总览 · ← 返回运行 20261003-172000-search-t2-heart-extrap-chain-12h

节点 n31 在终选来历上

几何均值总量比均匀线性缩放替代逐细胞线性总量回拉（γ=0.45）：log 域近似恒定每细胞偏移，保住位移的跨基因 dp 排序；PLAN 的速度表达水平回归经 6 配置查分证否后按备选机制路径提交。

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 673b91fb60: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 fc5fc15878。改动的文件：solution/METHOD.md +94 −56、solution/README.md +18 −6、solution/run.py +279 −29

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 388286b..287ec2b 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,62 +1,100 @@-坐标15-NN两步支撑掩码扩散平滑（ε=0.4，保零模式、逐基因列缩放精确复原伪批量）移植为 rel 速度乘法外推的底物，γ 重耦合扫至 0.3。--## 方法（family T2HX-01，节点 27 → 29）--在节点 27 的按型 rel 速度乘法外推（v_t=(m_a−m_p)/(m_p+1)、软阈值 k=0.5、双硬门控、PBC=10、坐标/行序/细胞数/组成冻结）之上，加**唯一结构改动**：位移前对锚阶段表达矩阵做空间平滑（复刻节点 23/26 配方）：--1. 在锚坐标（输出的同一份坐标）上建 15-NN 图（cKDTree，去自身）；-2. 两步扩散 `Xc ← (1−ε)·Xc + ε·mean(15 邻居)`，ε=0.4；**支撑掩码**：原矩阵为 0 的位置逐步强制回 0（零模式逐位保留，只有实测正值移动）；-3. 逐基因单列缩放 `c_g = pb_orig_g / pb_smooth_g`，精确复原全局伪批量（实测最大相对差 0.0 < 1e-6）。--速度估计器、门控的 dp/dt_approx 都用**未平滑的原始**两阶段矩阵（保伪批量，使 DE/mmd 证据与节点 27 直接可比）；位移作用于平滑后底物 `x′ = x_smoothed·2^(α·v̂_t)`；γ 再闭合拉向平滑后细胞自身线性总量；PBC 界 = 10 × median(**原始**输入总量)，全部现场计算。回退：单输入/锚≠末输入 → 平滑与速度一并关闭，逐位 copy_last（同父节点）。--提交配置：steps=2, ε=0.4, k_NN=15, α=1, k_soft=0.5, γ=0.3, PBC=10, MIXMODE=rel, MIXREL=1, MIXEPS=1。--## 机制生效证据（PLAN mechanism_evidence 逐条）--- (a) 平滑生效：平滑前后逐基因伪批量最大相对差 0.0（<1e-6）；零模式逐位不变（数组比较 True）；坐标 15-NN 图上 Dirichlet 能量 35.34 → 30.82（−12.8%）；100% 正值条目被改动；列缩放中位 2.36×。-- (b) 目标指标（proxy_noscale A 半，提交配置）：**neighborhood_mmd raw 0.07896 ≤ 0.0800**（节点 27 为 0.08123，超过节点 26 的 0.07958，全树最好）；**variogram raw 0.03872 ≤ 0.0388**（节点 27 为 0.03877，rel 速度下唯一变差的指标被修复）。-- (c) 保持项：**de_direction 0.2864 ≥ 0.26**（父 0.2776）；**mmd_u 0.04763 ≤ 0.0530**（父 0.05158，全树最好）；de_score 0.1806（父 0.1528）。-- (d) 组分：local_spatial 59.20（父 58.12，+1.08）、cell_state 57.74（父 56.57，+1.17）、expression_change 56.85（父 56.25，+0.60，不降反升）、shape_scale 50.00 不变（坐标冻结）。榜分 55.95（父 55.23）。-- (e) 平台：γ∈{0, 0.2, 0.3, 0.4} → 56.00/55.85/55.95/55.85（平坦，全部高于父）；ε∈{0.3, 0.4}（γ=0.5）→ 55.82/55.78；提交点取 γ 平台内部、ε 平台内部，非刀尖。--## 查分记录（16 次，全部 proxy_noscale，A 半）--| ε | γ | k_soft | 榜分 | nb_mmd | variogram | mmd_u | de_dir | de_score |-|---|---|---|---:|---:|---:|---:|---:|---:|-| – (父节点27) | 0.75 | 0.5 | 55.23 | 0.08123 | 0.03877 | 0.05158 | 0.2776 | 0.1528 |-| 0.6 | 0.75 | 0.5 | 54.74 | 0.08766 | 0.03613 | 0.05768 | 0.2571 | 0.1667 |-| 0.6 | 0.5 | 0.5 | 55.06 | 0.08571 | 0.03630 | 0.05521 | 0.2564 | 0.1806 |-| 0.6 | 0.6 | 0.5 | 54.92 | 0.08631 | 0.03624 | 0.05598 | 0.2571 | 0.1667 |-| 0.6 | 0.9 | 0.5 | 54.51 | 0.08990 | 0.03596 | 0.06035 | 0.2543 | 0.1806 |-| 0.4 | 0.5 | 0.5 | 55.78 | 0.07959 | 0.03823 | 0.04911 | 0.2855 | 0.1667 |-| 0.4 | 0.75 | 0.5 | 55.46 | 0.08184 | 0.03730 | 0.05274 | 0.2831 | 0.1667 |-| 0.2 | 0.5 | 0.5 | 55.52 | 0.07959 | 0.04046 | 0.04856 | 0.2810 | 0.1389 |-| 0.3 | 0.5 | 0.5 | 55.82 | 0.07904 | 0.03938 | 0.04852 | 0.2836 | 0.1806 |-| 0.5 | 0.5 | 0.5 | 55.49 | 0.08158 | 0.03710 | 0.05094 | 0.2688 | 0.1528 |-| 0.4 | 0.4 | 0.5 | 55.85 | 0.07920 | 0.03849 | 0.04827 | 0.2860 | 0.1667 |-| 0.4 | 0.5 | 0.7 | 55.51 | 0.08150 | 0.03751 | 0.05219 | 0.2852 | 0.1667 |-| **0.4** | **0.3** | **0.5** | **55.95** | **0.07896** | **0.03872** | **0.04763** | **0.2864** | **0.1806** |-| 0.4 | 0.2 | 0.5 | 55.85 | 0.07882 | 0.03894 | 0.04713 | 0.2845 | 0.1528 |-| 0.4 | 0.0 | 0.5 | 56.00 | 0.07871 | 0.03931 | 0.04640 | 0.2855 | 0.1806 |-| 0.3 | 0.3 | 0.5 | 55.86 | 0.07852 | 0.04005 | 0.04689 | 0.2845 | 0.1667 |--读数：PLAN 初值 ε=0.6 使 neighborhood_mmd 变差（0.0857–0.0899 > 父 0.0812）——此时列缩放中位 4.35×、PBC 钳制 37% 细胞，平滑过强把与坐标配套的型内差异抹掉再放大；按 PLAN 步骤 (4) 重扫 ε 后，ε∈{0.3, 0.4}（列缩放 ~2.4×）同时改善 nb_mmd、variogram、mmd_u、de_direction、de_score。γ 的最优随平滑移动（父 0.75 → 提交 0.3，父 ANALYSIS lesson 3 再次成立）。选 γ=0.3 而非 γ=0：差 0.05 分远在噪声内，γ=0 在平台边界且会移除父系验证过的 CP10k 再闭合不变量修复；γ=0.3 在 [0, 0.4] 平台内部且恰好满足 PLAN variogram 目标。--## 对照验证（数组级）--- **SMOOTH_STEPS=0 + GAMMA=0.75**（其余同父）：输出与节点 27 预测**逐位一致**（`.X` 与坐标 array_equal True）——平滑模块关闭即父节点。-- SMOOTH_STEPS=0 + 提交 γ=0.3：输出与提交配置不同（平滑确实在运行）。-- `--ablate mechanism`：α=0 且平滑不执行，输出 = 锚阶段原样（copy_last，X 逐位一致；坐标经 write_t2 float32 舍入 ~1e-5，与父节点管线的舍入相同）。-- 确定性：seed 0/1/2 输出逐位相同（锚 n=24826 ≤ max_cells，分层 take 返回全部行）。-- 伪装视图安全：无绝对时间、无阶段名、无视图判断；NN 图、列缩放、速度、门控、PBC 界全部从输入现场计算。+几何均值总量比均匀线性缩放替代逐细胞线性总量回拉（γ=0.45）：log 域近似恒定每细胞偏移，保住位移的跨基因 dp 排序；PLAN 的速度表达水平回归经 6 配置查分证否后按备选机制路径提交。++# 节点 31（improve，父 = 节点 29，proxy_noscale A 半 56.23）++## 实际提交的方法族与机制++- family_id: T2HX-01（copy_last 基座上的按型速度乘法外推），底物/门控/坐标冻结全部承自节点 29。+- **提交机制：几何均值再闭合（geometric re-closure）**，`VEC_RECLOSE_GEO=1`（默认）、`VEC_GAMMA=0.45`（默认）。+  节点 24 以来的 γ 再闭合把每个位移后细胞的**线性总量**拉回其自身平滑后总量：由于拉回是逐细胞线性缩放，+  高表达基因以加性 log 偏移吸收、低表达基因以乘性方式吸收，**跨基因 dp 排序被扭曲**（实测再闭合后+  spearman(dp, dt_approx) 从 0.57 掉到 0.39）。几何变体改为施加统一线性缩放+  `s_i = (r_med / r_i)^γ`（`r_i = tot_new_i / tot_old_i`，r_med = 全体中位数）：log1p 域大值每细胞近似+  平移同一常数，位移的基因排序大部分保留（γ=0.3 时 spearman_post 0.4624 > 父 0.3941），同时位移后线性+  总量中位数仍落回未位移中位数（实测阶段的库容量不变量在中位数意义上保持）。零保持零、稀疏模式/坐标/+  行序/细胞数/组成全部不变；PBC=10、软阈值 k=0.5、门控、平滑 ε=0.4 底物均不变。++## PLAN 机制（速度表达水平回归）——已证否++PLAN 要求逐型把 v_t 对锚阶段 log1p 表达水平 OLS 回归取残差（std 复原）以改善 de_score。已按 PLAN 实现+（`VEC_VEL_REGRESS`/`VEC_REGRESS_MODE`/`VEC_REGRESS_W`/`VEC_REGRESS_KEEPMEAN`，代码保留、默认关），+离线诊断：33 个细胞型中 28 型 R² < 0.05（PLAN 风险 1 的放弃线），仅 Forebrain 0.12 / Hindgut 0.095 /+Unknown 0.095 / Neural Tube 0.083 有可去趋势；回归前后速度 Spearman 0.92–1.00。