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

节点 n44

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 429cb629c7: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 bea86622fc。改动的文件：solution/METHOD.md +9 −9、solution/run.py +111 −47

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 1d34aad..57008dc 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,12 +1,12 @@ ## 改了什么-相对父节点（node 22，score 49.10）的全部改动：-1. 完全移除 sign disagreement 机制（第 0 轮验证有效）。-2. 将基因收缩函数从 (snr/(snr+K))^power 改为线性斜坡+软阈值：gene_shrink = clip((snr - SOFT_THRESHOLD)/(K_RAMP - SOFT_THRESHOLD), 0, 1)^SHRINK_POWER。理由：旧函数饱和太慢（SNR=10 时仅保留 47% 位移），导致 de_recovery 持续偏低（48.70）。新函数在 SNR≤1.5 时完全归零（去除噪声，保护 covariation），在 SNR≥5.0 时保留 100% 位移（保留 DE 基因信号，提升 de_recovery）。-3. 参数：SOFT_THRESHOLD=1.5, K_RAMP=5.0, SHRINK_POWER=2.0, ALPHA_MAX=0.80（回退到第 0 轮值）。-4. 移除第 1 轮的自适应 K（K_MIN/K_MAX 按类型插值），因实验表明无增益（49.35 vs 49.37）。-5. 保留 include_external=False、单阶段退化为 copy_last、输出前 clip≥0。+相对父节点（node 29，score 49.28）的全部改动：+1. 加入组成重加权（composition reweighting）：按心脏家族（CARDIAC_SUBSTRINGS）上调权重 1.6，非心脏外胚层（Surface Ectoderm、EXEM、Paraxial Mesoderm）下调到 0.25，Neural Tube 完全去除。这是 node 40（score 53.10）相对本分支最大的结构性差异，使 proxy/proxy2 从 50.04 提升到 54.49。+2. 加入细胞尺度投影（cell-scale projection）：对每个类型，沿 delta 方向投影每个细胞，根据投影分位数给予 0.75–1.25 的缩放因子，使位移异质化，保护协方差结构。+3. 将基因收缩函数改回有理函数形式 (snr/(snr+K))^power，K=2.5，与最佳节点一致。移除线性斜坡+软阈值（实验证明无增益）。+4. 移除自适应 alpha（ALPHA_MIN/ALPHA_MAX/frac_significant），改用固定 DELTA_ALPHA=0.6。+5. 保留 include_external 逻辑：有官方阶段时排除外部输入，无官方阶段时（如 X3）使用全部输入。 ## 用到的知识与出处-- 第 0 轮实验：K=3.5 全局收缩得分 49.37，covariation 48.95 但 de_recovery 仅 48.70-- 第 1 轮实验：自适应 K 无增益（49.35），回退+- 实验表 node 40（score 53.10）：组成重加权 + 细胞尺度投影 + SHRINK_K=2.5 的组合 - 方法卡 k018：SNR 收缩框架-- 分析：(snr/(snr+K))^2.5 对 SNR=10 的基因仍收缩 53%，是 de_recovery 瓶颈；线性斜坡在 SNR≥5 时完全保留位移+- 父节点 ANALYSIS：de_recovery 瓶颈不在收缩函数形状，而在结构性改动（组成重加权）+- 方向库 T1-01（组成重加权）+ T1-02（伪批量位移+收缩）diff --git a/solution/run.py b/solution/run.pyindex c6b4fdf..62bf61c 100644--- a/solution/run.py+++ b/solution/run.py@@ -21,18 +21,64 @@ from src.task1_temporal.view_io import (     load_manifest,     panel_genes,     read_stage,-    sample_rows,     target_n_cells,     write_prediction, ) -ALPHA_MAX = 0.80-ALPHA_MIN = 0.1-SIGN_THRESHOLD = 0.05+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+DELTA_ALPHA = 0.6+SHRINK_K = 2.5 SHRINK_POWER = 2.0-SOFT_THRESHOLD = 1.5-K_RAMP = 5.0-SNR_SIGNIFICANT_THRESHOLD = 1.5+CELL_SCALE_MAX = 1.25+CELL_SCALE_MIN = 0.75+++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 weighted_sample(labels, n_target, rng, weight_fn=None):+    n = len(labels)+    if weight_fn is None:+        w = np.ones(n, dtype=np.float64)+    else:+        w = np.array([weight_fn(t) for t in labels], dtype=np.float64)+    total = w.sum()+    if total <= 0:+        w = np.ones(n, dtype=np.float64)+        total = float(n)+    p = w / total+    if n_target <= n:+        rows = rng.choice(n, size=n_target, replace=False, p=p)+    else:+        rows = rng.choice(n, size=n_target, replace=True, p=p)+    return rows+++def has_official_stage(manifest) -> bool:+    for inp in manifest.get("inputs", []):+        if inp.get("source") != "external":+            return True+    return False   def _to_dense(X):@@ -50,62 +96,80 @@ def main() -> None:      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)--    stages = inputs_by_time(manifest, include_external=False)-    if not stages:-        stages = inputs_by_time(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)-    rows = sample_rows(last.n_obs, target_n_cells(manifest, last.n_obs), rng)++    labels = labels_of(last)+    n_target = target_n_cells(manifest, last.n_obs)+    max_c = manifest.get("max_cells", None)+    min_c = manifest.get("min_cells", None)+    if max_c is not None:+        n_target = min(n_target, int(max_c))+    if min_c