总览 · ← 返回运行 20261002-202908-search-t1-scr-D

节点 n10

按 PLAN（family local_ot）在父节点 8 的 per-type EB 位移+PC 投影之上，加入同型两输入阶段间 entropic OT 耦合的 per-cell 零均值方向修正 κ·δ_j（δ 投影到同一 top-25 PC 子空间），κ=0.25，双 seed 约 +2.9。

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 8410309519: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 4a3449e6fc。改动的文件：solution/METHOD.md +32 −27、solution/README.md +5 −5、solution/run.py +201 −26

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex a60383b..01b01c7 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,46 +1,51 @@-按 PLAN 实现非 HVG 位移保留系数 γ（--proj-retain，默认 0.75）；实测 γ 对分数无影响（PLAN 风险 4 成立），改为在投影机制之上把位移幅度 α 从 1.5 调到 1.75（双 seed 一致 +0.3）。+按 PLAN（family local_ot）在父节点 8 的 per-type EB 位移+PC 投影之上，加入同型两输入阶段间 entropic OT 耦合的 per-cell 零均值方向修正 κ·δ_j（δ 投影到同一 top-25 PC 子空间），κ=0.25，双 seed 约 +2.9。 -## 方法（family: other / 父节点 5 + γ 保留 + α 重校准）+## 方法（family: local_ot，父节点 8 全部保留） -完整保留父节点 5：copy_last 抽样 + per-type EB 收缩位移（时间缩放 α·r，只作用非零元，max(0,·)）+ top-k PC 低秩协变投影（k=25, β=1.0, HVG 8000, ≤3000 细胞, 中心化, 独立 rng 流）。单输入视图严格退化 copy_last。+基座不变：copy_last 抽样 + per-type EB 收缩位移（时间缩放 α·r，α=1.75，只作用非零元，max(0,·)）+ top-25 PC 低秩协变投影（HVG 8000、β=1.0、γ=0.75）。单输入视图严格退化 copy_last。 -本节点改动两处：+新机制（本节点）：对每个在两个输入阶段都有 ≥10 个细胞的类型（不足或无源类型 δ=0），在 top-8000 HVG 空间做同型 entropic OT： -1. **γ 保留（PLAN 指定机制）**：投影混合改为 HVG 基因 Δ_final=(1−β)Δ+βV(VᵀΔ)，非 HVG 基因 Δ_final=γ·Δ（父节点 β=1 时为 0）。参数 `--proj-retain`（默认 0.75，环境变量 VEC_PROJ_RETAIN）。γ=0 对照与父节点输出**逐元素一致**（diff nnz=0，已验证）。-2. **α 重校准**：默认 `--alpha` 1.5 → 1.75。依据：γ 对分数完全平坦后，α 扫描（seed0：1.0/1.5/1.75/1.9/2.0/2.5/3.0 = 52.50/53.48/**53.83**/53.87/53.73/53.39/52.34）显示峰值在 1.75–1.9，α=1.75 双 seed 确认。+1. 平方欧氏代价（按均值归一化以稳定数值），`ot.sinkhorn(reg=0.05, numItermax=2000, stopThr=1e-9)`，均匀边际，每侧最多 3000 细胞（独立 rng 流 `[seed, 0x0712]`/`[seed, 0x0710]`，不扰动父节点的抽样与 PCA 流）；+2. 每个 stage-2 细胞 j 的重心位移 d_j = x_j − Σ_i π(i|j)x_i，减去型内均值得零均值修正 δ_j（型级位移不变，只加型内方向散布）；+3. δ_j 投影到与型级位移**同一个** V（top-25 PC）：C_j = (1−β)δ_j + β·V(Vᵀδ_j)（β=1 时即纯投影；线性性保证等价于"先混合 Δ_c+κδ 再投影"，per-cell 只存 25 维系数的等价形式）；+4. 输出细胞 i 的最终位移 = 父节点 df（型级）+ κ·α·r·C_j，仍只作用非零元。++参数：`--ot-kappa`（默认 0.25，env VEC_OT_KAPPA）、`--ot-reg 0.05`、`--ot-min 10`、`--ot-cap 3000`、`--ot-proj 1`、`--ot-proj-k 0`（0=复用型级 V；>0 用独立更大子空间，已试且更差）。  ## 机制生效证据 -- γ 机制确实改变输出：γ=0.75 相对父节点 diff nnz=91935；每型 diag 中 nonhvg_ratio 恰为 0.750（hvg_ratio 0.29–0.90 不变）。-- **但 γ 对分数无影响**（X3 A 半 seed0）：γ=0 → 53.48，γ=0.75 → 53.478，γ=1.0 → 53.474，hvg12000 → 53.474。四组分几乎不动（cell_state 61.76–61.81，covariation 47.60–47.65）。PLAN 风险 4 成立：非 HVG 基因的位移量（EB 收缩后本就很小）对评分贡献可忽略。保留 γ=0.75 作默认（无害，且比归零更接近父节点 4 的行为）。-- α 机制生效：α↑ → cell_state↑、covariation↓（α=2.0：63.14/46.58），α=1.75 处总分最优。+- **关闭对照**：κ=0（--ot-kappa 0）输出与父节点 8 **逐元素一致**（diff nnz=0，重构前后各验证一次）。+- δ 非退化：≥10 细胞的 7 个类型 mean‖δ‖/‖mean d‖ = 2.9–9.9，mean‖C‖ ≈ 11.3–12.8（OT diag 打印在 stderr）；太小的型（BEC 3 细胞、aSHF 源 7 细胞、pSHF/NCC-derived/ST 无源）按设计跳过，δ=0。+- 四组分变化（X3 A 半 seed0，κ=0 → κ=0.25）：cell_state 62.73 → **75.59**（+12.9，主要增益：型内方向散布恢复了目标群体的细胞级多样性）；covariation 47.16 → 43.56（−3.6，per-cell 修正部分破坏了型级低秩结构）；direction 50.86 → 50.49（−0.4，噪声内）；de_recovery 51.46 → 51.46（0）。PLAN 预期 direction 升，实际 direction 平、cell_state 大涨——增益来源与预期不同，如实记录。+- δ 必须投影：`--ot-proj 0`（原始 δ 直接加）灾难性恶化（κ=0.15 → 38.4，κ=0.3 → 30.1），确认低秩投影是该机制可用的前提；独立更大子空间也更差（ot-proj-k=50 → 55.93，100 → 54.07；=10 → 56.82，与复用 k=25 的 56.88 噪声内持平）。 -## 查分记录（X3 A 半，共 8 次 / 额度 20）+## 查分记录（X3 A 半，共用 17/20 次） -| 配置 | seed0 | seed1 | 均值 |-|---|---|---|---|-| 父节点 5（γ=0, α=1.5） | 53.48 | 54.40 | 53.94 |-| γ=0.75, α=1.5 | 53.478 | – | – |-| γ=1.0, α=1.5 | 53.474 | – | – |-| hvg12000, γ=0.75 | 53.474 | – | – |-| α=2.0, γ=0.75 | 53.73 | 54.63 | 54.18 |-| α=2.5 / 3.0 | 53.39 / 52.34 | – | – |-| α=1.9 | 53.87 | – | – |-| **α=1.75（提交）** | **53.83** | **54.67** | **54.25** |-| α=1.75, k=40 → α=2.0 k40 | 53.42 | – | – |+| 配置 | seed0 | seed1 |+|---|---|---|+| 父节点 8（κ=0，α=1.75） | 53.83 | 54.67（父节点记录） |+| κ=0.1 / 0.2 | 55.68 / 56.63 | – |+| **κ=0.25（提交默认）** | **56.88** | **57.31** |+| κ=0.3 / 0.35 / 0.4 | 56.71 / 56.14 / 55.46 | 56.99 / – / – |+| κ=0.5 / 0.7 | 53.59 / 49.06 | – |+| κ=0.25, ot-proj 0 | 30.1（κ=0.3）/ 38.4（κ=0.15） | – |+| κ=0.25, ot-proj-k 10/50/100 | 56.82 / 55.93 / 54.07 | – | -**诚实评估**：提交配置相对父节点 +0.35（seed0）/ +0.27（seed1），均值 +0.31，**低于 2 分噪声线**，双 seed 方向一致但幅度小。α 峰值 1.75–1.9 平台平坦（53.83–53.87），是在 X3 A 半上的局部校准，对 final（r=1，有效尺度不同）不一定最优。+κ=0.25 双 seed 均值 +2.87（seed0 +3.05，seed1 +2.64 vs 父记录），在 2 分噪声线之上；κ 平台 0.2–0.3 平坦（56.6–56.9），取 0.25。**诚实评估**：cov 组分净损 −3.6 被 cell_state +12.9 覆盖；该权衡是 X3 A 半上的实测，B 半与 final（r=1，OT 只跨 1 天而非 X3 的 0.25 天，δ 结构可能不同）不一定同幅度。  ## 验证过 / 没验证 -- 验证：γ=0 对照 == 父节点逐元素；默认参数输出 == 已查分 a175 预测逐元素；seed0 双跑确定；X3 vec-check ok；运行 ~5s、内存远低于限制；纯 CPU（EXECUTION.json gpu:false）。-- 视图无关：只改常数默认值，逻辑仍只用相对时间差与视图数据；独立 rng 流不变。-- 没验证：final / proxy / proxy2 视图（本地只有 X3 视图；单输入退化路径代码未动，父节点已验证）；B 半真值；α=1.75 在 r=1 视图上的最优性。+- 验证：κ=0 对照==父节点逐元素（重构前后各一次）；默认参数输出==已查分 κ=0.25 预测逐元素；seed0 双跑确定（diff=0）；X3 vec-check ok；运行 ~13s、内存 <1GB（corr 只存输出细胞×HVG）；纯 CPU。+- 视图无关：只用视图内表达/标签与相对时间差（r、α·r），OT 代价与耦合全部现场计算；独立 rng 流按 seed 派生；无任何视图/路径/绝对时间分支。+- 没验证：final / proxy / proxy2 视图（本地只有 X3；单输入退化路径未动，κ≠0 时 OT 也只在 ≥2 输入分支内）；B 半真值；reg=0.1（PLAN 备用，reg=0.05 无收敛问题故未试）；δ 的 L2 归一化（PLAN 风险 3，未需要）。  ## 知识来源 -全部沿用父节点（标签改名词表来自方法卡 §标签；EB 收缩为标准经验贝叶斯；PCA 只用视图内输入阶段表达数据）。无外部数据、无 prior/、无禁窗信息、无硬编码统计量。+沿用父节点：SYNONYM_PARENTS 改名词表来自方法卡 §标签（通用谱系知识）；EB 收缩为标准经验贝叶斯；PCA/HVG/OT 代价只用视图内输入阶段数据。无外部数据（external/ 未读）、无 prior/、无禁窗信息、无硬编码统计量。POT (`ot`) 为环境已有依赖。  ## 下一步建议 -γ/非 HVG 方向已证伪（分数平坦）。方向性增益需换机制：低秩投影 + OT 耦合组合（ANALYSIS 建议 2），或在 α 平台上做 per-type 幅度自适应（按 ||Δ̃||/||Δ|| 反比缩放，弱投影型多保留）。+- cov 组分是当前最弱（43.6）：试"型级投影保留、per-cell 修正后整体再做一次轻投影"或对 C_j 做幅度截断（按型内分位数 winsorize），或把 κ·C 的加入放在 max(0,·) 之前/之后对比。+- direction 停在 ~50.5：δ 的方向信息可能被 25 维子空间截断，但更大子空间实测更差——方向改进可能需要非 OT 机制（如基因级时间趋势）。+- α 与 κ 存在耦合（α·r 同乘两者），final r=1 时最优点可能漂移，终选前应在 r=1 形态的视图上复核 κ。diff --git a/solution/README.md b/solution/README.mdindex 622d58b..0513559 100644--- a/solution/README.md+++ b/solution/README.md@@ -1,7 +1,7 @@-# per-type EB displacement + low-rank covariation projection+# per-type EB displacement + PC projection + within-type OT per-cell correction -父节点 4（copy_last + 按 celltype 的 EB 收缩位移，只作用非零元）之上加一步：-每型位移向量投影到输入数据自身协变结构的 top-k PC 子空间（合并两输入阶段 ≤3000 细胞、-top-8000 HVG、中心化、svds k=25），Δ_final=(1−β)Δ+βΔ̃，β=1.0。-修复父节点 covariation 净损（45.3→47.7）。`--proj-beta 0` 为机制关闭对照（输出与父节点逐元素一致）。+父节点 8（copy_last + 按 celltype 的 EB 收缩位移 + top-25 PC 低秩投影）之上：+同型两输入阶段间 entropic OT（POT sinkhorn, reg=0.05, HVG 8000）给出每个 stage-2+细胞的零均值重心位移偏差 δ_j，投影到同一 PC 子空间后以 κ·α·r 加到该细胞的位移上+（--ot-kappa，默认 0.25）。κ=0 为机制关闭对照（输出与父节点逐元素一致）。 单输入视图退化 copy_last。详见 METHOD.md。diff --git a/solution/run.py b/solution/run.pyindex 064ffe9..cf10861 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,24 +1,29 @@ #!/usr/bin/env python3-"""Per-type EB-shrunk displacement + low-rank covariation projection.--Base (parent node 4): copy_last + per-type gene-mean shift between the two-latest input stages, gene-level EB shrinkage, time rescale alpha*r, applied at-existing nonzeros only, max(0, .).--New mechanism (this node): each type's shrunk displacement vector-Delta_c = alpha*r*lambda*delta_c is projected onto the top-k principal-directions V of the input data's own gene covariance (cells of both input-stages, centered, top-HVG subset), and mixed back:-    Delta_final = (1-beta)*Delta_c + beta * V (V^T Delta_c)   [HVG genes]-    Delta_final = gamma * Delta_c                             [non-HVG genes]-with gamma = --proj-retain (gamma=0 reproduces the parent element-wise).-This makes the displacement move along coordinated co-expression programs-instead of perturbing every gene independently, targeting the covariation-regression of the parent. beta=0 or k=0 bypasses the projection entirely and-reproduces the parent element-wise.--View independence: V is computed from the view's own input matrices; only-relative time differences are used. Single input stage -> exact copy_last.+"""Per-type EB displacement + PC projection + within-type OT per-cell correction.++Base (parent node 8): copy_last + per-type gene-mean shift between the two+latest input stages, gene-level EB shrinkage, time rescale alpha*r, low-rank+projection onto the top-k PCs of the input data's own covariance (HVG subset),+non-HVG retain gamma, applied at existing nonzeros only, max(0, .).++New mechanism (this node, family local_ot): for each cell type present in both+input stages (>= --ot-min cells each side), an entropic OT coupling+(ot.sinkhorn, reg --ot-reg on mean-normalized squared-Euclidean cost in the+same top-HVG space as the PC projection) matches stage-1 cells to stage-2+cells of the same type. Each stage-2 cell j gets a barycentric displacement+d_j = x_j - sum_i pi(i|j) x_i; the within-type zero-mean deviation+delta_j = d_j - mean_c(d) is a per-cell directional correction:+    Delta_j[hvg]  = (1-beta)*(Delta_c + kappa*delta_j)[hvg]+                    + beta * V V^T (Delta_c[hvg] + kappa*delta_j[hvg])+    Delta_j[rest] = gamma * Delta_c[rest]+kappa = --ot-kappa (0 -> element-wise identical to parent node 8, the+mechanism-off control). Since V V^T is linear, the per-cell term is+kappa * ((1-beta)*delta_j + beta * V z_j) with z_j = V^T delta_j precomputed+per stage-2 cell.