总览 · ← 返回运行 20261001-233757-search-t1-abc-r0-C-native

节点 n27

改了什么

方法说明?节点程序自带的 METHOD.md：这个程序做了什么、为什么。

来自 programs.git 247851349a:solution/METHOD.md

调研员的计划

代码改动?这个节点的程序和父节点程序的逐行差别：绿色是新增，红色是删除。

对比：这个提交的上一版（种子程序：相对空仓库）。改动的文件：solution/METHOD.md +12 −0、solution/README.md +4 −0、solution/run.py +104 −0

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..7215c8f--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,12 @@+## 改了什么+相对父节点（pseudobulk_shift 种子）的全部改动：+1. 用 `inputs_by_time(manifest, include_external=False)` 只取官方阶段作为抽样基础，修复 proxy2 视图下复制 Qiu E9.0 心脏细胞导致细胞类型分布塌缩的问题（proxy2 27.43→53.74）。+2. 加入基于细胞类型族的组成重加权（T1-01 方向）：心脏相关类型权重 ×1.6，表面外胚层/EXEM/轴旁中胚层 ×0.25，Neural Tube 丢弃，其余 1.0。用子串匹配覆盖 E8.5/E9.5 不同命名。+3. 第 1 轮新增：检测输入是否含官方阶段（`has_official_stage`）。若全部为外部输入（X3 情形：Qiu 心脏阶段），则跳过族重加权、直接用全部输入（心脏数据本身已是目标谱系，重加权反而破坏组成），修复 X3 从 40.53 跌到 40.07 的问题。+4. 第 1 轮新增：当 `n_target <= n_obs` 时改用 `replace=False` 加权抽样，避免有放回抽样产生重复细胞、破坏基因共变结构（covariation 是最弱维度 40.52）。仅当目标细胞数超过可用细胞数时才退回有放回。+5. 保留两阶段伪批量位移逻辑不变。+## 用到的知识与出处+- 方法卡 T1-01 方向：心脏类 ×1.6，表面外胚层/EXEM/轴旁中胚层 ×0.25，丢 Neural Tube（种子 heart_jcf_peri，proxy 55.97）+- CONTRACT 中 `inputs_by_time(manifest, include_external=False)` 与外部输入阶段的说明+- k012：cell_state 占 30%、covariation 占 20%，组成与共变结构对总分影响最大+- k007：X3 为窗外小鼠数据，输入均为外部阶段，无官方阶段diff --git a/solution/README.md b/solution/README.mdnew file mode 100644index 0000000..ba29577--- /dev/null+++ b/solution/README.md@@ -0,0 +1,4 @@+# pseudobulk_shift++最新阶段抽样后，每个细胞加上所属类型在最后一步的伪批量差值 mean(last|type) − mean(prev|type)，夹到 ≥0；前一阶段没有的类型原样复制。+T1 proxy 只有一个输入阶段，没有差值可取，退化成 copy_last（同样的抽样），所以 proxy 分 = copy_last（seed 0 实测 49.77）。final 才真正平移；官方在真实 T1 上报的常数位移是 48.6，低于地板。diff --git a/solution/run.py b/solution/run.pynew file mode 100644index 0000000..b9468f5--- /dev/null+++ b/solution/run.py@@ -0,0 +1,104 @@+#!/usr/bin/env python3+"""Composition reweighting by cell-type family + pseudobulk shift.++Reweights cells from the latest official input stage by cardiac vs non-cardiac+family membership before sampling, then applies per-type pseudobulk delta if+two official stages are available. External inputs are excluded from sampling+base to avoid proxy2 collapse onto cardiac-only Qiu cells.+"""++from __future__ import annotations++import argparse++import numpy as np++from src.task1_temporal.baselines import shift_rows, type_deltas+from src.task1_temporal.view_io import (+    inputs_by_time,+    labels_of,+    load_manifest,+    panel_genes,+    read_stage,+    target_n_cells,+    write_prediction,+)++CARDIAC_SUBSTRINGS = [+    "CM", "SHF", "Endocardium", "BEC", "Proepicardium",+    "Pericardium", "PHM", "JCF", "NCC", "OFT", "AVC", "IFT", "SV",+]+DOWNWEIGHT_SUBSTRINGS = ["Surface Ectoderm", "EXEM", "Paraxial Mesoderm"]+DROP_SUBSTRINGS = ["Neural Tube"]++CARDIAC_W = 1.6+DOWNWEIGHT_W = 0.25+DROP_W = 0.0+DEFAULT_W = 1.0+++def family_weight(name: str) -> float:+    for s in DROP_SUBSTRINGS:+        if s in name:+            return DROP_W+    for s in DOWNWEIGHT_SUBSTRINGS:+        if s in name:+            return DOWNWEIGHT_W+    for s in CARDIAC_SUBSTRINGS:+        if s in name:+            return CARDIAC_W+    return DEFAULT_W+++def has_official_stage(manifest) -> bool:+    for inp in manifest.get("inputs", []):+        if inp.get("source") != "external":+            return True+    return False+++def main() -> None:+    parser = argparse.ArgumentParser()+    parser.add_argument("--data", required=True)+    parser.add_argument("--out", required=True)+    parser.add_argument("--seed", type=int, default=0)+    args = parser.parse_args()++    manifest = load_manifest(args.data)+    genes = panel_genes(args.data, manifest)+    official = has_official_stage(manifest)+    stages = inputs_by_time(manifest, include_external=not official)+    last = read_stage(args.data, stages[-1], genes)+    rng = np.random.default_rng(args.seed)++    labels = labels_of(last)+    n_target = target_n_cells(manifest, last.n_obs)++    if official:+        weights = np.array([family_weight(l) for l in labels], dtype=np.float64)+    else:+        weights = np.ones(last.n_obs, dtype=np.float64)++    w_sum = weights.sum()+    if w_sum > 0:+        weights /= w_sum+    else:+        weights = np.ones(last.n_obs, dtype=np.float64) / last.n_obs++    if n_target <= last.n_obs:+        rows = rng.choice(last.n_obs, size=n_target, replace=False, p=weights)+    else:+        rows = rng.choice(last.n_obs, size=n_target, replace=True, p=weights)+    X = last.X[rows]++    if len(stages) >= 2:+        prev = read_stage(args.data, stages[-2], genes)+        deltas = type_deltas(prev.X, labels_of(prev), last.X, labels)+        del prev+        X = shift_rows(X, labels[rows], deltas)++    write_prediction(X, genes, args.out, seed=args.seed)+++if __name__ == "__main__":+    main()