6 次查分（≥3 种幅度组合、+共 20 次查分额度用满）de_score 全部低于父节点 0.1944，PLAN 首查目标 ≥0.20 未达成，de_direction 监控线+0.28 大多被击穿 → 按 improve 规则转备选机制（下表 #1–#6）。++**结论（供后续节点）**：速度与表达水平的共线分量携带真实 DE 信号而非纯 chance 伪影——去水平化+（回归残差、加性解码）一律降 de_score/de_direction；#11 的支撑掩码加性解码（level-free dp）使+de_direction 塌到 0.046，是本方向最强的证否证据。++## 全部 20 次查分（proxy_noscale A 半；父 = 56.23：de_score 0.1944 / de_dir 0.2837 / mmd_u 0.04669 / vario 0.03795 / nb_mmd 0.07569）++| # | 配置 | 榜分 | de_score | de_dir | mmd_u | vario | nb_mmd |+|---|---|---:|---:|---:|---:|---:|---:|+| 1 | 回归 type w=1 | 55.86 | 0.1806 | 0.2638 | 0.04562 | 0.03991 | 0.07936 |+| 2 | 回归 global w=1 | 55.84 | 0.1667 | 0.2523 | 0.04692 | 0.03875 | 0.07828 |+| 3 | 回归 type w=0.5 | 55.88 | 0.1806 | 0.2716 | 0.04655 | 0.03949 | 0.07921 |+| 4 | 回归 type w=−0.5（反向放大） | 55.81 | 0.1528 | 0.2758 | 0.04881 | 0.03773 | 0.07881 |+| 5 | 回归 type w=1 保均值 | 55.79 | 0.1806 | 0.2538 | 0.04610 | 0.04011 | 0.07917 |+| 6 | 回归 global w=0.5 | 55.87 | 0.1667 | 0.2703 | 0.04727 | 0.03874 | 0.07861 |+| 7 | 收缩 N0=150 | 55.63 | 0.1667 | 0.2839 | 0.04649 | 0.04124 | 0.08069 |+| 8 | 收缩 N0=30 | 55.88 | 0.1667 | 0.2888 | 0.04704 | 0.03938 | 0.07912 |+| 9 | 收缩 N0=50 | 55.91 | 0.1806 | 0.2908 | 0.04683 | 0.03974 | 0.07935 |+| 10 | 收缩 N0=400 | 54.89 | 0.1528 | 0.2187 | 0.04704 | 0.04380 | 0.08328 |+| 11 | 加性解码 α=0.5（支撑掩码） | 50.81 | −0.0139 | 0.0462 | 0.05509 | 0.05929 | 0.10479 |+| 12 | 幅度分层 q40/1.15/0.55（移植节点30） | 55.87 | 0.1806 | 0.2869 | 0.05019 | 0.03644 | 0.08029 |+| 13 | 幅度分层 q40/1.3/0.4 | 55.69 | 0.1667 | 0.2821 | 0.05148 | 0.03581 | 0.08128 |+| 14 | 线性再闭合基准=原始总量 | 55.76 | 0.1667 | 0.2678 | 0.04894 | 0.03834 | 0.07897 |+| 15 | 非对称 AUP1.15/ADN0.85 | 55.83 | 0.1528 | 0.2954 | 0.05049 | 0.03565 | 0.08044 |+| 16 | 非对称 AUP0.85/ADN1.15 | 55.24 | 0.1389 | 0.2491 | 0.04694 | 0.04294 | 0.08049 |+| 17 | 速度空间扩散 w=0.5 k=15 | 55.80 | 0.1667 | 0.2811 | 0.04566 | 0.04121 | 0.07907 |+| 18 | **几何再闭合 γ=0.3** | 56.18 | 0.1944 | 0.3132 | 0.04564 | 0.03949 | 0.07906 |+| 19 | **几何再闭合 γ=0.45（提交）** | **56.33** | **0.2222** | **0.3215** | **0.04542** | 0.03954 | 0.07932 |+| 20 | 几何再闭合 γ=0.15 | 56.07 | 0.1806 | 0.2999 | 0.04597 | 0.03941 | 0.07885 |++备选机制族（#7–#17）：全局速度收缩（4 档幅度）、加性解码、幅度分层（2 档）、再闭合基准、方向非对称+（2 档）、速度空间扩散——全部 ≤ 55.91。几何再闭合是唯一超过父节点的机制（#18–#20，γ 单调上升）。++## 提交配置的分组变化（#19 vs 父）++- expression_change（父最弱组，57.02）→ **58.15**：de_score 0.1944→0.2222（+2 个量化步，全树最佳，+  超过 PLAN 首查目标 0.20）；de_direction 0.2837→0.3215（全树最佳，连续量、远超以往配置 ±0.01 的抖动）。+- cell_state 58.06 → 58.07：mmd_u 0.04669→0.04542（全树最佳），variogram 0.03795→0.03954（略差）。+- local_spatial 59.83 → 59.09：nb_mmd 0.07569→0.07932（唯一恶化项，−0.74 分，被表达组 +1.13 覆盖）。+- shape_scale 50.00 不变（proxy_noscale 固定 + 坐标冻结）。+- 总分 56.23 → 56.33（+0.10，在 ~1 分噪声内；但机制级证据是结构性的：dp 排序保持度 spearman_post+  0.39→0.46，两个 DE 指标同创树内最佳且方向一致，非单指标尖峰）。++## 机制生效证据与对照++- 机制改变的细胞：全部 24,826 个锚细胞的表达被位移（moved_frac=1.0），几何再闭合改变其中每个细胞的+  缩放因子（与线性再闭合逐细胞不同）；坐标/行序/细胞数/零模式逐位不变（数组级验证）。+- 关闭对照：`--ablate <任意名>`（含 harness 将传入的 `VEC_VEL_REGRESS`，未识别名按主要机制=几何再闭合+  处理）或 `VEC_RECLOSE_GEO=0` → 回到同 γ=0.45 的线性再闭合，输出与主运行不同（mechanism_active 可判）。+- `VEC_GAMMA=0.3 VEC_RECLOSE_GEO=0` 时输出与父节点 29 逐位一致（X 与坐标 array_equal=True，已验证）；+  回归/收缩/分层/扩散/加性解码全部默认关，默认输出 = #19 查分文件逐位一致（已验证）。  ## 验证过 / 未验证 -- 验证过：proxy_noscale A 半上的全部上表配置；平滑的伪批量复原、零模式、能量下降；平滑关 = 节点 27 逐位；ablate = copy_last；三种子确定。-- 未验证：B 半与真实 E10.5 外推（本榜本地尺子已知高估，历史 54.2→官网 49.6；判据按 PLAN 用指标级 raw——nb_mmd/variogram/mmd_u/de_direction 四项全部同向改善且超父节点，而非总分 +0.72）；ε 在 (0.4, 0.6) 之间更细的网格；插值榜上的可迁移性。-- 运行时：CPU ~3 s，峰值内存 ~1.5 GB（EXECUTION.json gpu:false）。+- 已验证：默认输出逐位 = 已查分的 #19；seed 0/1/2 逐位一致（锚 24,826 ≤ max_cells 25,179，分层 take+  返回全部行，机制无随机性）；伪装视图（时间统一 +1、manifest 键序打乱、路径更换）输出逐位一致；+  vec-check 通过（主运行与 --ablate）；运行 ~3s / <2GB（limits 30min/28GB 内）；EXECUTION.json gpu:false。+- 未验证：γ>0.45（0.6/0.75）——γ 扫描在测试范围内单调上升，0.45 是已测最大值而非内点最优，额度用尽+  无法外探，**后续节点可先试 γ=0.6**；B 半与官网外推榜的迁移（本地外推尺子已知高估，历史 54.2→49.6，+  +0.10 的总分差不可外推，但 de_score/de_direction/mmd_u 三项 raw 同向创树内最佳是更可信的信号）；+  nb_mmd −0.0036 的代价在真实括号上是否更大。+- 单输入阶段/锚≠末输入退路同父：平滑与速度一并关闭，逐位 copy_last。  ## 知识来源 -无外部生物学知识：平滑、速度估计、门控均为纯算法部件，所有统计量（NN 图、型内均值、列缩放、总量中位数）从视图输入现场计算。未使用任何保留阶段/保留基因型的测量值。+无生物学先验知识被使用（sources=[]）：本节点全部改动为算法级（再闭合的缩放几何），只依赖输入数据现场+计算的量（各细胞总量、中位总量比），不含任何由已发布阶段测量值硬编码的常数。++## 旋钮（默认 = 提交配置）++`VEC_RECLOSE_GEO`(1)、`VEC_GAMMA`(0.45)；继承：`VEC_SMOOTH_STEPS`(2)、`VEC_SMOOTH_EPS`(0.4)、+`VEC_SMOOTH_K`(15)、`VEC_ALPHA`(1.0)、`VEC_K`(0.5)、`VEC_PBC`(10)、`VEC_MIXMODE`(rel)、+`VEC_MIXREL`(1.0)、`VEC_MIXEPS`(1.0)、`VEC_GATE_SPEARMAN`(0.3)；已证否机制保留可复用：+`VEC_VEL_REGRESS`(0)、`VEC_REGRESS_MODE/W/KEEPMEAN/MINGENES`、`VEC_SHRINK_N0`(0)、`VEC_SHRINK_MODE`、+`VEC_DECODE`(mult)、`VEC_STRAT_Q`(0)/`UP`/`DN`、`VEC_VDIFF_W`(0)/`K`、`VEC_RECLOSE_REF`(smoothed)、+`VEC_AUP/ADN`(1.0)、`VEC_DEBUG`。diff --git a/solution/README.md b/solution/README.mdindex 01b140b..140a710 100644--- a/solution/README.md+++ b/solution/README.md@@ -1,8 +1,20 @@-# 空间平滑底物 + rel 速度乘法外推（T2:heart:val_extrap，T2HX-01，节点29）+# 几何均值再闭合 + 空间平滑底物 rel 速度乘法外推（T2:heart:val_extrap，T2HX-01，节点31） -copy_last 基座（坐标/行序/细胞数/组成冻结）上：锚阶段表达先在**坐标 15-NN 图上做两步支撑掩码扩散平滑**（ε=0.4，零模式逐位保留，逐基因列缩放精确复原伪批量，节点23/26 配方），再按共有细胞型的相对差速度 v_t=(m_a−m_p)/(m_p+1)（m=**未平滑**原始型内均值 log1p）做乘法位移 x′=x_smoothed·2^(α·v̂_t)（α=1，k=0.5 MAD 软阈值）；门控（std 比 ≥0.01、spearman ≥0.3，dp 对原始锚伪批量）通过后按 γ=0.3 把每细胞线性总量拉回其平滑后自身总量，PBC=10×median(原始总量) 钳亮尾。单输入/锚≠末输入/`--ablate` 时逐位回退 copy_last（平滑也关闭）。+copy_last 基座（坐标/行序/细胞数/组成冻结）承自节点 29：锚阶段表达先在坐标 15-NN 图上做两步支撑掩码+扩散平滑（ε=0.4，零模式逐位保留，逐基因列缩放复原伪批量），再按共有细胞型的相对差速度+v_t=(m_a−m_p)/(m_p+1) 做乘法位移 x′=x_smoothed·2^(α·v̂_t)（α=1，k=0.5 MAD 软阈值），门控+（std 比 ≥0.01、spearman ≥0.3）通过后做再闭合与 PBC=10 钳亮尾。 -- 运行：`python run.py --data <view> --out <pred.h5ad> --seed <int>`；CPU ~3s、~1.5GB（`EXECUTION.json gpu:false`）。-- 旋钮（默认=提交配置）：`VEC_SMOOTH_STEPS`(2)、`VEC_SMOOTH_EPS`(0.4)、`VEC_SMOOTH_K`(15)、`VEC_ALPHA`(1.0)、`VEC_K`(0.5)、`VEC_GAMMA`(0.3)、`VEC_PBC`(10.0)、`VEC_MIXMODE`(rel)、`VEC_MIXREL`(1.0)、`VEC_MIXEPS`(1.0)、`VEC_GATE_SPEARMAN`(0.3)、`VEC_DEBUG`。-- 对照：`VEC_SMOOTH_STEPS=0 VEC_GAMMA=0.75` 输出逐位 = 节点27。-- proxy_noscale A 半 **55.95**（父节点27 = 55.23）；nb_mmd 0.07896、variogram 0.03872、mmd_u 0.04763、de_direction 0.2864 全部优于父节点。16 次查分记录见 METHOD.md。+**节点 31 改动**：γ 再闭合从"逐细胞线性总量回拉"换成**几何均值总量比的统一线性缩放**+`s_i=(r_med/r_i)^γ`，γ=0.45（0.15/0.3/0.45 单调扫描的已测最大值）。线性回拉按表达水平不均地吸收+缩放、扭曲跨基因 dp 排序（spearman_post 0.39）；几何缩放在 log 域近似每细胞恒定偏移，保住位移的+基因排序（spearman_post 0.46）。proxy_noscale A 半 56.33（父 56.23）：de_score 0.2222、+de_direction 0.3215、mmd_u 0.04542 均创树内最佳，nb_mmd 0.07932 略差于父 0.07569。++- 运行：`python run.py --data <view> --out <pred.h5ad> --seed <int>`；CPU ~3s、<2GB（`EXECUTION.json gpu:false`）。+- 关闭对照：`--ablate <任意名>` 或 `VEC_RECLOSE_GEO=0` → 同 γ 的线性再闭合；+  `VEC_GAMMA=0.3 VEC_RECLOSE_GEO=0` → 输出逐位 = 节点 29。+- 已证否（默认关，代码保留）：速度表达水平回归（PLAN，6 配置全降 de_score）、全局速度收缩、+  加性解码、幅度分层移植、方向非对称、速度空间扩散、原始总量再闭合基准——20 次查分明细见 METHOD.md。+- 确定性：seed 0/1/2 逐位一致；伪装视图（时间平移/键序打乱/换路径）逐位一致；单输入/锚≠末输入退路+  逐位 copy_last。diff --git a/solution/run.py b/solution/run.pyindex f727353..db1c01f 100644--- a/solution/run.py+++ b/solution/run.py@@ -24,8 +24,9 @@ Mechanism (on by default, alpha = 1.0):  Coordinates, row order and cell count are never modified. ---ablate <name> (any name) turns the mechanism off: output is bit-for-bit-copy_last (same anchor stage, same rows, same coordinates).+--ablate <name> (any name) turns THIS node's mechanism (the geometric-mean+re-closure, see (6) below) off and keeps the rest of the pipeline+unchanged (linear re-closure at the submitted gamma).  Node 24 additions on top of node 21: @@ -67,6 +68,54 @@ Node 25 addition on top of node 24 (the submitted mechanism):     de_direction 0.208 -> 0.228, variogram 0.0419 -> 0.0387, mmd_u 0.0575 ->     0.0566, neighborhood_mmd 0.0847 -> 0.0849 (flat). +Node 31 addition on top of node 29 (the submitted mechanism):++(6) Geometric-mean re-closure (VEC_RECLOSE_GEO, default 1; VEC_GAMMA+    default 0.45). The node-24 linear-total re-closure pulls each shifted+    cell's LINEAR total back toward its own smoothed total; because the+    pull-back is a per-cell linear scale, bright genes absorb it as an+    additive log shift while dim genes absorb it multiplicatively, which+    distorts the cross-gene dp ranking (spearman(dp, dt_approx) after+    re-closure 0.39 vs 0.57 before, measured). The geometric variant+    applies instead a uniform linear scale s_i = (r_med / r_i)**gamma with+    r_i = tot_new_i/tot_old_i: large log1p entries shift by an almost+    constant per-cell amount, the gene ranking of the displacement largely+    survives (spearman_post 0.46 at gamma=0.3), and the shifted linear+    total median still lands at the unshifted median (library-size+    invariant of measured stages preserved in the median). Zeros stay+    zero, sparsity pattern and coordinates untouched. Proxy A-half+    (20 queries): gamma 0.15/0.3/0.45 -> 56.07/56.18/56.33 vs node-29+    parent 56.23; at gamma=0.45 de_score 0.2222 (tree best, parent+    0.1944), de_direction 0.3215 (parent 0.2837), mmd_u 0.04542 (parent+    0.04669); nb_mmd 0.07932 slightly worse (parent 0.07569). gamma=0.45+    is the tested maximum of a monotone sweep (quota exhausted before+    0.6 could be probed).++    Off-control: VEC_RECLOSE_GEO=0 or --ablate <any> reverts ONLY the+    re-closure to the parent's linear-total pull-back at the same gamma;+    VEC_VEL_REGRESS/VEC_SHRINK_N0/VEC_STRAT_Q/VEC_VDIFF_W/VEC_DECODE=add+    are all off by default (falsified alternatives, kept for reuse).+    With --ablate the pipeline is node 29 except gamma=0.45 in the linear+    re-closure.++    FALSIFIED (do not retry): (a) the PLAN's velocity expression-level+    regression (per-type/global OLS residual of v_t on anchor log1p+    pseudobulk, std-rescaled, 6 scored configs incl. w=0.5/1/-0.5 and+    keep-mean): de_score 0.1528-0.1806 < parent 0.1944 in ALL configs --+    measured per-type R^2 is 0.00-0.12, and the level-covariant part of+    the velocity carries real de signal, not just chance-corrected+    artefact. (b) support-masked additive decode in log1p space (level-free+    dp): de_direction collapses to 0.046 -- dp's level coupling is+    signal, not artefact. (c) shrinkage of per-type velocity toward the+    global pseudobulk velocity (N0 30-400): de_direction up to 0.2908 but+    nb_mmd/mmd_u down, total <= 55.91. (d) node-30 magnitude+    stratification ported onto this base: 55.87/55.69. (e) asymmetric+    AUP/ADN: best de_direction 0.2954 + best variogram 0.03565 at+    AUP1.15/ADN0.85 but de_score 0.1528, total 55.83. (f) spatial+    velocity diffusion on coordinate 15-NN: best mmd_u 0.04566 but+    nb_mmd 0.0791, total 55.80. (g) linear re-closure toward ORIGINAL+    (pre-smoothing) totals: 55.76.