++View independence: everything (EB deltas, PC basis, HVG set, OT couplings) is+computed from the view's own input matrices; only relative time differences+are used. Single input stage -> exact copy_last. """  from __future__ import annotations@@ -97,6 +102,108 @@ def covariation_basis(Xa: sparse.csr_matrix, Xb: sparse.csr_matrix,     return hvg, V  +def hvg_only(Xa: sparse.csr_matrix, Xb: sparse.csr_matrix,+             n_hvg: int, n_cells: int, rng: np.random.Generator) -> np.ndarray:+    """Variance-based top-HVG indices over the pooled input cells (no SVD)."""+    X = sparse.vstack([Xa.tocsr(), Xb.tocsr()], format="csr")+    n = X.shape[0]+    if n_cells < n:+        sel = rng.choice(n, size=n_cells, replace=False)+        sel.sort()+        X = X[sel]+    m1 = np.asarray(X.mean(axis=0), dtype=np.float64).ravel()+    m2 = np.asarray(X.multiply(X).mean(axis=0), dtype=np.float64).ravel()+    var = np.maximum(m2 - m1 * m1, 0.0)+    g = int(min(n_hvg, X.shape[1]))+    hvg = np.argpartition(-var, g - 1)[:g]+    hvg.sort()+    return hvg+++def ot_cell_corrections(X1: sparse.csr_matrix, lab1: np.ndarray,+                        X2: sparse.csr_matrix, lab2: np.ndarray,+                        hvg: np.ndarray, V: np.ndarray | None,+                        beta: float, proj_delta: bool, out_pos: np.ndarray,+                        reg: float, min_cells: int, cap: int,+                        rng: np.random.Generator, Vk: np.ndarray | None = None):+    """Within-type entropic OT barycentric displacements, returned as per-cell+    correction data for the OUTPUT cells (rows of the sampled stage-2 cells).++    Returns (corr, diag): corr is a dense (n_out, len(hvg)) float32 array+    holding the per-cell correction ((1-beta)*delta + beta*V V^T delta if+    proj_delta else raw delta), and diag a list of per-type diagnostics. Types+    below the cell threshold (or missing on the stage-1 side) stay zero.+    out_pos maps stage-2 cell index -> output row (or -1).+    """+    import ot as pot++    corr = np.zeros((int((out_pos >= 0).sum()), len(hvg)), dtype=np.float32)+    diag = []+    types2 = [str(t) for t in np.unique(lab2)]+    for t in sorted(types2):+        m2 = np.flatnonzero(lab2 == t)+        # stage-1 cells of the same type (or synonym parents), else skip+        if np.any(lab1 == t):+            m1 = np.flatnonzero(lab1 == t)+        else:+            parents = [p for p in SYNONYM_PARENTS.get(t, ()) if np.any(lab1 == p)]+            if not parents:+                diag.append((t, len(m2), 0, 0.0, 0.0, "no-source"))+                continue+            m1 = np.concatenate([np.flatnonzero(lab1 == p) for p in parents])+        if len(m1) < min_cells or len(m2) < min_cells:+            diag.append((t, len(m2), len(m1), 0.0, 0.0, "too-few"))+            continue+        if len(m1) > cap:+            m1 = np.sort(rng.choice(m1, size=cap, replace=False))+        if len(m2) > cap:+            sel2 = np.sort(rng.choice(m2, size=cap, replace=False))+        else:+            sel2 = m2+        A = np.asarray(X1[np.ix_(m1, hvg)].todense(), dtype=np.float64)+        B = np.asarray(X2[np.ix_(sel2, hvg)].todense(), dtype=np.float64)+        # squared euclidean cost, mean-normalized for numerical stability+        sa = (A * A).sum(axis=1)+        M = sa[:, None] + (B * B).sum(axis=1)[None, :] - 2.0 * (A @ B.T)+        np.maximum(M, 0.0, out=M)+        mu = M.mean()+        if not np.isfinite(mu) or mu <= 0:+            diag.append((t, len(m2), len(m1), 0.0, 0.0, "degenerate-cost"))+            continue+        M /= mu+        a = np.full(len(m1), 1.0 / len(m1))+        b = np.full(len(sel2), 1.0 / len(sel2))+        try:+            pi = pot.sinkhorn(a, b, M, reg, numItermax=2000, stopThr=1e-9)+        except Exception as e:  # convergence failure -> skip type (kappa corr 0)+            diag.append((t, len(m2), len(m1), 0.0, 0.0, f"sinkhorn:{type(e).__name__}"))+            continue+        if not np.all(np.isfinite(pi)):+            diag.append((t, len(m2), len(m1), 0.0, 0.0, "sinkhorn-nan"))+            continue+        # barycentric displacement d_j = B_j - sum_i pi(i|j) A_i+        colsum = pi.sum(axis=0)+        colsum[colsum <= 0] = 1.0+        W = (pi / colsum[None, :]).T  # (n_sel2, n_m1), rows sum to 1+        Dsp = B - W @ A+        Dmean = Dsp.mean(axis=0, keepdims=True)+        Delta = Dsp - Dmean  # zero-mean within type+        Vd = Vk if Vk is not None else V+        if proj_delta and Vd is not None:+            Z = Delta @ Vd+            C = (1.0 - beta) * Delta + beta * (Z @ Vd.T)+        else:+            C = Delta+        op = out_pos[sel2]+        keep = op >= 0+        corr[op[keep]] = C[keep].astype(np.float32)+        nd = float(np.linalg.norm(Dmean))+        ratio = float(np.mean(np.linalg.norm(Delta, axis=1))) / max(nd, 1e-12)+        diag.append((t, len(m2), len(m1), ratio, float(np.mean(np.linalg.norm(C, axis=1))),+                     "ok" if len(sel2) == len(m2) else f"capped({len(sel2)})"))+    return corr, diag++ def main() -> None:     parser = argparse.ArgumentParser()     parser.add_argument("--data", required=True)@@ -118,6 +225,20 @@ def main() -> None:                         default=os.environ.get("VEC_PROJ_NORM", "") == "1")     parser.add_argument("--no-time-scale", action="store_true",                         default=os.environ.get("VEC_NO_TIME_SCALE", "") == "1")+    parser.add_argument("--ot-kappa", type=float,+                        default=float(os.environ.get("VEC_OT_KAPPA", "0.25")))+    