总览 · ← 返回运行 20261002-034201-search-t1-abc-r1-B-population

节点 n3

官方最新阶段分层按型抽样复制；外部阶段只在无官方输入时作底；伪批量平移默认关闭（X3 实测任何幅度都掉分）。

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 30dc67de4c: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：父节点版本 15cb10398e。改动的文件：solution/METHOD.md +49 −0、solution/README.md +2 −3、solution/run.py +57 −16

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..611900d--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,49 @@+官方最新阶段分层按型抽样复制；外部阶段只在无官方输入时作底；伪批量平移默认关闭（X3 实测任何幅度都掉分）。++# 方法++在父节点（pseudobulk_shift 种子）上做三处修改：++1. **基座只用官方阶段**（`inputs_by_time(manifest, include_external=False)`）。父节点在 proxy2 上把外部+   Qiu E9.0 心脏细胞（2174 个、仅心脏谱系、部分基因用 E8.5 均值补齐）当成"最新阶段"整体复制，预测群体+   塌缩成心脏细胞，这是 proxy2=27.43 的主因。外部输入只在没有官方阶段时使用（X3 视图两个输入均为外部，+   无 `source` 标记时按其自身 manifest 处理）。+2. **分层抽样**：按基座 `obs["celltype"]` 的比例配额（最大余数法）抽 `target_n_cells` 个细胞，保持类型+   组成精确不变，替代均匀行抽样。无 celltype 列时回退 `sample_rows`。+3. **平移关闭**：保留按时间间隔比缩放（(t_target−t_last)/(t_last−t_prev)，夹到 [0,3]）再乘 ALPHA 的+   每类型伪批量差值机制，但 ALPHA 默认 0（环境变量 `VECSHIFT_ALPHA` 可覆盖，代码内常数，与 seed 无关，+   输出仍确定）。++## ALPHA=0 的依据（X3 实测，vec-score）++| 有效平移幅度 | X3 分 | covariation |+|---|---|---|+| 0（分层复制） | 50.0 | 50.0 |+| 0.5 | 42.4 | 21.9 |+| 1.0（父节点） | 40.5 | – |+| 2.0 | 37.6 | 9.7 |++单调递减：对数空间 clip(x+delta, 0) 造成的人为零峰破坏 variogram/协变结构，且 direction、de_recovery+也未因平移提高。官方方法卡同样报告 T1 上常数位移 48.6 < copy_last。故 final 视图（E8.5+E9.5→E10.5）+也退化为分层 copy_last(E9.5)。++## 查分记录（A 半）++- proxy seed0：50.35（covariation 51.9，父 50.04 / cov 22.6）；seed1：50.69+- proxy2 seed0：50.35（父 27.43）+- X3 seed0：50.0（父 40.53）；alpha 扫描见上表+- 预计节点分 ≈ (50.35+50.35+50.0)/3 ≈ 50.2++## 验证过 / 没验证++- 验证：三个视图 vec-check 通过；同 seed 两次运行输出 md5 相同；X3 上 ALPHA∈{0,0.25,0.5,1,2} 扫描。+- 没验证：proxy/proxy2 只有一个官方阶段，平移分支在这两个视图上从不触发，ALPHA 的选择只由 X3 和官方+  卡的 T1 常数位移结果支持；没有验证按基因子集（如仅 top-DE 基因）或免 clip 的平移是否可能超过 50。+- 生物学先验：仅使用了"外部输入不是全胚、不能直接当预测输出"这一 CONTRACT 说明；未使用保留阶段/基因型+  的任何信息，未读 `uns.celltype_palette`。++## 下一步建议++- 在 proxy 上寻找能真正超过 copy_last 的 E8.5→E9.5 变换（当前 de_recovery≈50 表示与基线持平）；+  例如只平移高置信 DE 基因、保稀疏结构的加性修正、或用 prior/（Reactome、TF 调控）约束平移方向。+- 若引入平移，需避免 clip 零峰：可在计数空间做乘法缩放再 log1p，保持稀疏与方差结构。diff --git a/solution/README.md b/solution/README.mdindex ba29577..2fdff65 100644--- a/solution/README.md+++ b/solution/README.md@@ -1,4 +1,3 @@-# pseudobulk_shift+# copy_last_official_stratified -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。+官方最新输入阶段按细胞类型分层抽样复制；外部输入阶段只在无官方阶段（X3）时作底；伪批量平移机制保留但 ALPHA 默认 0（X3 实测任何幅度均掉分，见 METHOD.md）。diff --git a/solution/run.py b/solution/run.pyindex f3a0f25..d43402b 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,13 +1,21 @@ #!/usr/bin/env python3-"""pseudobulk_shift: latest stage + per-cell-type pseudobulk delta of the last step.+"""copy_last(official) + stratified per-type sampling; optional time-scaled shift (ALPHA=0 default). -The delta is mean(last|type) - mean(prev|type) over the two latest inputs,-computed on the full stages and added once to a subsample of the latest stage-(clipped at 0). Types missing from the earlier stage are copied unchanged.--With a single input stage (T1 proxy: E8.5 only) there is no step to take a-delta from, so this falls back to copy_last with the same sampling. The proxy-therefore cannot tell this seed from copy_last; that gap is expected.+Changes vs parent (pseudobulk_shift seed):+1. Base stage = latest OFFICIAL input (include_external=False). The parent+   copied the external Qiu E9.0 heart-only cells on proxy2, collapsing the+   predicted population to heart lineages (proxy2 = 27.43). External inputs+   are only used as a delta source when there are no official stages (X3).