+ Node 29 addition on top of node 27 (the submitted mechanism):  (5) Spatial support-masked diffusion smoothing of the anchor substrate@@ -128,9 +177,12 @@ Gates (std(dp) >= 0.01 * std(dt_approx), spearman(dp, dt_approx) >= 0.3) are computed on the PRE-renormalisation shift, exactly as in node 21, so the re-closure and clamp steps can never be blocked by them. -Experiment knobs (defaults = submitted configuration): VEC_SMOOTH_STEPS (2),+Experiment knobs (defaults = submitted configuration): VEC_VEL_REGRESS (1),+VEC_REGRESS_MODE (type), VEC_REGRESS_W (1.0), VEC_REGRESS_MINGENES (50),+VEC_SMOOTH_STEPS (2), VEC_SMOOTH_EPS (0.4), VEC_SMOOTH_K (15), VEC_ALPHA (1.0),-VEC_C (0.0 = pure multiplicative), VEC_K (0.5), VEC_GAMMA (0.3),+VEC_C (0.0 = pure multiplicative), VEC_K (0.5), VEC_GAMMA (0.45),+VEC_RECLOSE_GEO (1), VEC_PBC (10.0), VEC_MIXMODE (rel), VEC_MIXREL (1.0), VEC_MIXEPS (1.0), VEC_DOMAIN (log), VEC_SOFTD (0), VEC_CAPD (0), VEC_CVEL (-1 = legacy max(c,1)), VEC_VMEAN (log; used only in est mode), VEC_WMIX (0.0),@@ -158,7 +210,7 @@ from src.task2_spatial.view_io import ( ALPHA = float(os.environ.get("VEC_ALPHA", "1.0")) PSEUDOCOUNT = float(os.environ.get("VEC_C", "0.0")) SOFT_K = float(os.environ.get("VEC_K", "0.5"))-GAMMA = float(os.environ.get("VEC_GAMMA", "0.3"))+GAMMA = float(os.environ.get("VEC_GAMMA", "0.45")) DOMAIN = os.environ.get("VEC_DOMAIN", "log")  # "log" (node 24) or "lin" (node 25) CAPD = float(os.environ.get("VEC_CAPD", "0.0"))  # hard log-domain displacement cap (0 = off) SOFTD = float(os.environ.get("VEC_SOFTD", "0.0"))  # soft (tanh) displacement knee, log1p units (0 = off)@@ -172,6 +224,21 @@ AUP = float(os.environ.get("VEC_AUP", "1.0"))  # exponent multiplier for v>0 gen ADN = float(os.environ.get("VEC_ADN", "1.0"))  # exponent multiplier for v<0 genes PBC = float(os.environ.get("VEC_PBC", "10.0"))  # post-brightness clamp: max total / median original total (0 = off) GATE_SPEARMAN = float(os.environ.get("VEC_GATE_SPEARMAN", "0.3"))+VEL_REGRESS = os.environ.get("VEC_VEL_REGRESS", "0") not in ("0", "", "false", "False")+SHRINK_N0 = float(os.environ.get("VEC_SHRINK_N0", "0"))  # shrinkage pseudo-count (0 = off, output = node 29)+SHRINK_MODE = os.environ.get("VEC_SHRINK_MODE", "min")  # n_eff = "min"(n_a,n_b) or "prod" sqrt(n_a*n_b)+DECODE = os.environ.get("VEC_DECODE", "mult")  # "mult" (parent) or "add": support-masked additive in log1p space+STRAT_Q = float(os.environ.get("VEC_STRAT_Q", "0.0"))  # magnitude-stratification: top-|A*v| fraction boosted (0 = off)+STRAT_UP = float(os.environ.get("VEC_STRAT_UP", "1.15"))  # multiplier for the top fraction+STRAT_DN = float(os.environ.get("VEC_STRAT_DN", "0.55"))  # multiplier for the rest+RECLOSE_REF = os.environ.get("VEC_RECLOSE_REF", "smoothed")  # gamma re-closure target totals: "smoothed" (parent) or "original" (pre-smoothing anchor totals)+RECLOSE_GEO = os.environ.get("VEC_RECLOSE_GEO", "1") not in ("0", "", "false", "False")  # re-close toward the geometric-mean total ratio (uniform linear scale, log-domain library-size invariant)+VDIFF_W = float(os.environ.get("VEC_VDIFF_W", "0.0"))  # spatial velocity diffusion weight (0 = off)+VDIFF_K = int(os.environ.get("VEC_VDIFF_K", "15"))  # coordinate neighbours for the velocity diffusion+REGRESS_MODE = os.environ.get("VEC_REGRESS_MODE", "type")  # "type" (per-type OLS) or "global" (pooled slope)+REGRESS_W = float(os.environ.get("VEC_REGRESS_W", "1.0"))  # blend weight of the residual velocity (1 = pure residual)+REGRESS_MINGENES = int(os.environ.get("VEC_REGRESS_MINGENES", "50"))  # types with fewer positive-pb genes skip regression+REGRESS_KEEPMEAN = os.environ.get("VEC_REGRESS_KEEPMEAN", "0") not in ("0", "", "false", "False")  # re-add mean(v_t) to the rescaled residual SMOOTH_STEPS = int(os.environ.get("VEC_SMOOTH_STEPS", "2")) SMOOTH_EPS = float(os.environ.get("VEC_SMOOTH_EPS", "0.4")) SMOOTH_K = int(os.environ.get("VEC_SMOOTH_K", "15"))@@ -246,24 +313,36 @@ def spatial_smooth(Xa: np.ndarray, coords: np.ndarray, steps: int, eps: float,     return Xs, info  +def _ols_residual(v: np.ndarray, x: np.ndarray, coef: np.ndarray | None) -> tuple[np.ndarray, np.ndarray, float]:+    """OLS of v on [1, x]; returns (residual, coef, r2). coef=None -> fit."""+    A = np.vstack([np.ones_like(x), x]).T+    if coef is None:+        coef, *_ = np.linalg.lstsq(A, v, rcond=None)+    resid = v - A @ coef+    ss_tot = float(((v - v.mean()) ** 2).sum())+    r2 = 1.0 - float((resid ** 2).sum()) / ss_tot if ss_tot > 0 else 0.0+    return resid, coef, r2++ def type_velocity(Xa: np.ndarray, la: np.ndarray, Xb: np.ndarray, lb: np.ndarray,-                  c: float, k: float) -> tuple[np.ndarray, list]:+                  c: float, k: float, regress: bool, shrink: bool) -> tuple[np.ndarray, list]:     """Per-cell log2-domain pseudobulk fold-change velocity for anchor cells."""     