总览 · ← 返回运行 20261002-034201-search-t1-abc-r1-B-population

节点 n14

官方最新阶段分层按型抽样复制 + 心脏解剖重加权（神经管/表面外胚层×0、旁轴×0.1、心脏谱系×1.3）+ 反向组成趋势外推 GAMMA=-1.2（仅词表可比的多阶段视图）+ 输出下限 3000；表达值不改。

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 fcdf90827c: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 30011bc035。改动的文件：solution/METHOD.md +28 −27、solution/run.py +164 −167

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 1197f4d..ac96fab 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,38 +1,39 @@-官方最新阶段分层抽样复制（表达不改）+ 心脏解剖重加权（神经管/表面外胚层丢弃、旁轴×0.1、心脏×1.3）+ 反向组成趋势外推 GAMMA=-1.2 + 输出数下限 3000。+官方最新阶段分层按型抽样复制 + 心脏解剖重加权（神经管/表面外胚层×0、旁轴×0.1、心脏谱系×1.3）+ 反向组成趋势外推 GAMMA=-1.2（仅词表可比的多阶段视图）+ 输出下限 3000；表达值不改。  ## 方法 -在父节点 5（分层复制 + 阻尼组成趋势外推）基础上，加入本 run 已被 node 6/7/9/10 反复验证为唯一稳定提分的部件——**心脏解剖组成重加权**，并微调其强度、加入输出细胞数下限。表达值从不修改，只重采样真实细胞。+父节点 12 的程序目录在本节点 workspace 中不存在，按其 ANALYSIS/变化量表**逐项重建**并在 proxy A 半验证（seed 0：57.64，父 57.09，噪声内；X3：52.17，父 53.53，噪声内）。机制： -流程（每次运行现算，全部来自 view 输入）：+1. 基座 = 最新**官方**输入阶段（proxy2 忽略 Qiu E9.0；X3 无官方输入时退回全部输入的最新者 E9.0）。+2. 组成重加权（键在官方类型名，未知名字权重 1.0，故 X3 上为 no-op）：+   - DROP（×0）：Neural Tube、Surface Ectoderm+   - Paraxial Mesoderm ×0.1；EXEM ×1.0+   - 心脏谱系（HEART_TYPES，同 `src.task1_temporal.reweight` 名单）×1.3+3. GAMMA=-1.2 反向组成趋势外推：仅当视图有两个类型词表可比的输入阶段（X3 的 E8.75/E9.0）时，按型乘 `(p_last/p_prev)^GAMMA`（比值夹在 [0.2,5]）；proxy（单阶段）、proxy2（词表不可比）自动跳过。+4. n_out = clip(max(3000, 池大小), min_cells, max_cells)；按权重最大余数法分配，型内均匀随机抽样（配额>池时有放回，X3 上约 27–38% 重复）。表达值从不修改，rng=default_rng(seed)，确定。 -1. **基座**：输出细胞一律取最新**官方**输入阶段（`official[-1]`）；无官方阶段的外部测试题（X3）退而取最新输入。proxy2 的 Qiu E9.0 心脏细胞只当参考、不作基座（沿用 node 2/3 结论）。-2. **组成趋势外推（GAMMA）**：当最新两个输入阶段的细胞类型词表可比（覆盖率 ≥ `OVERLAP_MIN=0.5`，如 X3 的 E8.75+E9.0、final 的官方 E8.5+E9.5），按 `p_target = p_last + GAMMA·(Δt_target/Δt_last)·(p_last − p_prev)` 外推每型比例，夹到 `MIN_FRAC=0.002` 再归一。词表不可比（proxy2：全胚 vs 心脏标签不相交）或只有一个阶段（proxy）时**跳过**，保留最新阶段原比例。实测 GAMMA 只作用于 X3（proxy 单阶段、proxy2 词表不可比都跳过）。-3. **解剖重加权**：对趋势后的比例乘以每型权重再归一——心脏谱系（CM 各型、SHF、PHM、心内膜/内皮、BEC、JCF、心包、心外膜）×`HEART_W=1.3`；旁轴中胚层 ×`PARAXIAL_W=0.1`；神经管、表面外胚层 ×0（丢弃，心脏为中心解剖时不在取样内）；EXEM 及其余 ×1.0；未知名字（X3 的 heart field 标签）×1.0，故重加权在 X3 上是 no-op。-4. **输出细胞数**：`n = clip(max(target_n×OUT_FRAC(0.95), N_FLOOR=3000), min_cells, max_cells)`。池足够大（proxy/proxy2，E8.5=16787）→ n≈4862，类型内**不放回**抽样、最大余数法分配、禁止重复取细胞；池小于目标（X3，E9.0=2174）→ n=3000，允许类型内重复补齐。-5. 按分配从各型抽真实细胞，表达原样输出，`write_prediction` 写成合规 h5ad。+## 本节点实验（A 半，均 seed 0/1） -## 关键参数（均在 A 半实测）+| 实验 | proxy | X3 | 结论 |+|---|---|---|---|+| 基线重建（=父机制） | 57.64 | 52.17 | 与父 57.09/53.53 噪声内一致 |+| PLAN 核心：型内 PC1(500 HVG) 等距分层抽样 | 56.65 / 57.05 (s0/s1) | — | **失败**：de_recovery 53.0 不动（de_score 恒 0.109），cell_state 60.9→58.1，gate（de_recovery +1.5）不满足，弃用 |+| HEART_W 1.3→1.6 | 56.16 / 56.10 | — | 更差，回退 |+| PARAXIAL_W 0.1→0 | 56.75 / 56.79 | — | 略差（均值 vs 57.64），回退 |+| GAMMA=0（关外推） | — | 47.32 | **GAMMA=-1.2 在 X3 上强正贡献 +4.9**，确认保留 |+| GAMMA=-2.0 | — | 52.61 | 与 -1.2 差 +0.44 在噪声内，按平台稳健原则保留 -1.2 |+| proxy2 心脏谱系向 Qiu E9.0 伪批量方向平移 α=0.25（median-centered，仅 covered 基因） | 41.78 (proxy2) | — | **灾难性失败**：cell_state 33、covariation 24——跨平台（10x vs sci-RNA-seq3）批次差主导 delta，α 再小也不可行，已禁用（ALPHA_EXT=0） | -- `GAMMA=-1.2`：X3 扫描 −1.0/−1.2/−1.4 → 52.58/53.60/53.67，−1.2~−1.4 为峰，取 node 9 已在 B 半验证的 −1.2。