节点 n2

1. 当任务含外部（异技术/异标签）输入阶段时，在读入阶段就 include_external=False 过滤，而不是让它参与 delta 计算或 copy_last：本节点 proxy2 +22.61、cell_state +16.42、covariation +18.17，因为外部阶段的错配标签会同时污染类型匹配和输出分布。
2. 外部阶段数据不能直接作为预测输出（CONTRACT 明令禁止），且其细胞类型标签与官方阶段不一致，用它算 per-type delta 会产生错配，纯过滤是安全的下限做法。
3. 单官方输入阶段时方案退化为 copy_last，proxy 组无法区分改进（本节点 proxy +0.00），验证 delta/收缩类改动必须看双阶段视图（proxy2/final）。
4. Engineer/PLAN 的叙述可能与实际 diff 不符（本节点声称移除噪声，但 diff 中根本没有噪声代码），复盘必须以 git diff 和变化量表为准。

1. 针对 final/X3 组：在双官方阶段路径上对 ALPHA 做网格搜索（0.3/0.5/0.7/1.0），delta 收缩强度是本方案唯一未调的超参。
2. 针对 covariation（当前 40.74，仍是最低组）：在 shift_rows 后按 per-type 残差重标定，使输出各类型的方差与 last 阶段一致，而不是只平移均值。
3. 针对 proxy2：尝试用外部 Qiu E9.0 做辅助信息（如仅用于估计心脏相关类型的时间趋势权重），但不直接输出其细胞；预期收益不确定，先在 proxy2 上小规模验证再决定是否保留。

ANALYSIS.json written

REVIEW.json written

I'll analyze the round 1 results carefully.

**Round 1 results:**
- Score: 43.70 → 47.75 (good improvement!)
- proxy: 45.01 → 50.40 (recovered, slightly better than parent's 50.04)
- proxy2: 45.01 → 50.40 (big improvement - now matches proxy!)
- X3: 41.06 → 42.43 (improved)
- cell_state: 48.99 → 49.38 (improved)
- covariation: 21.15 → 41.48 (massive improvement from removing noise!)
- de_recovery: 49.29 → 48.97 (slight decrease)
- direction: 49.81 → 49.58 (slight decrease)

**Analysis of remaining weaknesses:**

1. **de_recovery (48.97)** and **direction (49.58)** are the weakest groups now.
2. covariation jumped from 21.15 to 41.48 by removing noise - confirms that preserving exact cell values is crucial.
3. proxy and proxy2 both score 50.40 - they're identical because both have only 1 official input stage (E8.5), so both do copy_last.

**What can we improve?**

The current program on proxy/proxy2 is just copy_last (sample from E8.5). The proxy target is E9.5. So we're predicting E9.5 by copying E8.5 cells. This is the baseline.

**Key insight for improvement:** On proxy/proxy2, we have E8.5 as input and need to predict E9.5. The method card says the official pseudobulk_shift scored 48.6 on T1 (below copy_last). But that's for the final view (E9.5→E10.5). For proxy (E8.5→E9.5), copy_last is what we're doing.

Can we do better than copy_last on proxy? The method card mentions:
- Composition reweighting (T1-01) was the "run2 winner"
- The T1 card describes how cell type proportions change between stages

**Idea: Composition reweighting.** If we know that certain cell types grow faster than others between E8.5 and E9.5, we can reweight the sampling to match the expected E9.5 composition. But we can't use E9.5 data (it's the target). However, we could use general biological knowledge about proliferation rates.

Wait - but the rules say we can't hardcode statistics. We need to compute from inputs.

**Better idea:** Use the external Qiu E9.0 data in proxy2 to inform composition. Qiu E9.0 has heart cells at a time between E8.5 and E9.5. We could use the relative proportions of heart subtypes in Qiu to adjust the sampling weights of corresponding official cell types.

Actually, let me think differently. The **direction** metric likely measures whether the predicted expression moves in the right direction from input to target. With copy_last, direction is just "no movement" - which apparently scores ~50.

**Alternative approach: Use gene-level trends from the input data itself.**

Within E8.5, we can compute pseudotime or RNA velocity-like signals. Cells that are more "advanced" along a differentiation trajectory might give us a hint about which direction expression should move. But this is complex.

（只摘前 40 行）