+2. Stratified per-cell-type sampling instead of uniform row sampling: keeps+   the type composition of the base stage exactly (proportional quotas),+   preserving between-type covariation under subsampling.+3. Optional per-type pseudobulk shift when two base stages exist, scaled by+   (target gap / observed step gap) and shrunk by ALPHA. Measured on X3+   (two Qiu heart stages E8.75->E9.0, target E9.5): effective shift 0 ->+   50.0, 0.5 -> 42.4, 1.0 -> 40.5 (parent), 2.0 -> 37.6; the official card+   reports the same on T1 (constant shift 48.6 < copy_last). The shift+   monotonically hurts (clip-induced zero spikes destroy variogram /+   covariation), so ALPHA defaults to 0.0 (env VECSHIFT_ALPHA overrides). """  from __future__ import annotations@@ -28,6 +36,27 @@ from src.task1_temporal.view_io import (     write_prediction, ) +ALPHA = float(__import__("os").environ.get("VECSHIFT_ALPHA", "0.0"))+MAX_SCALE = 3.0+++def stratified_rows(labels: np.ndarray, n: int, rng: np.random.Generator) -> np.ndarray:+    """Row indices with per-type quotas proportional to type counts (largest remainder)."""+    types, counts = np.unique(labels, return_counts=True)+    quota = n * counts / counts.sum()+    take = np.floor(quota).astype(np.int64)+    rem = n - take.sum()+    if rem > 0:+        order = np.argsort(-(quota - take))+        take[order[:rem]] += 1+    rows = []+    for t, k in zip(types.tolist(), take.tolist()):+        if k <= 0:+            continue+        idx = np.flatnonzero(labels == t)+        rows.append(rng.choice(idx, size=k, replace=k > idx.size))+    return np.sort(np.concatenate(rows))+  def main() -> None:     parser = argparse.ArgumentParser()@@ -38,16 +67,28 @@ def main() -> None:      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)-    stages = inputs_by_time(manifest)-    last = read_stage(args.data, stages[-1], genes)+    stages = inputs_by_time(manifest, include_external=False)+    if not stages:+        stages = inputs_by_time(manifest, include_external=True)+    base_entry = stages[-1]+    base = read_stage(args.data, base_entry, genes)     rng = np.random.default_rng(args.seed)-    rows = sample_rows(last.n_obs, target_n_cells(manifest, last.n_obs), rng)-    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_of(last))+    n = target_n_cells(manifest, base.n_obs)+    has_types = "celltype" in base.obs.columns+    rows = stratified_rows(labels_of(base), n, rng) if has_types else sample_rows(base.n_obs, n, rng)+    X = base.X[rows]++    if len(stages) >= 2 and ALPHA > 0:+        prev_entry = stages[-2]+        dt_step = float(base_entry["time"]) - float(prev_entry["time"])+        dt_out = float(manifest["target"]["time"]) - float(base_entry["time"])+        prev = read_stage(args.data, prev_entry, genes)+        if has_types and "celltype" in prev.obs.columns and dt_step > 0:+            scale = float(np.clip(dt_out / dt_step, 0.0, MAX_SCALE)) * ALPHA+            deltas = type_deltas(prev.X, labels_of(prev), base.X, labels_of(base))+            deltas = {t: (scale * d).astype(np.float32) for t, d in deltas.items()}+            X = shift_rows(X, labels_of(base)[rows], deltas)         del prev-        X = shift_rows(X, labels_of(last)[rows], deltas)     write_prediction(X, genes, args.out, seed=args.seed)  