负号=向上一阶段组成收缩（父节点教训：符号方向须实测，不能靠直觉）。-- `HEART_W=1.3`：proxy 扫描 1.0/1.15/1.3/1.6/2.0 → 57.05/57.11/57.18/56.95/56.13，1.0–1.3 平台、2.0 掉分，取 1.3。-- `PARAXIAL_W=0.1`、神经管/表面外胚层 ×0、EXEM ×1.0：沿用 node 10 的细拆（EXEM 单调剂量峰在 ×1.0、表面外胚层丢弃）。-- `N_FLOOR=3000`：X3 扫描 2500/3000/3500 → 53.23/53.60/52.90，3000 为峰。-- `OUT_FRAC=0.95`、`MIN_FRAC=0.002`、`OVERLAP_MIN=0.5`：沿用父节点。+最终提交 = 基线重建配置（STRATIFY=False、ALPHA_EXT=0），与已打分的 `base_proxy` 逐字节一致。 -## 验证过什么+## 验证 / 未验证 -- **三视图 seed 0 均跑通 + `vec-check` ok**；seed 1、2 亦跑通、格式合规；同 seed 重跑输出逐元素相同（确定性）。运行 ~1–2 s，峰值内存远低于 28 GB 上限。-- **A 半查分**：proxy 57.18 / proxy2 57.18 / X3 53.60 → 节点均分 **≈55.99**（父节点 50.60，当前最佳 node 9 rank3 55.57）。分组：proxy cell_state 59.85、direction 59.68、covariation 55.26、de_recovery 53.0；X3 cell_state 57.65、covariation 50.59、direction/de_recovery ~52。-- **消融**：within-type 增殖偏置（prior GO/Reactome/hallmark 细胞周期基因打分，BETA=0.3）——X3 因 n>池 未触发（no-op），proxy **−5.0**（cell_state 59.4→51.1、covariation 54.9→48.7），按 PLAN gate 丢弃，代码已删除。+- 已验证：proxy、proxy2、X3 三视图 vec-check 全 ok，同 seed 输出确定（md5 一致），运行 ~5s、峰值 ~2GB。+- 未验证：final 视图（无沙箱）；GAMMA 在官方 E8.5+E9.5 两阶段上的效果（机制上会在 final 触发，方向按 X3 证据保留）。+- 生物学知识来源：心脏解剖重加权名单来自 `modeling/src/task1_temporal/reweight.py`（harness 提供的公开基线模块），未使用任何保留阶段测量信息。 -## 没验证 / 局限+## 下一步建议 -- 只用 A 半查分；节点正式分用 B 半，小幅差异（HEART_W 1.0↔1.3、GAMMA −1.2↔−1.4 皆 <2 分噪声）未必在 B 半重现，故取平台内稳健值而非 A 半 argmax。-- **final 视图（官方 E8.5+E9.5，目标 E10.5）未测**：那里 GAMMA 会真正作用于官方两阶段、且 E9.5 出现的新型（V-CM/Endocardium/BEC/aPHM/pPHM/Proepicardium 等）已写入 HEART_TYPES 名单以正确加权，但无沙箱可查分。-- proxy2 仍完全忽略 Qiu E9.0（只当参考），未尝试用它做心脏谱系的时间插值——留作后续。-- X3 输出含约 38% 重复细胞（池 2174 < n 3000）；实测净收益为正，但重复对 de_recovery 的长期影响未单独隔离。--## 生物学知识来源--仅用通用小鼠胚胎学定性事实：晚期以心脏为中心解剖 → 心脏谱系富集、神经管/表面外胚层/旁轴中胚层相对缩减（node 6/reweight.py 已记录的解剖动机）。类型名单取自**输入阶段**（已发布 E8.5/E9.5 注释），未使用任何保留阶段/基因型的测量。+1. X3 最弱组是 de_recovery/covariation 且 27–38% 细胞为重复——试按型内 k-NN 去重的重复策略或提高 N_FLOOR 至 4000（X3 max_cells=5000）。+2. proxy2 想利用 Qiu E9.0，必须先做跨平台校正（如只取 Qiu 内部 E8.75→E9.0 的方向，但 proxy2 未挂载 E8.75，可请求组委会或放弃）。+3. de_recovery 在纯重采样家族里恒定 53.0（de_score 0.109），说明该组天花板受"不改表达值"约束；要突破需合法的方向性表达修正（如 prior TF 调控网络约束下的保守位移），但官方 pseudobulk_shift 证据（48.6）表明全局位移有害。diff --git a/solution/run.py b/solution/run.pyindex 00b38d8..0909460 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,195 +1,192 @@-#!/usr/bin/env python3-"""Stratified copy of the latest usable stage + damped composition-trend extrapolation.--Output population = cells sampled from the latest *official* input stage-(fallback: latest input of any kind when the view has no official stage, e.g.-external test questions like X3). Expression is copied unchanged (the seed's-pseudobulk shift measured worse than copy on every ruler; alpha = 0).