总览 · ← 返回运行 20261002-034201-search-t1-abc-r1-A-era

节点 n2

以最新官方输入阶段为基底的 copy_last：外部输入阶段永不作基底；伪批量平移在所有三把尺子上实测有害，故 ALPHA=0 不平移。

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 b764d7f1fb: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 f87cd7b03b。改动的文件：solution/METHOD.md +28 −0、solution/README.md +4 −3、solution/run.py +31 −13

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..a1ed3d4--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,28 @@+以最新官方输入阶段为基底的 copy_last：外部输入阶段永不作基底；伪批量平移在所有三把尺子上实测有害，故 ALPHA=0 不平移。++# 方法++- 基底选择：`inputs_by_time(manifest, include_external=False)` 取最新**官方**阶段；视图无官方阶段时（X3 类外部测试题）退回全部阶段。父节点在 proxy2 上把外部 Qiu E9.0（仅心脏谱系、2174 细胞、基因不全）当基底复制，得 27.4；改成官方 E8.5 基底后 50.4。+- 抽样：`sample_rows` 无放回随机抽到 `target_n_cells`（保持输入组成，期望意义下分层），`np.random.default_rng(seed)`，确定性。+- 平移：保留 `ALPHA`/`S_MAX` 机制（factor = ALPHA × clip(Δt_target/Δt_step, 0, 2)，按细胞类型伪批量差值，clip≥0），但 **ALPHA=0**，即所有视图纯 copy_last。++# 实测（vec-score，A 半）++| 视图 | 父节点 | 本节点 |+|---|---|---|+| proxy | 50.04 | 50.40 |+| proxy2 | 27.43 | 50.40 |+| X3 | 40.53 | 50.00（factor 0）；42.43（0.5）、40.33（1.0，父行为）、38.86（1.5）、37.65（2.0） |++X3 上平移随 factor 单调变差（covariation 50→9.7，cell_state 50→39.7），与官方 T1 报告「常数位移 48.6 < copy_last」一致，故不启用平移。proxy/proxy2 的多数组已贴地板 50（de_recovery=50.0 即 de_score=0），copy_last 无 DE 信号可恢复。++# 验证过 / 未验证++- 验证：三个视图跑通 + vec-check ok；最终代码输出与已打分文件逐元素一致（maxdiff=0）；seed 确定性（单一 rng）。+- 未验证：任何需要目标阶段真值的假设；final 视图（E8.5+E9.5→E10.5）上 ALPHA=0 即 copy E9.5，与卡的 v1 起点一致但未按卡做「同名型加一次收缩差值」——卡与 X3 实测都提示平移在此评分体系下为负收益，若要试应从小 ALPHA（≤0.3）+ 只平移同名型开始。+- 未使用外部训练数据、prior/、任何保留阶段信息。生物学知识：无（纯统计复制）。++# 下一步建议++1. proxy2 用 Qiu E9.0 只做**心脏谱系类型**的小步平移（标签映射 FHF/SHF/Endocardial → aSHF/pSHF/Endothelium），其余类型复制；先验上比全局平移安全。+2. 组成/新类型层面（Hepatocyte、Blood 扩张等）无法从 proxy 校准，需按方法卡 v1 的规则式生成，在 final 上才见效。diff --git a/solution/README.md b/solution/README.mdindex ba29577..402bd24 100644--- a/solution/README.md+++ b/solution/README.md@@ -1,4 +1,5 @@-# pseudobulk_shift+# copy_last（官方基底，无平移） -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。+最新**官方**输入阶段无放回抽样复制；外部输入阶段（proxy2 的 Qiu E9.0）永不作基底。+伪批量平移机制保留但 ALPHA=0：X3 实测平移随系数单调变差（50.0 → 42.4/40.3/37.6），官方 T1 也报常数位移低于 copy_last。+A 半分：proxy 50.40 / proxy2 50.40 / X3 50.00。详见 METHOD.md。diff --git a/solution/run.py b/solution/run.pyindex f3a0f25..7b2e824 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,13 +1,18 @@ #!/usr/bin/env python3-"""pseudobulk_shift: latest stage + per-cell-type pseudobulk delta of the last step.+"""copy_last on the latest OFFICIAL input stage (stratified-free exact subsample). -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.+Rationale (measured on the three rulers of this node):+- proxy2 with the parent code copied the external Qiu E9.0 heart-only stage as+  base and scored 27.4; using the official E8.5 as base scores 50.4. External+  input stages (``source: external``) are never used as the copy base: partial+  gene panel, single lineage, different technology.+- X3 (two same-dataset stages E8.75/E9.0 -> E9.5): per-type pseudobulk shift+  hurts monotonically in the shift factor (50.0 at factor 0, 42.4 at 0.5,+  40.3 at 1.0, 37.6 at 2.0), killing covariation and cell_state. So no shift+  is applied anywhere (ALPHA = 0 below keeps the machinery for future nodes).+- proxy (single official stage) has no delta available anyway. -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.+Deterministic given --seed: one np.random.default_rng(seed), no global state. """  from __future__ import annotations@@ -28,6 +33,9 @@ from src.task1_temporal.view_io import (     write_prediction, ) +ALPHA = 0.0   # delta shrinkage; 0 = pure copy_last (measured best on all rulers)+S_MAX = 2.0   # cap on time-ratio extrapolation factor (only used if ALPHA > 0)+  def main() -> None:     parser = argparse.ArgumentParser()@@ -38,16 +46,26 @@ def main() -> None:      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)-    stages = inputs_by_time(manifest)+    stages = inputs_by_time(manifest, include_external=False)+    if not stages:+        stages = inputs_by_time(manifest, include_external=True)     last = read_stage(args.data, stages[-1], 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))-        del prev-        X = shift_rows(X, labels_of(last)[rows], deltas)++    if ALPHA > 0 and len(stages) >= 2:+        dt_step = float(stages[-1]["time"]) - float(stages[-2]["time"])+        dt_target = float(manifest["target"]["time"]) - float(stages[-1]["time"])+        s = float(np.clip(dt_target / dt_step, 0.0, S_MAX)) if dt_step > 0 else 0.0+        factor = ALPHA * s+        if factor > 0:+            prev = read_stage(args.data, stages[-2], genes)+            deltas = type_deltas(prev.X, labels_of(prev), last.X, labels_of(last))+            deltas = {t: (d * np.float32(factor)) for t, d in deltas.items()}+            X = shift_rows(X, labels_of(last)[rows], deltas)+            del prev+     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

完成。

- **做了什么**：修复父节点在 proxy2 上把外部 Qiu E9.0（心脏-only、基因不全）当复制基底的 bug，改为只用最新官方阶段作基底；X3 上扫了平移系数 {0, 0.5, 1.0, 1.5, 2.0}，实测单调有害，故 ALPHA=0 纯 copy_last（机制保留供后续节点）。
- **查分**（7/20 次）：proxy 50.40（父 50.04）、proxy2 50.40（父 27.43）、X3 50.00（父 40.53）→ 节点 A 半均分 ≈50.3 vs 父 39.34。三视图 vec-check 通过，输出与已打分文件逐元素一致，seed 确定性验证通过。
- **下一步最值得试**：proxy2 仅对心脏谱系类型（FHF/SHF/Endocardial → aSHF/pSHF/Endothelium 映射）做小步 Qiu 平移，其余复制；以及按方法卡 v1 的规则式组成/新类型生成（final 上才见效，proxy 无法校准）。

调研员

roach": "1) Replace random subsampling with stratified sampling: sample cells per type proportional to their frequency in the last stage, ensuring the output composition matches the input exactly (preserves global covariance structure). Use np.random.choice within each type stratum. 2) Remove the hard clip at 0. Instead apply a soft floor: x_new = log1p(exp(x + delta)) in log-space (softplus), which is smooth and preserves rank-order correlations. If data is already in log space, just remove the clip entirely (allow small negatives). 3) For proxy2 (two inputs available): apply shrinkage x_new = x + alpha * delta_c with alpha initial 0.5, search {0.3, 0.5, 0.7, 1.0}. Estimate best alpha by holding out 20% of Qiu E9.0 cells, applying delta from E8.5→Qiu E9.0 on the rest, and measuring correlation of pseudobulk per type. 4) For proxy (single input, no delta possible): fall back to stratified copy_last (same sampling improvement). 5) For X3: same logic as proxy/proxy2 depending on available stages. 6) Quick validation: run proxy first (fast, single stage), confirm no regression from 50.04; then proxy2 to check delta + shrinkage effect; use vec-score after each change. Total runtime should stay under 5s.",
  "expected_groups": ["covariation", "cell_state"],
  "risks": "Stratified sampling may fail if a type has very few cells (<5); Engineer should fall back to random for types with <5 cells. Removing clip might produce negative values that the scorer handles poorly; check by inspecting output range after first run. Shrinkage alpha tuned on proxy2 may not transfer to X3 (different dataset); mitigate by picking alpha conservatively (0.5 default). If covariation doesn't improve, the issue may be that the scorer measures temporal covariance change rather than static covariance; in that case the fix is to ensure the delta itself has realistic gene-gene correlation structure (use the observed covariance of the delta across cells within a type rather than a constant vector)."
}

审查员

REVIEW.json written