parser.add_argument("--ot-reg", type=float,+                        default=float(os.environ.get("VEC_OT_REG", "0.05")))+    parser.add_argument("--ot-min", type=int,+                        default=int(os.environ.get("VEC_OT_MIN", "10")))+    parser.add_argument("--ot-cap", type=int,+                        default=int(os.environ.get("VEC_OT_CAP", "3000")))+    parser.add_argument("--ot-proj", type=int,+                        default=int(os.environ.get("VEC_OT_PROJ", "1")))+    parser.add_argument("--ot-proj-k", type=int,+                        default=int(os.environ.get("VEC_OT_PROJ_K", "0")),+                        help="0: project delta onto the same V as the type "+                             "shift; >0: separate PC basis with this many PCs")     args = parser.parse_args()      manifest = load_manifest(args.data)@@ -187,8 +308,52 @@ def main() -> None:                                        args.proj_cells, rng_proj)          beta = float(args.proj_beta)++        # within-type OT per-cell corrections (own rng stream; kappa=0 skips+        # entirely so the output stays element-wise identical to the parent)+        kappa = float(args.ot_kappa)+        corr = None+        hvg_pos = None+        if kappa != 0.0:+            rng_ot = np.random.default_rng(np.array([args.seed, 0x0710], dtype=np.int64))+            out_pos = np.full(last.n_obs, -1, dtype=np.int64)+            out_pos[rows] = np.arange(len(rows))+            proj_delta = bool(args.ot_proj)+            if hvg is None:+                hvg = hvg_only(prev.X.tocsr(), last.X.tocsr(),+                               args.proj_hvg, args.proj_cells, rng_ot)+                V = None+            Vk = None+            if args.ot_proj_k > 0 and hvg is not None:+                # separate, larger PC basis (same HVG subset) for the delta+                # projection; own rng stream, does not disturb other draws+                rng_vk = np.random.default_rng(np.array([args.seed, 0x0712],+                                                        dtype=np.int64))+                Xp = sparse.vstack([prev.X.tocsr(), last.X.tocsr()], format="csr")+                n = Xp.shape[0]+                if args.proj_cells < n:+                    selv = rng_vk.choice(n, size=args.proj_cells, replace=False)+                    selv.sort()+                    Xp = Xp[selv]+                Dp = np.asarray(Xp[:, hvg].todense(), dtype=np.float64)+                Dp -= Dp.mean(axis=0, keepdims=True)+                kk = int(min(args.ot_proj_k, min(Dp.shape) - 1))+                v0 = rng_vk.standard_normal(min(Dp.shape))+                _, _, Vt = svds(Dp, k=kk, v0=v0)+                Vk = np.ascontiguousarray(Vt[::-1].T)+            corr, ot_diag = ot_cell_corrections(+                prev.X.tocsr(), lab1, last.X.tocsr(), lab2, hvg, V,+                beta, proj_delta, out_pos,+                args.ot_reg, args.ot_min, args.ot_cap, rng_ot, Vk)+            hvg_pos = np.full(len(genes), -1, dtype=np.int64)+            hvg_pos[hvg] = np.arange(len(hvg))+            print("OT diag (type, n2, n1, mean||delta||/||mean d||, mean||C||, status):",+                  file=sys.stderr)+            for row in ot_diag:+                print("  %-16s %4d %4d  %.3f  %.3f  %s" % row, file=sys.stderr)+         inhvg = None-        if use_proj:+        if hvg is not None:             inhvg = np.zeros(len(genes), dtype=bool)             inhvg[hvg] = True         diag = []@@ -196,29 +361,39 @@ def main() -> None:         labs = lab2[rows]         for t in np.unique(labs):             d = scale * delta_for(str(t))+            df = None             if use_proj:                 dv = d[hvg]                 tilde = V @ (V.T @ dv)                 nd, nt = float(np.linalg.norm(dv)), float(np.linalg.norm(tilde))                 if args.proj_norm and nt > 0:                     tilde *= nd / nt-                corr = float(np.corrcoef(dv, tilde)[0, 1]) if nd > 0 and nt > 0 else 0.0-                diag.append((str(t), int((labs == t).sum()), nt / max(nd, 1e-12), corr))+                corr_t = float(np.corrcoef(dv, tilde)[0, 1]) if nd > 0 and nt > 0 else 0.0+                diag.append((str(t), int((labs == t).sum()), nt / max(nd, 1e-12), corr_t))                 gamma = float(args.proj_retain)                 df = gamma * d                 df[hvg] = (1.0 - beta) * d[hvg] + beta * tilde                 nh, nn = float(np.linalg.norm(df[hvg])), float(np.linalg.norm(df[~inhvg]))                 d0h = float(np.linalg.norm(d[hvg]))                 d0n = float(np.linalg.norm(d[~inhvg]))-                d = df                 diag[-1] = diag[-1] + (nh / max(d0h, 1e-12), nn / max(d0n, 1e-12))-            d = d.astype(np.float32)+            else:+                df = d+            df = df.astype(np.float32)             tmask = np.flatnonzero(labs == t)             for i in tmask:                 lo, hi = Xs.indptr[i], Xs.indptr[i + 1]                 if hi > lo:                     cols = Xs.indices[lo:hi]-                    Xs.data[lo:hi] = np.maximum(Xs.data[lo:hi] + d[cols], 0.0)+                    dj = df[cols]+                    if corr is not None:+                        pos = hvg_pos[cols]+                        ok = pos >= 0+                        if ok.any():+                            dj = dj.copy()+                            dj[ok] += (kappa * scale+                                       * corr[i, pos[ok]].astype(np.float64))+                    Xs.data[lo:hi] = np.maximum(Xs.data[lo:hi] + dj, 0.0)         X = Xs         if use_proj and diag:             print("proj diag (type, n, ||proj||/||d||, corr, hvg_ratio, nonhvg_ratio):",

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

Done. Final state verified: `run.py` default κ=0.25 reproduces the scored prediction element-wise, deterministic across reruns, κ=0 control identical to parent, vec-check ok, runtime ~13s.