--Composition (cell-type proportions): when the two latest input stages share a-comparable cell-type vocabulary (same dataset/annotation, e.g. X3's E8.75 +-E9.0, or T1 final's official E8.5 + E9.5), the per-type proportion trend-r = (p2 - p1) / dt is extrapolated to the target time with a damping factor-GAMMA, clipped and renormalised; cells are then drawn per type from the latest-stage (with replacement inside a type when needed). When the two latest stages-are not comparable (T1 proxy2: official whole-embryo E8.5 vs external heart-only-Qiu E9.0 with disjoint labels) or there is only one stage (T1 proxy), the-proportions of the latest stage are kept as-is.--On top of the trend, an anatomical reweighting of the target composition is-applied (parent node 6/7/9/10 winning ingredient): the later stage is a-heart-centred dissection, so heart-lineage types are up-weighted (HEART_W),-paraxial mesoderm is down-weighted (PARAXIAL_W), and neural tube / surface-ectoderm (absent from the heart-centred later sample) are dropped. Weights are-keyed on *input-stage* type names (published E8.5/E9.5 annotations); unknown-names (e.g. X3's heart-field labels) get weight 1.0, so the reweighting is a-no-op there and only the GAMMA trend acts. The output size is floored at-N_FLOOR cells; when the pool is smaller (X3, 2174 cells) cells are reused to-reach it -- measured best at 3000 for X3.--Nothing about held-out stages/genotypes is used; all proportions, rates and-type names are computed/read from the view's inputs at runtime. The only-external knowledge is the qualitative lineage fact that a heart-centred later-dissection enriches cardiac lineages and depletes neural-tube / surface-ectoderm-/ paraxial tissue (textbook mouse embryology, not a held-out measurement).-"""+"""T1 candidate: resample the latest official stage with anatomical reweighting. +Reconstructed node-12 mechanism + within-type expression-stratified sampling.+Expression values are never modified; only real cells are resampled.+""" from __future__ import annotations  import argparse+from pathlib import Path  import numpy as np+from scipy import sparse -from src.task1_temporal.view_io import (-    inputs_by_time,-    is_external,-    labels_of,-    load_manifest,-    panel_genes,-    read_stage,-    target_n_cells,-    write_prediction,-)--GAMMA = -1.2          # damping of composition trend (X3 A-half: -1.2~-1.4 best)-OUT_FRAC = 0.95-N_FLOOR = 3000       # output size floor (X3 peak at 3000)-MIN_FRAC = 0.002     # floor for a type kept in the target composition-OVERLAP_MIN = 0.5    # min fraction of latest-stage cells whose type is also in the previous stage-HEART_W = 1.3        # multiplier on heart-lineage type fractions (proxy A-half: flat 1.0-1.3, 2.0 worse)-PARAXIAL_W = 0.1+from src.task1_temporal import view_io as vio -# Anatomical reweighting of the target composition. General (textbook) knowledge-# of E8.5->E9.5 mouse dissection: the embryo is dissected heart-centred at the-# later stage, so neural tube / surface ectoderm / paraxial mesoderm /-# extra-embryonic mesoderm shrink while heart-lineage tissues grow. Names are-# those of the *input* stages (published E8.5/E9.5 annotations); no held-out-# measurement is used. Unknown names (e.g. external datasets) get weight 1.0.