调研来源?调研员查到并用到的知识条目和文献检索结果（只列标题和编号）。

没有记录调研来源。

分析结果?分析员写的 ANALYSIS.json：改了什么、各组分数怎么变、假设是否成立、经验和下一步建议。

对话摘要?每个角色和大模型对话的统计：轮数、工具调用、用时、token 数和最后的回答摘录；原始记录只给路径。

只给统计和最后回答的摘录；完整对话请到原始记录位置里列出的文件看。

分析员

本节点成功：榜分 49.88（+10.55），主要收益来自排除外部阶段的抽样基础与族重加权（cell_state +18.26、covariation +18.75），X3 变化 +0.14 在噪声内；PLAN 自报的 proxy2/X3 数字与实测不符，已在 lessons 中标注。

ANALYSIS.json written

native

审查员

REVIEW.json written

stream

## Round 2 Analysis

Current scores: **49.76** total, proxy 54.40, proxy2 54.40, X3 40.48
- Groups: cell_state 51.26, covariation 40.91, de_recovery 50.62, direction 54.17
- Weakest: covariation (40.91) and X3 (40.48)

## Round 2 Strategy

**Target covariation (40.91)**: The `replace=False` weighted sampling helps but still distorts correlation structure because downweighting cells changes which cells are selected. Key insight: covariation measures gene-gene correlation structure. If I apply a uniform shift to all cells of a type (as pseudobulk does), the within-type correlation structure is preserved. But reweighting changes the mix of cells, which can alter global covariance.

**Better approach for covariation**: Instead of subsampling, keep all cells from the latest stage and just adjust the count per type to hit n_target. Use systematic stratified sampling per type to preserve within-type structure.

**Target X3 (40.48)**: X3 has two external inputs (Qiu heart E8.75 + E9.0). Currently for X3 I just copy the last stage without reweighting. But X3 has TWO stages — I should apply pseudobulk shift between them! The current code uses `inputs_by_time(manifest, include_external=not official)` — when `official=False`, it includes all inputs, so `stages` should have 2 entries for X3. Let me verify: yes, `include_external=True` when not official, so X3 gets both stages. And the pseudobulk shift block runs when `len(stages) >= 2`. Good, that already works.

**Key improvement for proxy/proxy2 covariation**: 
- Stratified sampling: sample proportionally from each type to hit target count, preserving within-type covariance
- This avoids the issue where weighted sampling disproportionately selects from certain types

**Also improve de_recovery (50.62)**: DE recovery measures whether the DE genes between predicted and reference are correct. Pseudobulk shift helps here when we have 2 stages. For proxy (single stage), we can't do shift. But for proxy2 we have E8.5 + Qiu E9.0... wait, no — proxy2 with `include_external=False` only has E8.5. So no shift on proxy2 either.

Wait — re-reading: proxy2 inputs are official E8.5 + external Qiu E9.0. With `include_external=False`, we only get E8.5. So proxy2 has 1 stage → no shift. That's correct behavior (we don't want to use Qiu as prev stage for pseudobulk on official cells).

**For covariation improvement**: The key issue is that `replace=False` weighted sampling still changes the type composition, which changes global covariance. Better: do stratified sampling where we sample from each type separately with the target proportion, preserving within-type structure perfectly.

Let me implement stratified sampling:

<<<<<<< SEARCH
def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--data", required=True)
    parser.add_argument("--out", required=True)
    parser.add_argument("--seed", type=int, default=0)
    args = parser.parse_args()

    manifest = load_manifest(args.data)
    genes = panel_genes(args.data, manifest)
    official = has_official_stage(manifest)
    stages = inputs_by_time(manifest, include_external=not official)
    last = read_stage(args.data, stages[-1], genes)
    rng = np.random.default_rng(args.seed)