总览 · ← 返回运行 20261003-171955-search-t2-embryo-interp-chain-12h

节点 n31 在终选来历上

在父节点27位移管线上，把逐类型Δ改为范数保持、支持集受限的基因相关扩散方向校正 Δ'=renorm(Δ+η·C·Δ)（仅|Δ|≥0.15基因，η=1，C=括号两阶段全细胞的基因-基因Pearson相关），使位移方向对齐组织内协同变化程序；PLAN字面的通路投影因位移在伪批量中精确对消而被证否，改交此备选解码。

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 bfb3e3fe86: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 5af70edcb8。改动的文件：solution/METHOD.md +78 −30、solution/run.py +444 −4

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex e8a5573..fe976a3 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,41 +1,89 @@-实现软阈值基因加权位移（PLAN T2EI-01）：表达位移的基因参与从硬阈值 1{|Δ|≥0.25} 改为连续幂次权重 w_g=clip(|Δ_g|/τ,0,1)^p（τ=0.25、p=1.0 提交），off/on 对照显示 de_direction 微升、de_score 不动，净效应 +0.03~0.04 榜分，在噪声（~1 分）内；提交保持机制打开。+在父节点27位移管线上，把逐类型Δ改为范数保持、支持集受限的基因相关扩散方向校正 Δ'=renorm(Δ+η·C·Δ)（仅|Δ|≥0.15基因，η=1，C=括号两阶段全细胞的基因-基因Pearson相关），使位移方向对齐组织内协同变化程序；PLAN字面的通路投影因位移在伪批量中精确对消而被证否，改交此备选解码。 -## 方法（family T2EI-01，PLAN：软阈值基因加权位移攻 DE 排序）+- family_id: T2EI-01+- parent: node 27（A 半 seed0 = 61.48，B 半 62.55）+- mechanism_off_control: `--ablate mechanism`（等价 T2_PROJ_ETA=0）→ Δ'=Δ，输出与父节点 27 逐位一致（已验证，含 seed 0/1） -保留父节点 13/19/22 全部管线（mix 分层混抽、procrustes3d、scale_damp=0.5、α=5 类型级收敛位移、λ=6 投影加权、β=0.2 型内 NN 配对收缩、全局 aniso damp=1.25、jitter、RMS 回缩；VARISO / TYPE_ANISO 关闭）。唯一改动在 mix_converge 的 Δ 构造处（run.py）：+## 1. 提交机制（CORRPROJ，默认打开） -- 硬阈值（off）：`d = where(|d|≥0.25, d, 0)` —— |Δ|<0.25 的基因位移为零，dp 中被压平，破坏 de_direction 秩偏相关；阈值附近全有/全无。-- 软权重（on，提交默认）：`d = d · clip(|d|/τ, 0, 1)^p`，环境变量 T2_SOFT_ENABLE（默认 1）、T2_SOFT_TAU（默认 0.25）、T2_SOFT_P（默认 1.0）。位移仍是 x ← x ± α·(t 或 1−t)·w·d，nnz-only 与 clip≥0 保留，稀疏结构不变（nnz 比例 0.059 → 0.059，逐位验证）。λ 投影加权与 β 配对的 DE 掩码逻辑未改（β 的 de 掩码仍用硬阈值，只作用于配对距离与收缩维度选择）。-- off 对照：T2_SOFT_ENABLE=0（或 T2_SOFT_P=0）走原硬阈值分支，seed 0 输出与父节点逐位一致（off 榜分 61.4533 = 父节点 A 半 61.45，验证通过）。-- 单输入阶段（无括号）走上层既有 copy_last 退路，软权重代码不执行。确定性：无新增随机源。+位置：`mix_converge` 的逐类型位移步之前（软权重、α=5、λ=6、β=0.2、nnz-only、clip≥0 全部保留不变）。 -## 机制生效证据（seed 0，proxy E6.75+E8.0→E7.25，t=0.4）+对每个共有类型（两侧各 ≥10 细胞）的 Δ = μ_b − μ_a： -- 参与位移基因数（sum(w_g>0.01)，10 个共有类型均值）：硬阈值 166.4 → 软权重 478.2；位移幅度按 |Δ| 幂次递减（maxAbsD 1.93，meanAbsD 0.0028）。-- 四组分（A 半查分）：expression_change 60.97→61.03（de_direction raw 0.3853→0.3879、skill 0.638→0.639；de_score raw 0.25 完全不动）；cell_state 48.93→48.92（mmd_u +0.00003 微劣、variogram 0.01442→0.01432 微升，均在否决线内）；local_spatial / shape_scale 逐位不变（坐标未动，nbr raw 0.05506→0.05505 来自表达经邻域平均的间接影响，<0.002 阈值）。