+GAMMA = -1.2            # reversed composition-trend extrapolation (multi-stage, comparable vocab only)+N_FLOOR = 3000          # output cell-count floor+HEART_W = 1.3+PARAXIAL_W = 0.1+DROP_TYPES = {"Neural Tube", "Surface Ectoderm"} HEART_TYPES = {     "OFT/RV-CM", "IFT-CM", "AVC-CM", "SV-CM", "LV-CM", "RV-CM", "V-CM",     "Endothelium", "Endocardium", "BEC",     "aSHF", "pSHF", "aPHM", "pPHM",     "JCF", "Pericardium", "Proepicardium", }-ZERO_TYPES = {"Neural Tube", "Surface Ectoderm"}-PARAXIAL_TYPES = {"Paraxial Mesoderm"}+STRATIFY = False        # within-type expression-stratified sampling (PC1 axis)+N_HVG = 500+ALPHA_EXT = 0.0        # heart-lineage shift toward external later-stage pseudobulk direction (proxy2)+++def external_heart_delta(view, manifest, genes, base, labels, base_entry):+    """Direction (later external stage - official heart pseudobulk) on covered genes; median-centred."""+    ext = [e for e in vio.external_inputs(manifest) if e["time"] > base_entry["time"]]+    if not ext:+        return None+    e = ext[-1]+    a = vio.read_stage(view, e, genes, missing="zero")+    cov = a.var["covered"].to_numpy(dtype=bool)+    heart_mask = np.isin(labels, list(HEART_TYPES))+    if heart_mask.sum() < 50 or cov.sum() < 1000:+        return None+    m_ext = np.asarray(a.X.mean(axis=0), dtype=np.float64).ravel()+    m_off = np.asarray(base.X[heart_mask].mean(axis=0), dtype=np.float64).ravel()+    d = np.zeros(len(genes), dtype=np.float32)+    dd = (m_ext[cov] - m_off[cov]).astype(np.float32)+    d[cov] = dd - np.median(dd)+    return d, heart_mask   def type_weight(t: str) -> float:-    if t in ZERO_TYPES:+    if t in DROP_TYPES:         return 0.0-    if t in PARAXIAL_TYPES:+    if t == "Paraxial Mesoderm":         return PARAXIAL_W     if t in HEART_TYPES:         return HEART_W     return 1.0  -def apportion(fracs: np.ndarray, avail: np.ndarray, n_total: int) -> np.ndarray:-    """Largest-remainder allocation of n_total over fracs, capped at avail per type.--    Never asks for more cells than a type has (duplicated cells measurably hurt-    the DE-based groups); leftover quota goes to the types with the largest-    remaining fractional deficit that still have spare cells.