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

节点 n11

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 4c2e81d462: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 +64 −10

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex e9f77d1..46c39bc 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。+将固定 ALPHA=0.5 替换为逐类型自适应收缩：计算每个类型中 SNR>2 的基因比例（frac_significant），据此在 [0.1, 0.75] 间插值类型级 alpha；同时用 (snr/(snr+2))^1.5 作为逐基因收缩因子，对低信噪比基因更强压缩。相比父节点的固定 alpha + 线性收缩，这更激进地抑制噪声基因位移（改善 covariation），同时对有真实变化的类型允许更大幅度平移（改善 direction 和 X3）。 ## 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：α∈[0,1] 不能默认 1，逐基因收缩让噪声基因少动。实验表 Program 1（48.83）验证了自适应 alpha + power shrinkage 组合在 covariation（46.18）和 X3（46.39）上的优势。diff --git a/solution/run.py b/solution/run.pyindex dabccbd..31fbe4a 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,10 +1,11 @@ #!/usr/bin/env python3-"""pseudobulk_shift with official-stage filtering.+"""Pseudobulk shift with per-gene SNR shrinkage and conservative alpha. -Uses only official (non-external) stages for sampling and delta computation.-This fixes proxy2 where the latest input is Qiu E9.0 (different cell type-labels, heart-only subset) which must not be used as the prediction base.-Falls back to all stages when no official stages exist (e.g. X3 test).+Uses only official (non-external) stages. Applies per-gene empirical Bayes+shrinkage to the pseudobulk delta: genes with large signal relative to+sampling noise are shifted more, noisy genes are shrunk toward zero. This+preserves covariation structure by avoiding large spurious shifts that+would be clipped at zero. Falls back to copy_last with a single input. """  from __future__ import annotations@@ -12,8 +13,8 @@ from __future__ import annotations import argparse  import numpy as np+from scipy import sparse -from src.task1_temporal.baselines import shift_rows, type_deltas from src.task1_temporal.view_io import (     inputs_by_time,     labels_of,@@ -25,6 +26,17 @@ from src.task1_temporal.view_io import (     write_prediction, ) +ALPHA_MAX = 0.75+ALPHA_MIN = 0.1+SIGN_THRESHOLD = 0.05+GENE_SHRINK_POWER = 1.5+++def _to_dense(X):+    if sparse.issparse(X):+        return np.asarray(X.todense(), dtype=np.float32)+    return np.asarray(X, dtype=np.float32)+  def main() -> None:     parser = argparse.ArgumentParser()@@ -43,14 +55,56 @@ 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]+    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))-        del prev-        X = shift_rows(X, labels_of(last)[rows], deltas)+        labels_last = labels_of(last)+        labels_prev = labels_of(prev)++        Xd_last = _to_dense(last.X)+        Xd_prev = _to_dense(prev.X)++        types_prev = set(np.unique(labels_prev))+        labels_sampled = labels_last[rows]++        for t in np.unique(labels_sampled):+            mask_s = labels_sampled == t+            if t not in types_prev:+                continue+            mask_last = labels_last == t+            mask_prev = labels_prev == t+            n_last = int(mask_last.sum())+            n_prev = int(mask_prev.sum())++            mean_last = Xd_last[mask_last].mean(axis=0)+            mean_prev = Xd_prev[mask_prev].mean(axis=0)+            raw_delta = mean_last - mean_prev++            var_last = Xd_last[mask_last].var(axis=0)+            var_prev = Xd_prev[mask_prev].var(axis=0)+            noise_se = np.sqrt(+                var_last / max(n_last, 1) + var_prev / max(n_prev, 1) + 1e-12+            )+            snr = np.abs(raw_delta) / (noise_se + 1e-12)+            significant = snr > 2.0+            frac_significant = significant.mean()++            if frac_significant < SIGN_THRESHOLD:+                alpha = ALPHA_MIN+                gene_shrink = np.ones_like(raw_delta)+            else:+                alpha = ALPHA_MIN + (ALPHA_MAX - ALPHA_MIN) * min(+                    frac_significant / 0.3, 1.0+                )+                gene_shrink = (snr / (snr + 2.0)) ** GENE_SHRINK_POWER++            delta = (raw_delta * alpha * gene_shrink).astype(np.float32)+            X[mask_s] += delta[np.newaxis, :]++        del prev, Xd_last, Xd_prev +    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

审查员

REVIEW.json written

stream

Looking at the results, the main remaining gap is **direction** (49.48 vs 50.96 for copy_last) and **X3** (46.32). The key insight: X3 has inputs at E8.75 and E9.0 (0.25-day gap) predicting E9.5 (0.5-day gap), so we're under-extrapolating by 2x. Adding a time-aware extrapolation factor should boost direction on X3 specifically.

<<<<<<< SEARCH
    if len(stages) >= 2:
        prev = read_stage(args.data, stages[-2], genes)
        labels_last = labels_of(last)
        labels_prev = labels_of(prev)

        Xd_last = _to_dense(last.X)
        Xd_prev = _to_dense(prev.X)

        types_prev = set(np.unique(labels_prev))
        labels_sampled = labels_last[rows]

        for t in np.unique(labels_sampled):
            mask_s = labels_sampled == t
            if t not in types_prev:
                continue
            mask_last = labels_last == t
            mask_prev = labels_prev == t
            n_last = int(mask_last.sum())
            n_prev = int(mask_prev.sum())

            mean_last = Xd_last[mask_last].mean(axis=0)
            mean_prev = Xd_prev[mask_prev].mean(axis=0)
            raw_delta = mean_last - mean_prev

            var_last = Xd_last[mask_last].var(axis=0)
            var_prev = Xd_prev[mask_prev].var(axis=0)
            noise_se = np.sqrt(
                var_last / max(n_last, 1) + var_prev / max(n_prev, 1) + 1e-12
            )
            snr = np.abs(raw_delta) / (noise_se + 1e-12)
            significant = snr > 2.0
            frac_significant = significant.mean()

            if frac_significant < SIGN_THRESHOLD:
                alpha = ALPHA_MIN
                gene_shrink = np.ones_like(raw_delta)
            else: