总览 · ← 返回运行 20261001-114429-search-t1-g18-continue

节点 n8 终选程序?按该运行锁定的规则最终选出的程序；可能是候选节点，也可能由护栏回退到基线。在终选来历上

冻结父本类型配额，类内按增殖打分做成熟端加权无放回抽样（θ=0.5，E-S 算法），3-seed 均值 56.06→58.13。

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 8936c91c49: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 39e48cc98a。改动的文件：solution/METHOD.md +49 −0、solution/README.md +4 −3、solution/mature.py +148 −0、solution/run.py +26 −6

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..299dc91--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,49 @@+冻结父本类型配额，类内按增殖打分做成熟端加权无放回抽样（θ=0.5，E-S 算法），3-seed 均值 56.06→58.13。++## 方法++- 类型配额与父节点 2（heart_jcf_peri）逐位一致：复用 `src.task1_temporal.reweight` 的+  `type_weights`（心脏×1.6、边缘×0.25、丢 Neural Tube）+ `largest_remainder`，N=4000。+- 新自由度：**类内选谁**。每细胞成熟分 `m_i = −z_c(cycle_i)`：+  - `cycle_i` = 面板基因全细胞 z 分数后取 `HALLMARK_E2F_TARGETS`（199 基因，prior/msigdb GMT 现场读取）集内均值；+  - 在每个类型内部再 z 化；凋亡项（HALLMARK_APOPTOSIS，161 基因）实测降分，λ=0 出货。+- 类内 Efraimidis–Spirakis 加权无放回抽样：`v_i = exp(clip(θ·m_i, ±log4))`，`key_i = log(u_i)/v_i`，每类取 top n_c。+  单一 `default_rng(seed)`，u_i 按原始行号一次性生成（θ 间配对比较）。θ=0 且 λ=0 时走父本 `heart_reweight` 原路径（逐位复现 55.97）。+- 表达矩阵不做任何修改，输出 100% 真实细胞。++## 关键参数++θ=0.5（出货），λ_apopt=0，N=4000，clip=±log4。++## 查分结果（proxy，T1:val）++| 配置 | seed 0 | seed 1 | seed 2 | 均值 |+|---|---|---|---|---|+| 父本 A0（θ=0） | 55.97 | 56.32 | 55.90 | **56.06** |+| θ=+0.25 | 57.40 | - | - | - |+| θ=+0.5 | 57.89 | 58.34 | 58.16 | **58.13** |+| θ=+0.75 | 57.62 | - | - | - |+| θ=−0.25 | 52.14 | - | - | - |+| θ=+0.5, λ=0.5 | 57.13 | - | - | - |++θ=+0.5 分组（seed 0）：de_recovery 52.53 / direction 60.75 / cell_state 61.44 / covariation 55.67。+相对 A0：cell_state +4.9、direction +2.2、covariation +1.0、de_recovery −0.5（3-seed 均值口径），+出货门槛（均值 ≥ +2.0 且无组跌 >3.0）满足。++## 验证过的++- A0 逐位复现 55.97（锚点）；ship 配置对 refactor 前后逐位一致；vec-check 通过；运行 ~6 s、内存与父本同量级。+- 符号方向确定：负 θ 大幅降分，正 θ 在 0.25–0.75 区间稳定优于父本，峰值 ≈0.5。+- 深度诊断：corr(cycle, log10 depth)=0.366 < 0.5，未做深度残差化；θ=0.5 输出深度均值仅 −1.6%。+- 方差守卫：随机 3000 基因上，θ=0.5 vs A0 的全基因方差比 1.045，无类内塌缩。+- 单输入阶段（proxy）与两输入（final）走完全相同代码路径（只读 `inputs_by_time[-1]`），θ 可直接迁移。++## 没验证的++- final 视图（E8.5+E9.5，目标 E10.5）上的实际分数；E9.5 类型集更大（21 型），θ=0.5 的幅度是否仍最优未测。+- θ 在 0.5–0.75 之间的细调（有意不做，防 proxy 过拟合）。+- prior GMT 缺失时的退化路径（代码上回退为类内均匀抽样=父本语义，未在真实缺失场景跑过；final 视图 prior 应存在）。++## 下一步建议++在 final 两阶段视图上，可用 E8.5→E9.5 的每类型伪批量差值替代/叠加增殖打分作为 m_i（proxy 无法验证此扩展）；θ 细网格 {0.4,0.5,0.6} × 3 seed。diff --git a/solution/README.md b/solution/README.mdindex e66c7c3..92fd13c 100644--- a/solution/README.md+++ b/solution/README.md@@ -1,4 +1,5 @@-# heart_jcf_peri+# heart_jcf_peri + 类内成熟端抽样（node 8） -run2（20260927-172643-grok-t1）的 T1 胜者：对最新输入阶段按细胞类型重加权再抽 4000 个真实细胞，不改表达。丢 Neural Tube；心脏类（CM、内皮/心内膜、SHF/PHM、JCF、心包、心外膜原基）×1.6；表面外胚层、EXEM、轴旁中胚层 ×0.25；其余 ×1。-proxy 预期 55.97（seed 0，与 run2 的预测逐位相同）。final 读 E9.5，类型集合已含 E9.5 的名字（V-CM、Endocardium、BEC、aPHM、pPHM、Proepicardium）。+父本类型配额（心脏×1.6 / 边缘×0.25 / 丢 Neural Tube，largest_remainder，4000 细胞）逐位不动；+类内用 prior/msigdb `HALLMARK_E2F_TARGETS` 增殖打分（每类型内 z 化取负）做 Efraimidis–Spirakis+加权无放回抽样，θ=0.5。表达不修改。proxy 3-seed 均值 58.13（父本 56.06）。diff --git a/solution/mature.py b/solution/mature.pynew file mode 100644index 0000000..a1ab620--- /dev/null+++ b/solution/mature.py@@ -0,0 +1,148 @@+"""Within-type mature-end weighted resampling.++Keeps the parent's type quotas bit-for-bit (largest_remainder on+counts * type_weights from src.task1_temporal.reweight), but inside each type+selects cells by a proliferation-derived maturity score m_i = -z_c(cycle_i)+(- lam * z_c(apopt_i)), using Efraimidis-Spirakis weighted sampling without+replacement: key_i = log(u_i) / v_i, v_i = exp(clip(theta * m_i, +-log4)).++Scores come only from the view's prior/ GMT gene sets (HALLMARK_E2F_TARGETS,+HALLMARK_APOPTOSIS) and the input stage expression. theta = 0 reduces to+uniform within-type sampling.+"""++from __future__ import annotations++import sys+from pathlib import Path++import numpy as np+from scipy import sparse++from src.task1_temporal.reweight import DROP_TYPES, largest_remainder, type_weights++CLIP = np.log(4.0)+PROLIF_SET = "HALLMARK_E2F_TARGETS"+APOPT_SET = "HALLMARK_APOPTOSIS"+++def _read_gmt(path: Path) -> dict[str, list[str]]:+    out = {}+    with path.open() as fh:+        for line in fh:+            parts = line.rstrip("\n").split("\t")+            if len(parts) >= 3:+                out[parts[0]] = parts[2:]+    return out+++def _gmt_path(view, manifest: dict) -> Path | None:+    for entry in manifest.get("prior", []):+        p = Path(view) / entry["path"]+        if p.is_dir():+            cand = p / "hallmark_mouse.gmt"+            if cand.exists():+                return cand+        elif p.suffix == ".gmt" and p.name.startswith("hallmark"):+            return p+    return None+++def set_scores(X, genes: list[str], view, manifest: dict, lam_apopt: float,+               verbose: bool = True):+    """Per-cell z-scored set means for proliferation and apoptosis.++    Returns (None, None) when the prior GMT is unavailable -> caller must fall+    back to uniform within-type sampling.+    """+    gmt_file = _gmt_path(view, manifest)+    if gmt_file is None:+        return None, None+    gmt = _read_gmt(gmt_file)+    if PROLIF_SET not in gmt:+        return None, None+    gene_idx = {g: i for i, g in enumerate(genes)}+    sets = {}+    for name, key in ((PROLIF_SET, "cycle"), (APOPT_SET, "apopt")):+        if key == "apopt" and (lam_apopt == 0.0 or name not in gmt):+            sets[key] = np.zeros(0, dtype=int)+            continue+        idx = [gene_idx[g] for g in gmt[name] if g in gene_idx]+        sets[key] = np.asarray(idx, dtype=int)+        if verbose:+            print(f"[mature] {name}: {len(idx)}/{len(gmt[name])} genes on panel",+                  file=sys.stderr)++    sub = X[:, np.concatenate([sets["cycle"], sets["apopt"]])].toarray().astype(np.float64)+    mu = sub.mean(axis=0)+    sd = sub.std(axis=0)+    sd[sd < 1e-12] = 1.0+    z = (sub - mu) / sd+    nc = len(sets["cycle"])+    cycle = z[:, :nc].mean(axis=1)+    apopt = z[:, nc:].mean(axis=1) if sets["apopt"].size else np.zeros(X.shape[0])++    if verbose:+        depth = np.log10(np.maximum(np.asarray(X.sum(axis=1)).ravel(), 1e-9))+        c = np.corrcoef(cycle, depth)[0, 1]+        print(f"[mature] corr(cycle, log10 depth) = {c:.3f}", file=sys.stderr)+    return cycle, apopt+++def mature_reweight_full(X, labels: np.ndarray, genes: list[str], view,+                         manifest: dict, n_cells: int, seed: int, theta: float,+                         lam_apopt: float = 0.0, heart_weight: float = 1.6,+                         edge_weight: float = 0.25, verbose: bool = True):+    cycle, apopt = set_scores(X, genes, view, manifest, lam_apopt, verbose)++    types = [str(t) for t in np.unique(labels) if str(t) not in DROP_TYPES]+    counts = np.array([(labels == t).sum() for t in types], dtype=np.float64)+    alloc = largest_remainder(counts * type_weights(types, heart_weight, edge_weight),+                              n_cells)++    # maturity score, z-normalised within each type+    m = np.zeros(X.shape[0], dtype=np.float64)+    pools = {}+    for t in types:+        idx = np.flatnonzero(labels == t)+        pools[t] = idx+        if cycle is None:+            continue+        for s, w in ((cycle, -1.0), (apopt, -lam_apopt)):+            if w == 0.0:+                continue+            sc = s[idx]+            sd = sc.std()+            if sd < 1e-12:+                continue+            m[idx] += w * (sc - sc.mean()) / sd++    # paired uniforms: one draw per input cell, shared across all theta+    rng = np.random.default_rng(seed)+    u = rng.uniform(size=X.shape[0])+    u = np.clip(u, 1e-300, 1.0 - 1e-15)++    if verbose and theta != 0.0:+        print(f"[mature] theta={theta} lam={lam_apopt}", file=sys.stderr)++    blocks = []+    for t, n in zip(types, alloc):+        n = int(n)+        if n <= 0:+            continue+        pool = pools[t]+        if n >= pool.size:+            choice = rng.choice(pool, size=n, replace=True)+            blocks.append(X[choice])+            continue+        if theta == 0.0:+            v = np.ones(pool.size)+        else:+            v = np.exp(np.clip(theta * m[pool], -CLIP, CLIP))+        key = np.log(u[pool]) / v+        pick = pool[np.argsort(-key)[:n]]+        blocks.append(X[np.sort(pick)])++    out = sparse.vstack(blocks, format="csr").astype(np.float32)+    out.eliminate_zeros()+    return outdiff --git a/solution/run.py b/solution/run.pyindex a752b21..83ba1b5 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,10 +1,14 @@ #!/usr/bin/env python3-"""heart_jcf_peri: run2's T1 winner. Reweight the latest input stage by cell type.+"""heart_jcf_peri quotas + within-type mature-end resampling (node 8). -Neural Tube dropped; heart types (CM, endothelium/endocardium, SHF/PHM, JCF,-pericardium, proepicardium) x1.6; surface ectoderm, EXEM, paraxial mesoderm-x0.25; everything else x1. 4000 real cells, no expression shift. Proxy reads-E8.5, final reads E9.5; the type set already names both stages' labels.+Type quotas are bit-for-bit the parent's (heart x1.6, edge x0.25, drop Neural+Tube, largest_remainder). Inside each type, cells are drawn by+Efraimidis-Spirakis weighted sampling without replacement with weights+v_i = exp(clip(theta * m_i, +-log4)), where m_i = -z_c(E2F-proliferation score)+(- lam * z_c(apoptosis score)) is z-normalised within the type. Gene sets come+from the view's prior/msigdb hallmark GMT; no expression is modified.+theta = 0 (and lam = 0) falls back to the parent's heart_reweight exactly.+Same code path for 1-input (proxy) and 2-input (final) views. """  from __future__ import annotations@@ -23,6 +27,8 @@ from src.task1_temporal.view_io import ( )  N_CELLS = 4000+SHIP_THETA = 0.5+SHIP_LAM = 0.0   def main() -> None:@@ -30,13 +36,27 @@ def main() -> None:     parser.add_argument("--data", required=True)     parser.add_argument("--out", required=True)     parser.add_argument("--seed", type=int, default=0)+    parser.add_argument("--theta", type=float, default=SHIP_THETA)+    parser.add_argument("--lam", type=float, default=SHIP_LAM)     args = parser.parse_args()      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)     last = read_stage(args.data, inputs_by_time(manifest)[-1], genes)     n = target_n_cells(manifest, N_CELLS)-    X = heart_reweight(last.X, labels_of(last), n_cells=n, seed=args.seed)+    labels = labels_of(last)+    if args.theta == 0.0 and args.lam == 0.0:+        X = heart_reweight(last.X, labels, n_cells=n, seed=args.seed)+    else:+        import os+        import sys++        sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))+        from mature import mature_reweight_full++        X = mature_reweight_full(last.X, labels, genes, args.data, manifest,+                                 n_cells=n, seed=args.seed, theta=args.theta,+                                 lam_apopt=args.lam)     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

做了什么：冻结父本（heart_jcf_peri）类型配额，新增类内成熟端加权抽样——用 prior/msigdb E2F 增殖打分（类内 z 化取负）做 Efraimidis–Spirakis 无放回抽样，θ=0 时逐位退化回父本。

查分结果（proxy，共耗 9 次查询）：
- 父本 A0 3-seed 均值 56.06（55.97/56.32/55.90，seed 0 逐位复现）
- θ=+0.5：57.89/58.34/58.16，均值 **58.13**（+2.07，出货）；cell_state +4.9、direction +2.2、covariation +1.0、de_recovery −0.5
- θ=+0.25→57.40，+0.75→57.62，−0.25→52.14（符号明确）；凋亡项 λ=0.5→57.13（负结果，弃用）
- 守卫通过：深度 corr 0.366、方差比 1.045、运行 1.9s、vec-check ok

下一步最值得试：final 视图上用 E8.5→E9.5 每类型伪批量差值叠加/替代增殖打分作 m_i；θ 细网格 {0.4,0.5,0.6}×3 seed；de_recovery 是唯一受损组，可试仅对非心脏类型施加 θ 或对 DE 基因加权打分。

调研员

PLAN.json written

审查员

REVIEW.json written