V = np.zeros_like(Xa)     types_a = np.asarray(la).astype(str)     types_b = np.asarray(lb).astype(str)     per_type = []+    raw: dict[str, tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]] = {}     for t in sorted(set(types_a.tolist())):         ia = np.flatnonzero(types_a == t)         ib = np.flatnonzero(types_b == t)         if len(ib) == 0:-            per_type.append((t, len(ia), 0, 0.0, 0.0))+            per_type.append((t, len(ia), 0, 0.0, 0.0, None))             continue+        pb_a = Xa[ia].mean(axis=0)         if MIXMODE == "rel":             # PLAN literal mix: dimensionless displacement between the arith             # log2 fold-change of the pseudobulks and the relative arithmetic             # difference (pb_a - pb_p)/(pb_p + eps).-            pb_a = Xa[ia].mean(axis=0)             pb_p = Xb[ib].mean(axis=0)             v_arith = np.log2(pb_a + c) - np.log2(pb_p + c)             v_rel = (pb_a - pb_p) / (pb_p + MIXEPS)@@ -273,12 +352,85 @@ def type_velocity(Xa: np.ndarray, la: np.ndarray, Xb: np.ndarray, lb: np.ndarray             # the arith log2-of-mean (pseudobulk fold-change) velocity.             # WMIX = 0 is bit-for-bit the legacy estimator, WMIX = 1 pure arith.             v_log = np.log2(Xa[ia] + c).mean(axis=0) - np.log2(Xb[ib] + c).mean(axis=0)-            v_arith = np.log2(Xa[ia].mean(axis=0) + c) - np.log2(Xb[ib].mean(axis=0) + c)+            v_arith = np.log2(pb_a + c) - np.log2(Xb[ib].mean(axis=0) + c)             w = 1.0 if VMEAN == "arith" else WMIX             v_t = (1.0 - w) * v_log + w * v_arith+        raw[t] = (ia, ib, pb_a, v_t)++    # Alternative mechanism (node 31, submitted): shrink each type's velocity+    # toward the GLOBAL (all-cell) pseudobulk velocity with weight+    # s_t = n_eff/(n_eff + N0), n_eff = min(n_a, n_b) (SHRINK_MODE=prod:+    # sqrt(n_a*n_b)). Types measured with few cells in either stage have noisy+    # per-type velocities; the global velocity is estimated from all cells and+    # carries the coherent direction of change. N0 = 0 -> s_t = 1 -> identity+    # (bit-for-bit node 29). Types absent from prev keep v = 0 (parent+    # behaviour: no evidence, no displacement).+    if shrink:+        pb_a_all = Xa.mean(axis=0)+        pb_p_all = Xb.mean(axis=0)+        if MIXMODE == "rel":+            v_arith_g = np.log2(pb_a_all + c) - np.log2(pb_p_all + c)+            v_glob = (1.0 - MIXREL) * v_arith_g + MIXREL * (pb_a_all - pb_p_all) / (pb_p_all + MIXEPS)+        else:+            w = 1.0 if VMEAN == "arith" else WMIX+            v_log_g = np.log2(Xa + c).mean(axis=0) - np.log2(Xb + c).mean(axis=0)+            v_arith_g = np.log2(pb_a_all + c) - np.log2(pb_p_all + c)+            v_glob = (1.0 - w) * v_log_g + w * v_arith_g+        shrink_info = []+        for t, (ia, ib, pb_a, v_t) in raw.items():+            n_a, n_b = len(ia), len(ib)+            if SHRINK_MODE == "prod":+                n_eff = float(np.sqrt(n_a * n_b))+            else:+                n_eff = float(min(n_a, n_b))+            s_t = n_eff / (n_eff + SHRINK_N0)+            raw[t] = (ia, ib, pb_a, s_t * v_t + (1.0 - s_t) * v_glob)+            shrink_info.append((t, n_a, n_b, s_t))+        # types in the anchor but absent from prev never enter `raw`; the+        # parent leaves them at v = 0 -- keep that behaviour (V stays 0).+        if DEBUG:+            for t, n_a, n_b, s_t in shrink_info:+                print(f"  shrink {t}: n_a={n_a} n_b={n_b} s_t={s_t:.4f}", flush=True)++    # Velocity expression-level regression (node 31). Global mode pools all+    # (gene, type) pairs to fit one slope; type mode fits each type separately.+    coef_global = None+    if regress and REGRESS_MODE == "global" and raw:+        xs = np.concatenate([pb_a for _, _, pb_a, _ in raw.values()])+        vs = np.concatenate([v_t for _, _, _, v_t in raw.values()])+        _, coef_global, _ = _ols_residual(vs, xs, None)+        if DEBUG:+            print(f"regress: mode=global slope={coef_global[1]:+.5f} intercept={coef_global[0]:+.5f}",+                  flush=True)++    for t, (ia, ib, pb_a, v_t) in raw.items():+        rinfo = None+        n_pos = int((pb_a > 0).sum())+        if regress and n_pos < REGRESS_MINGENES:+            regress_t = False  # PLAN risk 4: too few expressed genes -> skip+        else:+            regress_t = regress+        if regress_t:+            x_reg = pb_a+            if REGRESS_MODE == "global":+                resid, _, r2 = _ols_residual(v_t, x_reg, coef_global)+            else:+                resid, coef_t, r2 = _ols_residual(v_t, x_reg, None)+            sd_v = float(v_t.std())+            sd_r = float(resid.std())+            scale = sd_v / sd_r if sd_r > 1e-12 else 1.0+            v_res = resid * scale+            if REGRESS_KEEPMEAN:+                v_res = v_res + float(v_t.mean())+            v_new = (1.0 - REGRESS_W) * v_t + REGRESS_W * v_res+            from scipy.stats import spearmanr+            sp_pre_post = float(spearmanr(v_t, v_new).statistic)+            rinfo = {"r2": r2, "std_scale": scale, "spearman": sp_pre_post,+                     "n_genes": int(v_t.size)}+            v_t = v_new         v_t, zeroed = soft_threshold(v_t, k)         V[ia] = v_t-        per_type.append((t, len(ia), len(ib), float(np.linalg.norm(v_t)), zeroed))+        per_type.append((t, len(ia), len(ib), float(np.linalg.norm(v_t)), zeroed, rinfo))     return V, per_type  @@ -290,9 +442,20 @@ def main() -> None:     parser.add_argument("--ablate", default=None)     args = parser.parse_args() -    alpha = 0.0 if args.ablate is not None else ALPHA-    pseudocount = PSEUDOCOUNT if args.ablate is None else 0.0-    gamma = 0.0 if args.ablate is not None else GAMMA+    alpha = ALPHA+    pseudocount = PSEUDOCOUNT+    gamma = GAMMA+    # Off-control: --ablate <any> (the harness passes the PLAN's+    # VEC_VEL_REGRESS; unrecognised names are treated as the primary+    # mechanism, which for the SUBMITTED configuration is the+    # geometric-mean re-closure) disables regression/shrinkage/velocity+    # diffusion/geo-re-closure; everything else (smoothing, rel velocity,+    # soft threshold, alpha, gamma=0.45 linear re-closure, PBC, gates,+    # frozen coordinates/rows/composition) is untouched.+    regress = VEL_REGRESS and args.ablate is None+    shrink_on = SHRINK_N0 > 0.0 and args.ablate is None+    vdiff_w = VDIFF_W if args.ablate is None else 0.0+    reclose_geo = RECLOSE_GEO and args.ablate is None      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)@@ -326,7 +489,25 @@ def main() -> None:                       f"zero_pattern_preserved={sinfo.get('zero_pattern_preserved')} "                       f"col_scale_med={sinfo.get('col_scale_med', 1.0):.4f}", flush=True)             c_vel = pseudocount if pseudocount > 0 else (CVEL if CVEL >= 0 else 1.0)-            V, per_type = type_velocity(Xa, stage.labels[rows], Xb, prev.labels, c_vel, SOFT_K)+            V, per_type = type_velocity(Xa, stage.labels[rows], Xb, prev.labels, c_vel, SOFT_K, regress, shrink_on)+            if vdiff_w > 0:+                # Spatial velocity diffusion: blend each cell's type velocity+                # with the mean type velocity over its VDIFF_K nearest+                # COORDINATE neighbours, making the displacement field locally+                # coherent (targets neighborhood_mmd, the heaviest metric).+                from scipy import sparse+                from scipy.spatial import cKDTree+                m = V.shape[0]+                kk = int(min(VDIFF_K, max(m - 1, 1)))+                cpt = np.asarray(coords, dtype=np.float64)[:, :3]+                tree = cKDTree(cpt)+                _, vidx = tree.query(cpt, k=kk + 1)+                vnb = np.atleast_2d(vidx)[:, 1:]+                vrows = np.repeat(np.arange(m), vnb.shape[1])+                Wv = sparse.csr_matrix(+                    (np.full(vrows.size, 1.0 / vnb.shape[1]), (vrows, vnb.ravel())),+                    shape=(m, m))+                V = (1.0 - vdiff_w) * V + vdiff_w * (Wv @ V)             pba = Xa.mean(axis=0)             dt_approx = pba - Xb.mean(axis=0)             if DOMAIN == "lin":@@ -338,6 +519,41 @@ def main() -> None:                 Xlin = np.expm1(Xs) * np.power(2.0, np.where(V > 0, alpha * AUP, alpha * ADN) * V)                 Xp = np.log1p(np.maximum(Xlin, 0.0))                 Xraw = Xp+            elif DECODE == "add":+                # Support-masked ADDITIVE decode in log1p space: x' = x + A*v~+                # on positive entries, zeros stay exactly zero (sparsity+                # pattern bit-preserved). dp_gene ~= A*v~ * frac_positive,+                # i.e. the pseudobulk change is (nearly) free of the+                # expression-level factor that the multiplicative-in-log1p+                # decode carries (dp ~ x * (2**v - 1)); de_score's chance+                # correction subtracts exactly the level-driven ranking, so a+                # level-free dp should convert more of the velocity signal+                # into hit rate. Negative shifts that would push x below 0+                # are clipped at 0 (clip_frac monitored in DEBUG).+                A = np.where(V > 0, alpha * AUP, alpha * ADN)+                Xraw = Xs + np.where(Xs > 0, A * V, 0.0)+                Xp = np.maximum(Xraw, 0.0)+                Xp = np.where(Xs > 0, Xp, 0.0)+            elif STRAT_Q > 0:+                # Magnitude-stratified displacement (ported from node 30,+                # proven on the sibling branch): pool all gated (type, gene)+                # velocity entries, rank by |A*v| descending, the top+                # STRAT_Q fraction is boosted (x STRAT_UP) and the rest+                # compressed (x STRAT_DN); a single global factor restores the+                # total displacement magnitude (conservation). Reweights WHICH+                # genes carry the extrapolation without changing the velocity+                # estimator itself.+                A = np.where(V > 0, alpha * AUP, alpha * ADN)+                mag = np.abs(A * V)+                nz = mag > 0+                if nz.any():+                    thr = float(np.quantile(mag[nz], 1.0 - STRAT_Q))+                    boost = np.where(nz & (mag >= thr), STRAT_UP,+                                     np.where(nz, STRAT_DN, 1.0))+                    f = float(mag[nz].sum() / max((mag * boost)[nz].sum(), 1e-12))+                    A = A * boost * f+                Xraw = Xs * np.power(2.0, A * V)+                Xp = np.maximum(Xraw, 0.0)             else:                 A = np.where(V > 0, alpha * AUP, alpha * ADN)                 if pseudocount > 0:@@ -376,9 +592,12 @@ def main() -> None:                       f"std(dp)/std(dt_approx)={std_ratio:.4f} "                       f"clip_frac={clip_frac:.4f} moved_frac={moved:.4f} spearman(dp,dt_approx)={sp:.4f}",                       