+1. C = 括号两阶段全部细胞合并后的基因-基因 Pearson 相关矩阵（498×498，对角置 0），只依赖视图输入数据，现场计算；+2. Δ_diff = Δ + η·C·Δ（一步扩散，η=1）；+3. 支持集限制：只有 |Δ_g| ≥ 0.15（T2_PROJ_SUPPORT）的基因取 Δ_diff，其余基因保持测得的 Δ_g——不招募新基因、不翻转弱基因符号；+4. 范数保持：Δ' = Δ_diff·‖Δ‖/‖Δ_diff‖——只改方向不改幅度（不加此步时扩散把幅度放大 ~10 倍，表达值爆到 147（父节点 max 32.6），mmd_u 从 0.0117 崩到 0.0210）。 -## 网格结果（A 半，共 10 次查分）+参数（run.py 默认值，环境变量可覆写）：T2_PROJ_MODE=corr、T2_PROJ_ETA=1.0、T2_PROJ_SUPPORT=0.15、T2_PROJ_NORM=1、T2_PROJ_CORR_KERNEL=all。无新增随机源，seed 确定（已验证 seed 0 重跑逐位一致）。 -| 配置 | 榜分 | de_score | de_dir | mmd_u | variogram | nbr |-|---|---:|---:|---:|---:|---:|---:|-| off（=父13/19/22） | 61.453 | 0.2500 | 0.3853 | 0.01164 | 0.01442 | 0.05506 |-| p=0.5, τ=0.25 | 61.483 | 0.2500 | 0.3878 | 0.01168 | 0.01430 | 0.05505 |-| **p=1.0, τ=0.25（提交）** | **61.483** | 0.2500 | 0.3879 | 0.01167 | 0.01432 | 0.05505 |-| p=2.0, τ=0.25 | 61.474 | 0.2500 | 0.3874 | 0.01167 | 0.01435 | 0.05505 |-| p=1.0, τ=0.15 | 61.490 | 0.2500 | 0.3883 | 0.01168 | 0.01428 | 0.05505 |-| p=1.0, τ=0.50 | 61.411 | 0.2500 | 0.3858 | 0.01170 | 0.01450 | 0.05518 |-| p=1.0, α=8 | 60.808 | 0.2500 | 0.3796 | 0.01532 | 0.01352 | 0.05477 |-| p=1.0, α=3 | 60.430 | 0.2143 | 0.3843 | 0.01221 | 0.01594 | 0.05886 |+## 2. PLAN 被证否的过程（结构原因 + 离线筛选） -## 结论（诚实评估）+PLAN 的字面机制（通路 boost 进 Δ）实现后离线筛选发现**结构性对消**：分层混抽的 (1−t, t) 权重使逐细胞位移在输出伪批量中精确抵消，而 de_score / de_direction 只看 dp = pb(pred) − pb(ref)。因此任何只改 Δ 方向（不改幅度、不改 nnz/clip 非线性）的校正对 dp 几乎不可见： -- 机制方向正确但效应量远低于预期：de_direction raw +0.0026（PLAN 继续阈值 0.02 未达到），de_score 在所有软权重配置下纹丝不动——de_score 只看 |dp| 排序取 top-N 的命中率，小 |Δ| 基因的软位移量太小（α·t·Δ·(|Δ|/τ)^p，|Δ|≪τ 时近乎零），改变不了 top-N 集合，说明 de_score 缺口不在阈值压平而在位移方向本身（α·Δ_type 与真值 dt 的基因级错位）。-- α 网格证实 α=5 仍是局部最优（α=3 de_score 崩到 0.214，α=8 mmd_u 崩到 0.0153）。-- τ=0.5（对大 Δ 基因也衰减）净负：位移整体变弱等价于 α 减小。-- 净榜分 +0.03（61.453→61.483），T2 噪声 ~1 分，A/B 半之间不保证重现；所有否决线（mmd_u < +0.001、variogram、nbr < +0.002）均满足且 variogram/nbr 微改善，故按 PLAN"提交时保持打开"提交 p=1.0、τ=0.25。若 B 半回归，教训是：DE 两项对幅度不敏感、只对排序敏感，任何不改变 dp 排序的幅度级微调都无效。+- 通路版（Reactome top-k 二元 boost，k∈{20,30,50}，η∈{1,2}；GO-BP k∈{30,100}，η∈{1,2,4}，graded/graded_max 两种解码）共 8 个离线配置：|dp| top-50 集合变化 ≤2 个基因（PLAN 自己的放弃线 <2），cos(Δ,Δ')≥0.90，boost 均值 0.005–0.085——方向无效；+- 其中 1 个配置送查分存档：reactome η=2 k=20 → 60.86（劣于父节点，见 §5 表 #17）；+- 相关版字面实现（Δ'=Δ+ηCΔ，无范数保持）：η=0.1 → 60.09（de_score 0.2857 但 mmd_u 崩），说明其 dp 变化全部来自幅度爆炸的 clip 整流，不是方向。 -## 验证过 / 没验证+按任务书规则（≥3 种解码/幅度组合、≥6 次查分）判定 PLAN 字面方向无效后，改交针对同一弱项（expression_change / de_score 排序错位）的备选解码：把"相关先验校正位移方向"从幅度爆炸中剥离（范数保持 + 支持集限制），这是同一机制家族的合法解码变体。 -- 验证：off 与父逐位一致；on 的稀疏结构保持；p/τ/α 网格 8 配置；vec-check ok；默认配置输出与 p1.0 网格文件逐位一致。-- 没验证：seed 1/2 复跑（时间预算内查分优先给了网格）；τ<0.15（PLAN 风险 4 的密集位移区）；β 配对掩码同步改软权重（预期同样在噪声内）。-- 知识来源：无外部生物学知识；全部为评分指标性质推理（DE 只看排序/符号）与代理网格实测。+## 3. 机制生效证据++- **实际改变了什么**：每个共有类型的 Δ 方向被转动（cos(Δ,Δ') 均值 0.86、最小 0.82，>0.5 且 <1，符合 PLAN 的预期区间）；受影响细胞 = 全部 10 个共有类型的输出细胞（a 侧 3000 + b 侧 2000，位移量随 Δ' 方向变）；坐标、细胞抽取、nnz 结构不变（nnz 0.0590→0.0567，来自 Δ' 方向改变后 clip 位置的微小变化）。+- **dp 层**：|dp| top-50 变化 2–3 个基因（过了 PLAN 的 ≥2 筛选线），de_score raw 0.25→0.2857（seed 0）、0.2143→0.25（seed 1）。+- **四组分（A 半，seed 0，本节点 vs 父节点）**：expression_change 62.16→62.31（de_score +0.17，de_direction −0.07）；cell_state 50.36→50.29（variogram +0.12，mmd_u −0.08）；local_spatial 60.37→60.61（neighborhood_mmd raw 0.05505→0.05158，+0.40 points）；shape_scale 77.31→77.31（坐标未动，符合设计）。总分 61.48→62.06。+- **跨 seed 一致性**：seed 1 配对比较 61.55 vs 父节点 60.93（+0.63）；seed 0 +0.58。收益主要来自 neighborhood_mmd（+0.34~0.40）、de_score（+0.17）、variogram（+0.12~0.19），代价是 mmd_u（−0.06~0.08）与 de_direction（−0.04~0.07）——两次查分模式一致，非单 seed 尖峰。++## 4. 调参轨迹（support / η / 核）++| 配置 | A 半总分 |+|---|---|+| support 0.25, η=1 | 61.89 |+| support 0.25, η=2 | 61.86 |+| **support 0.15, η=1（提交）** | **62.06** |+| support 0.1, η=1 | 62.04 |+| support 0.15, within-type 核 | 61.85 |++support 0.1–0.15 是平台（差 0.03，噪声内），取 0.15；η 在 1→2 已饱和（范数保持下方向变化封顶）；within-type 残差核（去掉类型间组成结构）略差于全细胞核，说明组成间共变结构本身有用。++## 5. 全部查分记录（A 半 proxy，seed 0 除注明外；共 18 次，余 2）++| # | 配置 | 总分 | de_score raw | de_dir raw | mmd_u raw | variogram raw | nbhd raw |+|---|---|---|---|---|---|---|---|+| 1 | 父节点/off（逐位一致验证） | 61.48 | 0.2500 | 0.3879 | 0.01167 | 0.01432 | 0.05505 |+| 2 | corr 字面 η=0.1（爆幅度） | 60.09 | 0.2857 | 0.3716 | 0.02104 | 0.01206 | 0.05663 |+| 3 | corrnorm η=1（无 support 限制） | 61.41 | 0.2143 | 0.3757 | 0.01258 | 0.01406 | 0.05211 |+| 4 | pbproj 通路 graded_max η=1 | 59.39 | 0.2143 | 0.3810 | 0.01508 | 0.01765 | 0.05969 |+| 5 | TF 重排（CollecTRI）上调 η=0.6 | 61.36 | 0.1786 | 0.3913 | 0.01165 | 0.01527 | 0.05185 |+| 6 | TF 重排 下调 η=−0.6 | 60.60 | 0.2143 | 0.3596 | 0.01218 | 0.01391 | 0.05988 |+| 7 | tshift 斜率时序 κ=0.5 | 59.34 | 0.2500 | 0.3742 | 0.01635 | 0.01635 | 0.06100 |+| 8 | leakfix（f≡t 去位移泄漏） | 58.31 | 0.2143 | 0.3721 | 0.01766 | 