-    """-    fracs = np.asarray(fracs, dtype=np.float64)-    avail = np.asarray(avail, dtype=np.int64)-    n_total = int(min(n_total, avail.sum()))-    raw = fracs * n_total-    n = np.minimum(np.floor(raw).astype(np.int64), avail)-    left = n_total - int(n.sum())-    while left > 0:-        deficit = np.where(n < avail, raw - n, -np.inf)-        order = np.argsort(-deficit)-        moved = False-        for j in order[:left]:-            if deficit[j] == -np.inf:-                break-            room = int(min(avail[j] - n[j], left))-            n[j] += room-            left -= room-            moved = True-        if not moved:-            break-    return n---def main() -> None:-    parser = argparse.ArgumentParser()-    parser.add_argument("--data", required=True)-    parser.add_argument("--out", required=True)-    parser.add_argument("--seed", type=int, default=0)-    args = parser.parse_args()--    manifest = load_manifest(args.data)-    genes = panel_genes(args.data, manifest)-    stages = inputs_by_time(manifest)--    official = [e for e in stages if not is_external(e)]-    pool_entry = official[-1] if official else stages[-1]-    pool = read_stage(args.data, pool_entry, genes)-    pool_labels = labels_of(pool)-    t_target = float(manifest["target"]["time"])-    t_last = float(pool_entry["time"])--    types, counts = np.unique(pool_labels, return_counts=True)-    fracs = counts.astype(np.float64) / counts.sum()--    # composition trend from the previous stage, only if vocabularies are comparable-    prev_entry = None-    for e in reversed(stages):-        if e is not pool_entry and e["time"] < t_last:-            prev_entry = e-            break-    if prev_entry is not None:-        prev = read_stage(args.data, prev_entry, genes)-        prev_labels = labels_of(prev)-        prev_types, prev_counts = np.unique(prev_labels, return_counts=True)-        prev_frac = dict(zip(prev_types.tolist(), (prev_counts / prev_counts.sum()).tolist()))-        shared = np.isin(types, prev_types)-        covered = fracs[shared].sum()-        dt = t_last - float(prev_entry["time"])-        if covered >= OVERLAP_MIN and dt > 0 and t_target > t_last:-            extrap = GAMMA * (t_target - t_last) / dt-            new = fracs.copy()-            for i, t in enumerate(types):-                if shared[i]:-                    new[i] = fracs[i] + extrap * (fracs[i] - prev_frac[t])-            new = np.clip(new, MIN_FRAC, None)-            fracs = new / new.sum()-        del prev--    # anatomical reweighting of the target composition-    w = np.array([type_weight(str(t)) for t in types], dtype=np.float64)-    if w.sum() > 0:-        adj = fracs * w-        fracs = adj / adj.sum()-+def largest_remainder(weights: np.ndarray, n: int) -> np.ndarray:+    weights = np.clip(np.asarray(weights, dtype=np.float64), 0, None)+    if n <= 0 or weights.sum() <= 0:+        return np.zeros(len(weights), dtype=int)+    raw = weights / weights.sum() * n+    out = np.floor(raw).astype(int)+    short = int(n - out.sum())+    for i in np.argsort(-(raw - out))[:short]:+        out[i] += 1+    return out+++def gamma_multipliers(manifest, view, genes, base_entry, labels):+    """(p_last/p_prev)^GAMMA per type when a comparable earlier stage exists, else 1."""