节点 n34

1. 在官方输入分支强制每类型精确配额的分层抽样, 配额超过可用细胞数时需要有放回补齐, 重复细胞破坏类内共变结构: covariation 41.31→40.18, proxy 54.49→53.516; 回退加权无放回抽样后数字精确恢复到节点27水平(proxy 54.4924, covariation 41.3133), 证明退化确由该改动引入
2. manifest.get('max_cells', n_target) 在键缺失时返回 n_target 使 clamp 变空操作; 必须先判 None 再 int() 转换, 并用 max(n_target,1) 兜底, 才能保证输出细胞数落在 [min_cells, max_cells] 内
3. 同一套抽样逻辑下 proxy 与 proxy2 完全相同(54.4924), 说明两把尺子在本任务上无区分度, 优化时以 proxy 为准即可
4. 回退后耗时 1.2s→3.0s(加权抽样需构建全量概率向量), 内存峰值不变 1.26GB, 时间成本可忽略

1. 针对 covariation(最弱组 41.31, 占分20%): 在保持无放回加权抽样的前提下微调 family_weight 的衰减系数(如表面外胚层/EXEM/轴旁从 ×0.25 试 ×0.4, 心脏从 ×1.6 试 ×1.4), 观察类型比例偏移与类内共变的权衡
2. 针对 X3(40.67, 非官方分支): 该分支目前是纯均匀抽样, 可尝试与官方分支相同的 family_weight 加权无放回抽样, 验证 X3 是否同样受益于类型比例偏移
3. 对比实验必须保留父节点各组的参照分数(本节点变化量表 ref 全为 null), 否则只能靠 Engineer 转述, 建议后续节点在 DELTA 中填充 ref 值

ANALYSIS.json written

REVIEW.json written

Looking at the results across rounds:
- Node 27 (original weighted sampling): proxy 54.49, X3 40.67, score 49.88
- Round 0 (stratified): proxy 53.516, X3 40.20, score 49.08
- Round 1 (weighted + clamp fix): proxy 54.397, X3 40.48, score 49.76

The X3 score is the big weak point (40.48 vs node 32's 50.00). X3 is the external test question — no official stages, only external data (Qiu heart E8.75 + E9.0). Node 32 got X3=50, suggesting it handles X3 fundamentally differently.

For X3: `official` is False (no official stages), `inputs_by_time(manifest, include_external=True)` returns all inputs. Currently we just do uniform sampling. Node 32 probably applies some kind of temporal extrapolation using the two external inputs.

Also, covariation is still weak (40.91). The covariation metric depends on gene-gene correlation structure. One approach: instead of subsampling from the latest stage, we could use more cells (closer to max_cells) to better preserve the empirical covariance structure. Currently `target_n_cells` determines the count — perhaps we should bias toward the higher end of the allowed range.

Let me think about what to change:

1. **X3 improvement**: When there are 2+ stages (even external), apply pseudobulk shift. The current code does apply shift when `len(stages) >= 2`, which should work for X3 too if there are 2 inputs. But `stages = inputs_by_time(manifest, include_external=not official)` — when `official=False`, this becomes `include_external=True`, so both Qiu E8.75 and E9.0 should be returned. Then `len(stages) >= 2` triggers the shift. But wait — the labels from Qiu data use Qiu's celltype names ("First heart field", etc.), and the shift uses type_deltas which matches by type name between prev and last. Since both are from the same Qiu dataset, types should match. So the shift should already be happening for X3.

   The issue might be that for X3, we're just doing uniform sampling (no weighting). Maybe we should apply the same kind of reweighting for X3 too? Or maybe the issue is that X3's target is E9.5 and we need to extrapolate from E8.75→E9.0 to E9.5 (extrapolating 0.5 days beyond the last input). Currently we only do one step of shift.

   Actually, for X3: E8.75 → E9.0 is the last step, target is E9.5. So we should extrapolate the delta by a factor of (9.5-9.0)/(9.0-8.75) = 0.5/0.25 = 2.0. Currently the code just adds one delta (equivalent to extrapolating to E9.25). Let me add time-scaled extrapolation!

2. **Covariation**: The covariation score measures gene-gene correlations. Subsampling preserves this well when done without replacement. The issue might be that with n_target < n_obs, we're losing some structure. Or the shift operation might distort correlations. Let me try adding a small amount of noise reduction — actually, let me focus on the X3 extrapolation first as it's the biggest gap.

3. **Proxy improvement**: proxy is 54.397 vs node 27's 54.49. Small difference (within noise ~2). The weights might need tuning but noise makes this hard to optimize.

Key change for this round: **time-scaled extrapolation**. Instead of just adding `delta` (one step), scale it by `dt_target / dt_step` where `dt_target = target_time - last_time` and `dt_step = last_time - prev_time`. For proxy (E8.5→E9.5, target E9.5): scale = (9.5-9.5)/(9.5-8.5) = 0... wait no. For proxy: inputs are [E8.5], target is E9.5. Only one input, so no shift. For final: inputs are [E8.5, E9.5], target E10.5. scale = (10.5-9.5)/(9.5-8.5) = 1.0. For X3: inputs [E8.75, E9.0], target E9.5. scale = (9.5-9.0)/(9.0-8.75) = 2.0.

This is a meaningful improvement for X3! Currently the code applies the full delta without scaling, which corresponds to extrapolating one step (to E9.25). With scaling by 2.0, we extrapolate properly to E9.5.

But wait — the method card says pseudobulk_shift scored 48.6 on T1, below copy_last. And "再加一次 E8.5→E9.5 差值要乘收缩系数 α∈[0,1]". So large extrapolation factors are risky. For X3 with scale=2.0, maybe I should apply some shrinkage: use `alpha * scale` where alpha < 1. Let me try scale with a cap or shrinkage factor.

Actually, let me re-read: the shift is `delta = mean(last|type) - mean(prev|type)`, and we add it once. For X3 with dt_step=0.25 and dt_to_target=0.5, the correct linear extrapolation would multiply delta by 2.0. But this is risky. Let me use a compromise: multiply by min(scale, 1.5) or apply shrinkage.

Let me try: `extrapolation_factor = dt_to_target / dt_step`, capped at some maximum (say 2.0), and apply it to the delta. For proxy (single input), no shift. For final (E8.5, E9.5 → E10.5), factor = 1.0 (same as before). For X3 (E8.75, E9.0 → E9.5), factor = 2.0.

I'll cap at 2.0 and also apply a global shrinkage of 0.8 to be safe (based on the method card warning about over-extrapolation).

Actually, let me be more careful. The current code for proxy with one input does copy_last (no shift). That's already good (proxy=54.397 ≈ copy_last level). The main improvement opportunity is X3.

Let me implement time-scaled extrapolation:

```python

（只摘前 40 行）