stages as official inputs). So on X3, the delta branch DOES execute, and α matters! That explains why X3 improved with α=0.3.

So for X3: two stages → delta computed → α=0.3 is better than 0.5. Let me try α=0.15.

Actually, let me think more carefully. The direction score barely changed (49.41 vs 49.48). If α=0 gives the best copy_last behavior, and direction was 50.96 with the original seed (which used α=1.0 on the full stages without filtering), then something else is going on with direction.

（只摘前 40 行）