flush=True)-                for t, na, nb, vn, zfrac in per_type:-                    print(f"  type {t}: n_anchor={na} n_prev={nb} |v_t|={vn:.4f} zeroed_frac={zfrac:.3f}",-                          flush=True)+                for t, na, nb, vn, zfrac, rinfo in per_type:+                    line = f"  type {t}: n_anchor={na} n_prev={nb} |v_t|={vn:.4f} zeroed_frac={zfrac:.3f}"+                    if rinfo is not None:+                        line += (f" regress_r2={rinfo['r2']:.4f} std_scale={rinfo['std_scale']:.3f}"+                                 f" spear_pre_post={rinfo['spearman']:.4f}")+                    print(line, flush=True)             if std_ratio < 0.01:                 if DEBUG:                     print("DE protection gate: fallback to copy_last", flush=True)@@ -395,18 +614,49 @@ def main() -> None:                     # gamma (0 = off, 1 = exact re-closure). Zeros stay zero.                     lin = np.expm1(Xp)                     tot_new = lin.sum(axis=1, keepdims=True)-                    tot_old = np.expm1(Xs).sum(axis=1, keepdims=True)-                    scale = np.power(np.divide(tot_old, np.maximum(tot_new, 1e-12),-                                               out=np.ones_like(tot_new), where=tot_new > 1e-12), gamma)-                    Xp = np.log1p(lin * scale)-                    if DEBUG:-                        r = (tot_new / np.maximum(tot_old, 1e-12)).ravel()-                        dp2 = Xp.mean(axis=0) - pba-                        sp2 = float(np.corrcoef(_rank(dp2), _rank(dt_approx))[0, 1])-                        print(f"renorm: gamma={gamma} shift_tot_ratio median={np.median(r):.4f} "-                              f"iqr=[{np.percentile(r, 25):.3f},{np.percentile(r, 75):.3f}] "-                              f"std_ratio_post={np.std(dp2) / (np.std(dt_approx) + 1e-12):.4f} "-                              f"spearman_post={sp2:.4f}", flush=True)+                    if reclose_geo:+                        # Geometric-mean re-closure: instead of pulling each+                        # cell's LINEAR total back (which distorts the+                        # cross-gene dp ranking: bright genes get an additive+                        # log shift, dim genes a multiplicative one), apply a+                        # uniform linear scale s_i = (r_med/r_i)**gamma with+                        # r_i = tot_new/tot_old. In log1p space large entries+                        # shift by a constant per cell, so the gene ranking of+                        # the displacement survives; the shifted linear-total+                        # median lands at the unshifted median.+                        tot_old_g = np.expm1(Xs).sum(axis=1, keepdims=True)+                        r = tot_new / np.maximum(tot_old_g, 1e-12)+                        r_med = float(np.median(r))+                        sg = np.power(np.divide(np.full_like(r, r_med), np.maximum(r, 1e-12)), gamma)+                        Xp = np.log1p(lin * sg)+                        if DEBUG:+                            dp2 = Xp.mean(axis=0) - pba+                            sp2 = float(np.corrcoef(_rank(dp2), _rank(dt_approx))[0, 1])+                            print(f"renorm_geo: gamma={gamma} r_med={r_med:.4f} "+                                  f"std_ratio_post={np.std(dp2) / (np.std(dt_approx) + 1e-12):.4f} "+                                  f"spearman_post={sp2:.4f}", flush=True)+                    else:+                        if RECLOSE_REF == "original":+                            # Node-29 ANALYSIS suggestion: pull each shifted cell+                            # back toward its OWN ORIGINAL (pre-smoothing) linear+                            # total instead of the smoothed one, so the per-cell+                            # library-size distribution of the measured anchor+                            # stage is preserved exactly (smoothing + column scale+                            # conserve only the global pseudobulk).+                            tot_old = np.expm1(Xa).sum(axis=1, keepdims=True)+                        else:+                            tot_old = np.expm1(Xs).sum(axis=1, keepdims=True)+                        scale = np.power(np.divide(tot_old, np.maximum(tot_new, 1e-12),+                                                   out=np.ones_like(tot_new), where=tot_new > 1e-12), gamma)+                        Xp = np.log1p(lin * scale)+                        if DEBUG:+                            r = (tot_new / np.maximum(tot_old, 1e-12)).ravel()+                            dp2 = Xp.mean(axis=0) - pba+                            sp2 = float(np.corrcoef(_rank(dp2), _rank(dt_approx))[0, 1])+                            print(f"renorm: gamma={gamma} shift_tot_ratio median={np.median(r):.4f} "+                                  f"iqr=[{np.percentile(r, 25):.3f},{np.percentile(r, 75):.3f}] "+                                  f"std_ratio_post={np.std(dp2) / (np.std(dt_approx) + 1e-12):.4f} "+                                  f"spearman_post={sp2:.4f}", flush=True)                 if PBC > 0:                     # Post-brightness clamp: after re-closure, a heavy tail of                     # cells (types with strongly positive velocity) still carries