+    entries = vio.inputs_by_time(manifest, include_external=False)+    if len(entries) < 2:+        entries = vio.inputs_by_time(manifest, include_external=True)+    if len(entries) < 2 or entries[-1]["path"] != base_entry["path"]:+        return None+    prev_entry = entries[-2]+    prev = vio.read_stage(view, prev_entry, genes)+    prev_labels = vio.labels_of(prev, prev_entry.get("celltype_col", "celltype"))+    a = set(np.unique(labels).tolist())+    b = set(np.unique(prev_labels).tolist())+    if len(a & b) < max(2, int(0.5 * min(len(a), len(b)))):+        return None  # vocabularies not comparable (e.g. proxy2 Qiu labels)+    pa = np.array([(prev_labels == t).mean() for t in sorted(a)])+    pl = np.array([(labels == t).mean() for t in sorted(a)])+    ratio = np.clip((pl + 1e-4) / (pa + 1e-4), 0.2, 5.0)+    return dict(zip(sorted(a), ratio ** GAMMA))+++def axis_scores(Xsub: np.ndarray, rng: np.random.Generator) -> np.ndarray:+    """1D diversity axis: PC1 of the (cells x HVG) block."""+    Z = Xsub - Xsub.mean(axis=0, keepdims=True)+    sd = Z.std(axis=0)+    if sd.max() <= 0:+        return np.zeros(Xsub.shape[0])+    _, _, vt = np.linalg.svd(Z, full_matrices=False)+    s = vt[0]+    proj = Z @ s+    if proj.sum() < 0:  # deterministic sign+        proj = -proj+    return proj+++def stratified_take(X: sparse.csr_matrix, pool: np.ndarray, k: int,+                    axis: np.ndarray, rng: np.random.Generator) -> np.ndarray:+    """Pick k cells from pool spread evenly along axis; if k>m reuse central cells."""+    m = pool.size+    if m == 0:+        return np.empty(0, dtype=np.int64)+    order = pool[np.argsort(axis[pool], kind="stable")]+    if k <= m:+        idx = np.round(np.linspace(0, m - 1, k)).astype(int)+        jitter = rng.integers(-1, 2, size=k)+        idx = np.clip(idx + jitter, 0, m - 1)+        return order[idx]+    # k > m: keep all, duplicate cells nearest the axis median (most average)+    extra = k - m+    med = np.median(axis[order])+    dup = np.argsort(np.abs(axis[order] - med), kind="stable")[:extra]+    return np.concatenate([order, order[dup]])+++def main():+    ap = argparse.ArgumentParser()+    ap.add_argument("--data", required=True)+    ap.add_argument("--out", required=True)+    ap.add_argument("--seed", type=int, default=0)+    args = ap.parse_args()+    view = args.data     rng = np.random.default_rng(args.seed)-    n_total = target_n_cells(manifest, pool.n_obs)-    n_total = max(int(n_total * OUT_FRAC), N_FLOOR, manifest["min_cells"])-    n_total = int(np.clip(n_total, manifest["min_cells"], manifest["max_cells"]))-    if n_total <= pool.n_obs:-        # enough real cells: no duplicates (capped apportion)-        alloc = apportion(fracs, counts, n_total)-    else:-        # pool smaller than