is not None:+        n_target = max(n_target, int(min_c))+    n_target = max(n_target, 1)++    if official:+        rows = weighted_sample(labels, n_target, rng, weight_fn=family_weight)+    else:+        if n_target <= last.n_obs:+            rows = rng.choice(last.n_obs, size=n_target, replace=False)+        else:+            rows = rng.choice(last.n_obs, size=n_target, replace=True)     X = _to_dense(last.X[rows])      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 = set(np.unique(labels_prev))-        labels_sampled = labels_last[rows]+        deltas = {}+        types_last = set(np.unique(labels))+        types_prev = set(np.unique(labels_prev))+        for t in types_last & types_prev:+            mask_l = labels == t+            mask_p = labels_prev == t+            n_l = int(mask_l.sum())+            n_p = int(mask_p.sum())+            raw_delta = Xd_last[mask_l].mean(axis=0) - Xd_prev[mask_p].mean(axis=0)+            var_l = Xd_last[mask_l].var(axis=0)+            var_p = Xd_prev[mask_p].var(axis=0)+            noise_se = np.sqrt(var_l / max(n_l, 1) + var_p / max(n_p, 1) + 1e-12)+            snr = np.abs(raw_delta) / (noise_se + 1e-12)+            gene_shrink = (snr / (snr + SHRINK_K)) ** SHRINK_POWER+            deltas[t] = (raw_delta * DELTA_ALPHA * gene_shrink).astype(np.float32) +        labels_sampled = labels[rows]         for t in np.unique(labels_sampled):-            mask_s = labels_sampled == t-            if t not in types_prev_set:+            if t not in deltas:                 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 > SNR_SIGNIFICANT_THRESHOLD-            frac_significant = significant.mean()--            if frac_significant < SIGN_THRESHOLD:-                alpha = ALPHA_MIN-                gene_shrink = np.ones_like(raw_delta)+            mask_s = labels_sampled == t+            delta_t = deltas[t]+            nonzero = np.abs(delta_t) > 1e-8+            if nonzero.sum() > 1 and mask_s.sum() > 2:+                centroid = Xd_last[labels == t][:, nonzero].mean(axis=0)+                proj_dir = delta_t[nonzero]+                proj_norm = np.linalg.norm(proj_dir) + 1e-12+                proj_dir = proj_dir / proj_norm+                cell_sub = X[mask_s][:, nonzero]+                centered = cell_sub - centroid[np.newaxis, :]+                proj_scores = centered @ proj_dir+                p_lo = np.percentile(proj_scores, 5)+                p_hi = np.percentile(proj_scores, 95)+                span = p_hi - p_lo + 1e-12+                norm_proj = (proj_scores - p_lo) / span+                scale = CELL_SCALE_MIN + (CELL_SCALE_MAX - CELL_SCALE_MIN) * norm_proj+                X[mask_s][:, nonzero] += delta_t[nonzero][np.newaxis, :] * scale[:, np.newaxis]+                zero_mask = ~nonzero+                if zero_mask.any():+                    X[mask_s][:, zero_mask] += delta_t[zero_mask][np.newaxis, :]             else:-                alpha = ALPHA_MIN + (ALPHA_MAX - ALPHA_MIN) * min(-                    frac_significant / 0.3, 1.0-                )-                gene_shrink = np.clip(-                    (snr - SOFT_THRESHOLD) / (K_RAMP - SOFT_THRESHOLD), 0.0, 1.0-                ) ** SHRINK_POWER--            delta = (raw_delta * alpha * gene_shrink).astype(np.float32)-            X[mask_s] += delta[np.newaxis, :]+                X[mask_s] += deltas[t][np.newaxis, :]          del prev, Xd_last, Xd_prev 

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

I'll read the rest of the PROMPT.md to see the current program and any remaining instructions.

<tool_call>
<function=read_file>
<parameter=path>
/home/spark-longxinyang/.local/state/vec-report/calls/earch-t1-abc-r0-C-native-n44-4f6059/work/PROMPT.md
</parameter>
<parameter=offset>
715
</parameter>
</function>
</tool_call>