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

Final state verified. Session summary:

**做了什么**：按 PLAN 实现速度表达水平回归（逐型/全局 OLS 残差、std 复原、幅度混合 w=0.5/1/−0.5、保均值共 6 配置）——离线诊断显示 28/33 型 R²<0.05，6 次查分 de_score 全部低于父节点 0.1944，PLAN 证否。按 improve 规则转备选机制，又测了 5 个族（全局速度收缩 4 档、支撑掩码加性解码、节点30幅度分层移植 2 档、AUP/ADN 非对称 2 档、速度空间扩散、原始总量再闭合基准），全部 ≤55.91。最后发现结构性问题：线性总量 γ 回拉按表达水平不均地吸收缩放、扭曲 dp 排序（spearman_post 0.39），改为**几何均值总量比统一线性缩放**（log 域近似恒定偏移，spearman_post 0.46），γ 扫描 0.15/0.3/0.45 单调上升，提交 γ=0.45。

**查分结果**（20/20 额度用满）：提交配置 proxy_noscale A 半 **56.33**（父 56.23）；de_score **0.2222**、de_direction **0.3215**、mmd_u **0.04542** 均创全树最佳（表达组 57.02→58.15，为父节点最弱组），代价 nb_mmd 0.07569→0.07932。默认输出逐位=已查分文件；seed 0/1/2、伪装视图逐位一致；`--ablate` 关闭几何再闭合（输出改变，mechanism_active 可判）；`VEC_GAMMA=0.3 VEC_RECLOSE_GEO=0` 逐位=父节点；vec-check 通过。

**下一步最值得试**：① γ=0.6/0.75（扫描在 0.45 处仍单调上升，是已测边界非内点最优）；② 几何再闭合 × 平滑 ε 重耦合（nb_mmd 是唯一恶化项，ε 微调或许可回收）；③ 本机制移植到插值两榜与全胚外推榜验证可迁移性；④ 用 B 半/官网复核——本地外推尺子已知高估，应以下一次官网提交验证 de_score/de_direction 增益是否兑现。

调研员

 ✓
- mechanism_off_control ✓

可选：
- sources（不用，因为没有用 lit-search）

好，方案完成。

让我再检查一下合规性：
- 不用禁窗数据 ✓
- 不硬编码统计量 ✓
- 不针对打分器漏洞 ✓（回归校正是改进估计质量，不是利用漏洞）
- 单输入退路 ✓
- 时间和内存可行 ✓（回归是简单操作）

好。

最后，让我确保方案不是"只是调一个参数"。回归校正是一个结构改动（引入了新的处理步骤和公式），不是调参数。

好，我满意这个方案。

输出最终 JSON。
</think>

```json
{
  "name": "速度表达水平回归校正改善DE排序",
  "motivation": "节点29 expression_change 57.02 是四个非固定组分中最弱（cell_state 58.06、local_spatial 59.83）；de_score raw 0.1944（skill 0.557）与 mmd_u 并列最低活跃指标。速度公式 (m_a−m_p)/(m_p+1) 的分母使速度与基因锚阶段表达水平系统性负相关，而 de_score 的 chance 校正恰好减去按 ±pb(ref) 排序的基线——部分本应贡献排序信号的速度方差被当作表达水平伪影扣掉。父节点 ANALYSIS 的 next_suggestions 第2条也指出 de_score 在 γ 平台上跳动（0.1528/0.1667/0.1806），说明排序信号尚未稳定释放。",
  "approach": "1) 在速度估计后、软阈值前，对每型 t 的速度向量 v_t 关于 log1p(pb_anchor_t)（逐基因）做 OLS 线性回归，取残差 v_res_t；2) 残差按该型原速度的标准差重缩放（保持位移幅度尺度不变）；3) 残差进入后续软阈值 k=0.5 → 乘法位移 α=1 → γ=0.3 再闭合 → PBC=10，管线其余不变（平滑 ε=0.4 底物、门控、坐标/行序/组成冻结）；4) 先试按型回归；若 de_score 未升，再试全局池化回归（所有型基因合并拟合一条斜率）；5) 单输入阶段 / 锚≠末输入退路同父节点：平滑与速度一并关闭，逐位 copy_last；6) vec-score 快筛：首查 de_score raw 目标 ≥0.20，同时监控 de_direction ≥0.28、mmd_u ≤0.048、nb_mmd ≤0.076，任一显著恶化即停止该方向。",
  "expected_groups": ["expression_change"],
  "risks": "1) 若真实发育变化本身与表达水平相关（如高表达基因确实在减速），回归会移除真实信号，de_score 反降——Engineer 应先打印回归 R²，R²<0.05 说明无趋势可去除，机制无效应尽早放弃并回退；2) 残差重缩放实现错误会改变速度幅度，需数组级验证每型标准差复原；3) 预期提升可能在 1 分噪声内，用种子 0/1/2 三次查分均值确认；4) 按型回归在小型（<30 基因有效表达）上可能过拟合，设最小基因数门限（如 50），低于则跳过该型回归。",
  "family_id": "T2HX-01",
  "mechanism": "逐型对基因速度向量关于锚阶段表达水平做线性回归取残差，去除速度与表达水平的系统性共线分量，让保留的速度只反映与表达水平无关的变化信号，从而在 de_score 的 chance 校正下获得更高命中率。",
  "vs_constant_shift": "常数位移/全局组成重加权对所有细胞或所有型施加统一变化量；本机制在基因维度上逐型去除速度的表达水平共线分量，每个基因被不同校正（回归残差不同），改变的是'哪些基因被认为在变化'的排序结构，而非统一幅度或方向。",
  "mechanism_evidence": "1) 回归 R² > 0（速度与表达水平确有共线分量可去除）；2) 回归前后速度向量 Spearman 相关 < 1（机制实际改变了速度排序）；3) de_score raw 从 0.1944 上升且 de_direction 不降；4) 全局 dp 与 dt 的秩偏相关（控制 pb(ref)）改善；5) 若 R² ≈ 0 或回归前后速度相关 > 0.99，说明机制未有效运行。",
  "mechanism_off_control": "环境变量 VEC_VEL_REGRESS=0 跳过回归步骤（速度保持原始值），其余管线不变，输出应与父节点 29 逐位一致（.X 与坐标 array_equal True）。",
  "sources": []
}
```

审查员

REVIEW.json written