0.02065 | 0.06075 |+| 9 | corr η=0.1 + bracket clamp1.3 | 61.23 | 0.2500 | 0.3729 | 0.01430 | 0.01198 | 0.05588 |+| 10 | corrnorm sup0.25 η=1 | 61.89 | 0.2857 | 0.3760 | 0.01199 | 0.01375 | 0.05279 |+| 11 | corrnorm sup0.25 η=2 | 61.86 | 0.2857 | 0.3760 | 0.01210 | 0.01379 | 0.05273 |+| 12 | **corrnorm sup0.15 η=1（提交）** | **62.06** | 0.2857 | 0.3813 | 0.01222 | 0.01347 | 0.05158 |+| 13 | corrnorm sup0.1 η=1 | 62.04 | 0.2857 | 0.3809 | 0.01235 | 0.01339 | 0.05168 |+| 14 | corrnorm sup0.15 within 核 | 61.85 | 0.2857 | 0.3845 | 0.01280 | 0.01352 | 0.05225 |+| 15 | 提交配置 seed 1 | 61.55 | 0.2500 | 0.3820 | 0.01173 | 0.01380 | 0.05201 |+| 16 | 父节点 seed 1（配对基线） | 60.93 | 0.2143 | 0.3888 | 0.01148 | 0.01471 | 0.05527 |+| 17 | 通路版 reactome η=2 k=20（PLAN 字面存档） | 60.86 | 0.2143 | 0.3793 | 0.01320 | 0.01418 | 0.05567 |++（#1 的另一次重复查分与 floor/max 探针占去其余额度，共用 19 次、余 1；shape 三项在所有表达侧变体中不变，坐标未动。#17 说明 PLAN 字面通路 boost 即使可测也劣于父节点，与离线筛选的"结构对消"结论一致。）++## 6. 知识来源++- **提交机制不用任何外部生物学知识**：C 完全由视图输入细胞现场计算（数据驱动的相关平滑），无硬编码统计量、无保留阶段信息、无文献数值。+- 已试但放弃的变体用到的资源（如实记录）：视图 prior/ 的 Reactome `gene_sets.gmt`、GO-BP `gene_sets_bp.gmt`（通路 boost，#4/#17，放弃）；视图 prior/ 的 CollecTRI `collectri_mouse.tsv.gz` + 教科书级 GRN 层级顺序知识（调控因子先于其靶基因变化；Davidson & Erwin 2006, Science, doi:10.1126/science.1121590）——仅用作 #5/#6 的方向依据，两个方向均劣于父节点，放弃。+- tshift（#7）的"阶段内异步性=伪时间"假设与 leakfix（#8）均为纯数据驱动，无外部知识。++## 7. 验证过 / 未验证++验证过：off 对照逐位一致（seed 0/1）；seed 0 重跑逐位一致（确定性）；伪装视图（全时间 +1 天、manifest 键序打乱、随机路径）输出与真实视图逐位一致（view-independent）；vec-check 通过（seed 0/1/2 及 ablate）；单输入阶段退路不经过本机制（bracket 为 None 时走 copy_last 分支）；prior/ 缺失时 corr 模式不依赖 prior（提交配置完全不读 prior/）；纯 CPU，峰值内存 0.61 GB，proxy 视图 1.6 s。++未验证：B 半分数（正式分）；真实 final 视图（括号两端与细胞数不同，C 与 Δ 都会变，机制无阶段特异性参数，预期可迁移但幅度未测）；final 视图更大细胞数下的 corr_kernel 稠密化内存（~每 10 万细胞 400 MB，28 GB 限内，但未实测）；support/η 平台在 final 括号上是否同位置。++## 8. 下一步建议++1. neighborhood_mmd 是本次收益最大的项（+0.4）：相关扩散让型内收敛后的细胞更像真实中间态；可试把扩散核换成"位移后残差"的相关（迭代一步），或只对 b 侧细胞扩散（a/b 不对称解码）。+2. mmd_u 的微损（−0.08）来自 Δ' 转动后个别基因 clip 位置变化；可试 clip 感知的扩散（对接近 0 的基因减弱校正）。+3. de_score 0.2857 疑似其局部上限（8 种独立重排里只有相关扩散族达到）；继续在 dp 重排上投入期望值低。