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

节点 n9

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 f79e69675f: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 17f3e11784。改动的文件：solution/METHOD.md +2 −2、solution/run.py +54 −5

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex e9f77d1..d750800 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,4 +1,4 @@ ## 改了什么-用 `inputs_by_time(manifest, include_external=False)` 过滤，只用官方阶段做抽样和差值计算的基底。当官方阶段为空时（如 X3 测试题全是外部输入）退回使用全部阶段。这修复了 proxy2 上的严重问题：父节点把 Qiu E9.0（心脏谱系、不同标签体系）当作最新阶段直接抽样输出，导致细胞类型和表达谱与官方 E9.5 真值严重不匹配，proxy2 只得 27.43。修复后 proxy2 退化为 copy_last（E8.5），预期接近 proxy 的 50 分水平。+相对父节点的全部改动：(1) 逐基因经验贝叶斯收缩（signal/(signal+noise_var)）替代原始 delta，乘全局系数 ALPHA_GLOBAL=0.6；(2) 移除了第 1 轮引入的 CELL_JITTER（它严重破坏 covariation，从 41.29 降至 31.97）；(3) 移除了移到 if 块外的高斯噪声（它使 proxy/proxy2 的 copy_last 退路变差，50.40→48.42）；(4) 不再添加任何额外噪声，纯 EB 收缩 + 常数位移，保留类型内协方差结构。单阶段时退化为无噪声的 copy_last。 ## 用到的知识与出处-CONTRACT 中"外部输入阶段"一节明确说明 `inputs_by_time(manifest, include_external=False)` 可获取仅官方阶段；方法卡 k018 指出单输入时退化为 copy_last 是预期行为。+方法卡 k018（逐基因经验贝叶斯收缩，噪声基因少移动）；实验表 Program 1（EB 收缩 + ALPHA=0.6 无噪声得 47.80，covariation 41.90）；第 1 轮实验证明 CELL_JITTER 和 if 块外噪声均有害（covariation 31.97，proxy 48.42）。diff --git a/solution/run.py b/solution/run.pyindex dabccbd..8235b5f 100644--- a/solution/run.py+++ b/solution/run.py@@ -13,7 +13,8 @@ import argparse  import numpy as np -from src.task1_temporal.baselines import shift_rows, type_deltas+from scipy import sparse+ from src.task1_temporal.view_io import (     inputs_by_time,     labels_of,@@ -25,6 +26,29 @@ from src.task1_temporal.view_io import (     write_prediction, ) +ALPHA_GLOBAL = 0.6+++def _to_dense(X):+    if sparse.issparse(X):+        return np.asarray(X.todense(), dtype=np.float32)+    return np.asarray(X, dtype=np.float32)+++def _per_type_stats(X, labels):+    Xd = _to_dense(X)+    types = np.unique(labels)+    means = {}+    variances = {}+    counts = {}+    for t in types:+        mask = labels == t+        sub = Xd[mask]+        means[t] = sub.mean(axis=0)+        variances[t] = sub.var(axis=0)+        counts[t] = int(mask.sum())+    return means, variances, counts+  def main() -> None:     parser = argparse.ArgumentParser()@@ -42,15 +66,40 @@ def main() -> None:      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]+    n_target = target_n_cells(manifest, last.n_obs)+    rows = sample_rows(last.n_obs, n_target, rng)+    X = _to_dense(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))+        labels_last = labels_of(last)+        labels_prev = labels_of(prev)++        means_last, vars_last, counts_last = _per_type_stats(last.X, labels_last)+        means_prev, vars_prev, counts_prev = _per_type_stats(prev.X, labels_prev)         del prev-        X = shift_rows(X, labels_of(last)[rows], deltas) +        deltas = {}+        for t in means_last:+            if t in means_prev:+                raw_delta = means_last[t] - means_prev[t]+                noise_var = (+                    vars_last[t] / max(counts_last[t], 1)+                    + vars_prev[t] / max(counts_prev[t], 1)+                )+                signal = raw_delta ** 2+                shrink = signal / (signal + noise_var + 1e-10)+                deltas[t] = (raw_delta * shrink * ALPHA_GLOBAL).astype(np.float32)+            else:+                deltas[t] = np.zeros(len(genes), dtype=np.float32)++        labels_sampled = labels_last[rows]+        for t in np.unique(labels_sampled):+            if t in deltas:+                mask = labels_sampled == t+                X[mask] += deltas[t]++    np.clip(X, 0.0, None, out=X)     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）。

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

ANALYSIS.json written

native

stream

Looking at Round 1 results, the problems are clear:
1. Noise outside the `if` block hurts proxy/proxy2 (50.40 → 48.42) because it degrades pure copy_last
2. CELL_JITTER destroys covariation (41.29 → 31.97) by adding random per-cell scaling

The fix: remove CELL_JITTER entirely, move noise back inside the `if` block, and reduce it further. Program 1 (no noise at all) achieved covariation 41.90, so minimal noise is best.

<<<<<<< SEARCH
ALPHA_GLOBAL = 0.6
NOISE_FRAC = 0.04
CELL_JITTER = 0.08
=======
ALPHA_GLOBAL = 0.6
>>>>>>> REPLACE

<<<<<<< SEARCH
        labels_sampled = labels_last[rows]
        for t in np.unique(labels_sampled):
            if t in deltas:
                mask = labels_sampled == t
                n_cells = int(mask.sum())
                jitter = 1.0 + rng.normal(0, CELL_JITTER, size=(n_cells, 1)).astype(np.float32)
                X[mask] += deltas[t] * jitter

    gene_std = np.std(_to_dense(last.X), axis=0) + 1e-8
    noise = rng.normal(0, NOISE_FRAC, size=X.shape).astype(np.float32)
    X += noise * gene_std[None, :]

    np.clip(X, 0.0, None, out=X)
=======
        labels_sampled = labels_last[rows]
        for t in np.unique(labels_sampled):
            if t in deltas:
                mask = labels_sampled == t
                X[mask] += deltas[t]

    np.clip(X, 0.0, None, out=X)
>>>>>>> REPLACE

```METHOD
## 改了什么