target (e.g. X3): largest-remainder, allow duplication-        raw = fracs * n_total-        alloc = np.floor(raw).astype(np.int64)-        order = np.argsort(-(raw - alloc))-        for j in order[: int(n_total - alloc.sum())]:-            alloc[j] += 1--    rows = []-    for i, t in enumerate(types):-        idx = np.flatnonzero(pool_labels == t)-        n = int(alloc[i])-        if n > 0:-            # no duplicates while the pool is large enough; X3 (pool < target) reuses cells-            rows.append(rng.choice(idx, size=n, replace=n > len(idx)))-    rows = np.sort(np.concatenate(rows))-    X = pool.X[rows]-    write_prediction(X, genes, args.out, seed=args.seed)++    manifest = vio.load_manifest(view)+    genes = vio.panel_genes(view, manifest)++    official = vio.inputs_by_time(manifest, include_external=False)+    entries = official if official else vio.inputs_by_time(manifest, include_external=True)+    base_entry = entries[-1]+    base = vio.read_stage(view, base_entry, genes)+    labels = vio.labels_of(base, base_entry.get("celltype_col", "celltype"))+    X = base.X++    n_pool = base.n_obs+    n_out = int(np.clip(max(N_FLOOR, n_pool), manifest["min_cells"], manifest["max_cells"]))++    types = [t for t in sorted(set(labels.tolist())) if type_weight(t) > 0]+    counts = np.array([(labels == t).sum() for t in types], dtype=np.float64)+    weights = np.array([type_weight(t) for t in types]) * counts+    mult = gamma_multipliers(manifest, view, genes, base_entry, labels)+    if mult is not None:+        weights = weights * np.array([mult.get(t, 1.0) for t in types])+    alloc = largest_remainder(weights, n_out)++    # HVG block for the stratification axis (computed from the input stage only)+    block = None+    if STRATIFY:+        mean = np.asarray(X.mean(axis=0)).ravel()+        sq = np.asarray(X.multiply(X).mean(axis=0)).ravel()+        var = np.maximum(sq - mean ** 2, 0)+        hvg_cols = np.argsort(-var)[:N_HVG]+        block = np.asarray(X[:, hvg_cols].todense(), dtype=np.float64)++    delta_info = external_heart_delta(view, manifest, genes, base, labels, base_entry) if ALPHA_EXT > 0 else None++    blocks = []+    heart_flags = []+    for t, k in zip(types, alloc):+        k = int(k)+        if k <= 0:+            continue+        pool = np.flatnonzero(labels == t)+        if STRATIFY and k > 0:+            ax = axis_scores(block[pool], rng)+            full = np.zeros(base.n_obs)+            full[pool] = ax+            sel = stratified_take(X, pool, k, full, rng)+        else:+            sel = np.sort(rng.choice(pool, size=k, replace=pool.size < k))+        blk = X[sel]+        if delta_info is not None and t in HEART_TYPES:+            d = delta_info[0]+            dense = np.clip(np.asarray(blk.todense(), dtype=np.float32) + ALPHA_EXT * d, 0, None)+            blk = sparse.csr_matrix(dense.astype(np.float32))+        blocks.append(blk)+    out = sparse.vstack(blocks, format="csr").astype(np.float32)+    out.eliminate_zeros()+    vio.write_prediction(out, genes, args.out, seed=args.seed)   if __name__ == "__main__":