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）。

用到的知识库条目

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

ANALYSIS.json written

工程师

`agent.search.run_candidate` 模块在本节点环境不可用；直接对三个真实视图目录跑 `run.py` 均已通过 vec-check，等价有效。

**总结**

- 做了什么：父节点在 proxy2 上复制了外部 Qiu E9.0 心脏细胞（群体塌缩 → 27.43）。改为：基座只取最新官方阶段；按 celltype 分层比例抽样；保留时间缩放的伪批量平移机制但默认 ALPHA=0（X3 实测有效幅度 0/0.5/1/2 → 50.0/42.4/40.5/37.6，单调掉分，clip 零峰破坏协变）。
- 查分（7 次额度）：proxy 50.35（seed1 50.69）、proxy2 50.35、X3 50.0，预计节点分 ≈50.2（父 39.34）。三视图 vec-check 通过、同 seed 输出 md5 一致。
- 下一步最值得试：能真正超过 copy_last 的 E8.5→E9.5 变换——只平移高置信 DE 基因、免 clip（计数空间乘法缩放再 log1p）保稀疏方差结构，或用 prior/（TF 调控、Reactome）约束平移方向；covariation 组是最敏感的信号。

调研员

use soft floor**: instead of clip(x+delta, 0), set floor = min(input expression) - 1.0 (in log space). Apply x_new = maximum(x + alpha*delta_c, floor). This avoids the zero-spike artifact while preventing unphysical negatives. Alternatively, simply remove clipping entirely if expression is already in log1p space (values can be negative).\n3. **Shrinkage alpha**: multiply delta by alpha, initial value 0.5, search {0.3, 0.5, 0.7} via vec-score. Rationale from k018: full alpha=1 overshoots; the one-input proxy cannot estimate alpha so we use a conservative prior and validate on proxy2/X3.\n\nSingle-input fallback (proxy): no delta exists, method reduces to stratified copy_last (same as parent but with stratified sampling ensuring exact type proportions). This should match or slightly exceed parent's proxy score.\n\nproxy2/final (two inputs): compute type_deltas as before, apply x + alpha*delta with soft floor. For types absent in prev stage, copy unchanged (as parent does).\n\nQuick validation order: (a) run on proxy to confirm no regression vs 50.04; (b) run on proxy2 to check if removing clip + alpha<1 recovers toward 40+; (c) run on X3. Total 3 vec-score queries for screening, leaving budget for alpha tuning (3 values × 3 rulers = 9 more queries max).",
  "expected_groups": ["covariation", "cell_state"],
  "risks": "1. Stratified sampling may not help if target type proportions differ substantially from input (composition shift is the real driver); Engineer should check if type fractions change between inputs on proxy2. 2. Removing clip may produce negative expression values that the scorer rejects; Engineer should verify write_prediction handles negatives gracefully, and fall back to clip-at-small-negative (-0.5) if needed. 3. Alpha=0.5 may still overshoot for noisy types with few cells; Engineer should skip delta for types with <10 cells in prev stage. 4. Improvement may be <2-point noise; confirm by running 2 seeds on proxy2 and requiring consistent direction."
}

审查员

REVIEW.json written