**做了什么**：按 PLAN（local_ot）在父节点 8 之上实现同型两输入阶段间的 entropic OT（POT sinkhorn, reg=0.05, HVG8000, ≥10 细胞/型），取每细胞重心位移的型内零均值偏差 δ_j，投影到同一 top-25 PC 子空间后以 κ·α·r 加入 per-cell 位移；κ=0 关闭对照与父节点逐元素一致。

**查分结果**（X3 A 半，用 17/20）：κ=0 基线 53.83（seed0）；κ=0.25 → **56.88（seed0）/ 57.31（seed1）**，双 seed +2.9，超噪声线。增益几乎全部来自 cell_state（62.7→75.6），covariation 净损 −3.6，direction 持平。δ 不经 PC 投影直接加会灾难性崩（30–38 分），更大投影子空间（k=50/100）也更差。

**下一步最值得试**：修复 cov 净损——对 C_j 做型内分位数截断（winsorize）或加入后再做一次轻投影；以及在 r=1 形态视图上复核 κ（α·r 同乘两者，final 最优点可能漂移）。

调研员

irection drops. Mitigate: κ≤0.5, check per-type coupling entropy; if direction drops >1 pt at κ=0.3, try κ=0.1. 2) Within-type OT with very few cells (<20) is degenerate → skip types below threshold. 3) δ_j might be dominated by a few outlier genes → L2-normalize δ_j before blending. 4) Total improvement may be <2 pt noise; Engineer should require seed-0 gain ≥1 pt before spending seed-1 confirmation. 5) POT import or convergence failure → fallback: skip OT, submit parent config (κ=0).",
  "family_id": "local_ot",
  "mechanism": "Within-type entropic OT coupling between the two input stages yields per-cell barycentric displacements; the deviation of each cell's OT displacement from its type mean provides a zero-mean directional correction that breaks the within-type constant-direction limitation of the per-type EB shift.",
  "vs_constant_shift": "Constant per-type shift gives every cell of a type the same Δ_c vector. Here, OT coupling produces a per-cell correction δ_j = d_j^OT − mean(d^OT) that varies by cell position in the coupling, so two cells of the same type with different neighbourhoods in the source stage receive different directional nudges. The type-mean component (Δ_c) is unchanged; only the within-type spread of directions is added.",
  "mechanism_evidence": "1) Per-type std of ||δ_j|| should be >0 for types with ≥10 cells (if 0, OT correction is degenerate). 2) Compare direction group score at κ=0 vs κ=0.3: direction should increase while cell_state stays within ±1. 3) Report per-type mean ||δ_j||/||Δ_c|| ratio to verify correction magnitude is meaningful but not dominant (target 0.1–0.5). 4) Covariation should stay ≥47.5 (PC projection still active). 5) Four-group breakdown at best κ vs κ=0.",
  "mechanism_off_control": "Set κ=0 (--ot-kappa 0 / VEC_OT_KAPPA=0): the OT correction term vanishes and the output must be element-wise identical to parent node 8 (diff nnz = 0). Engineer should verify this before scoring any κ>0 config.",
  "sources": []
}
```