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**：父节点 12 目录缺失，按 ANALYSIS 逐项重建其机制（分层抽样复制 + 心脏解剖重加权 + GAMMA=-1.2 + N_FLOOR=3000），验证与父分数噪声内一致后，测试了 PLAN 的型内 PC1 分层抽样及多个替代方向。

**查分结果**（8 次，余 9）：
- 基线重建：proxy 57.64（父 57.09），X3 52.17（父 53.53）— 重建成功
- PLAN 核心（PC1 分层抽样）：proxy 56.65/57.05 双种子，de_recovery 恒 53.0 不动、cell_state 掉 2.8 — **gate 失败，弃用**
- HEART_W=1.6：56.1，更差；PARAXIAL_W=0：56.8，略差 — 均回退
- GAMMA=0 在 X3：47.3（-4.9）— 确认 GAMMA 强正贡献；-2.0 与 -1.2 差在噪声内，保留 -1.2
- proxy2 用 Qiu E9.0 伪批量方向平移心脏细胞（α=0.25）：41.8 灾难性失败（跨平台批次主导），已禁用

**最终提交** = 基线重建配置，三视图 vec-check 全过、确定性验证（md5 一致）、~5s/2GB。

**下一步最值得试**：X3 重复细胞策略（k-NN 去重或 N_FLOOR→4000，X3 covariation/de_recovery 最弱）；de_recovery 53.0 是"不改表达值"家族的天花板，突破需合法的方向性表达修正。

调研员

A or numpy SVD (fast, <1s total); (b) sort the m cells along this axis; (c) pick k evenly-spaced indices (round(linspace(0, m-1, k))), ensuring coverage from low to high expression states. This preserves within-type DE gradients without modifying any expression value. Step 3 (X3 duplication): when k > m (X3: pool 2174 < n 3000), use all m cells, then fill k-m slots by duplicating cells nearest to the type centroid on the diversity axis (most 'average' cells), minimizing distortion of DE statistics. Step 4 (fallback): proxy has one input stage → same mechanism applies (no temporal info needed). proxy2 ignores Qiu E9.0 as before. X3 uses its own input stages; GAMMA and reweighting unchanged. Step 5 (validation): run vec-score on proxy and X3 A-half first. Gate: de_recovery must improve ≥1.5 on proxy vs node 12 baseline (53.0) to proceed; if <1, revert. If PC1 axis underperforms, try alternative axis = mean z-score of top 50 within-type DE genes (computed from input only). Step 6: if gate passes, run proxy2 and X3, confirm all three rulers. Parameters to scan: number of HVGs for PC1 {200, 500, 1000}; for X3 duplication strategy {centroid-nearest vs random (current)}. Single-stage fallback: identical mechanism, no temporal component needed.",
  "expected_groups": ["de_recovery", "covariation"],
  "risks": "1) DE genes may be orthogonal to within-type PC1, making stratification ineffective — Engineer should check if de_recovery improves ≥1.5 on proxy A-half before proceeding; if not, try top-DE-gene axis. 2) Evenly-spaced sampling is deterministic and removes sampling stochasticity — seed sensitivity drops but may lose beneficial randomness; mitigate by adding small jitter (±1 index) to spaced picks. 3) Improvement may be <2-point noise — run both seed 0 and seed 1 on proxy A-half to confirm direction consistency. 4) X3 centroid-duplication might reduce effective diversity vs random duplication — compare both strategies on X3 A-half before committing.",
  "sources": []
}