diff --git a/solution/run.py b/solution/run.pyindex cf4216b..a951b54 100644--- a/solution/run.py+++ b/solution/run.py@@ -8,6 +8,16 @@ exp(log r_a + SCALE_DAMP·t·Δlog r), and draws cells stratified by type: round(t·n) from the later stage, the rest from the earlier one. Coordinates travel with the cells. n is log-linear in t, clipped to the board range. +This node (31, CORRPROJ, family T2EI-01): before the soft weights, each shared+type's Δ is direction-corrected by one-step gene-gene correlation diffusion+Δ' = renormalize(Δ + η·C·Δ) restricted to genes with |Δ| ≥ PROJ_SUPPORT, where+C is the Pearson correlation of all bracket-stage cells (zero diagonal) and the+renormalization restores ||Δ||. C encodes which genes co-vary within the+tissue, so the displacement direction aligns with coordinated gene programs+instead of per-gene measured differences. η = 0 (or --ablate mechanism)+reproduces the parent node 27 bit-for-bit. See METHOD.md for the falsified+pathway-boost decodes and the pb-level variants.+ New mechanism (CONVERGE_ALPHA, family T2EI-01): for every cell type present in both bracketing stages (≥10 cells per side), compute its temporal change vector Δ = μ_b − μ_a from the stage pseudobulk means, keeping only genes with@@ -95,6 +105,156 @@ CONVERGE_NNZ_ONLY = os.environ.get("T2_CONVERGE_NNZ_ONLY", "1") == "1" SOFT_ENABLE = os.environ.get("T2_SOFT_ENABLE", "1") == "1" SOFT_TAU = float(os.environ.get("T2_SOFT_TAU", "0.25")) SOFT_P = float(os.environ.get("T2_SOFT_P", "1.0"))+# Pathway/correlation prior projection (family T2EI-01, this node): the per-type+# temporal change Δ = μ_b − μ_a is direction-corrected with a biological prior+# before entering the displacement step. SUBMITTED DEFAULT: mode "corr" with+# norm-preservation and support restriction (see module docstring / METHOD.md).+# Mode "pathway": read Reactome gene sets+# from the view's prior/ directory, build the pathway×gene membership matrix M+# on panel genes; per type, pathway activity s_p = mean(|Δ_g|, g∈p); take the+# top-k active pathways; boost_g = (# active pathways containing g) /+# (# all pathways containing g); Δ'_g = Δ_g·(1 + η·boost_g). Genes in no+# pathway keep Δ_g. Mode "corr" (fallback when prior/ is unusable): gene-gene+# Pearson correlation C of all bracket cells, Δ' = Δ + η·C·Δ (one-step+# diffusion). η = 0 disables the mechanism and reproduces the parent bit-for-+# bit. Everything downstream (soft weights, α, λ, β, nnz-only, clip≥0) is+# unchanged and receives Δ' in place of Δ.+PROJ_ETA = float(os.environ.get("T2_PROJ_ETA", "1.0"))+PROJ_MODE = os.environ.get("T2_PROJ_MODE", "corr")+PROJ_K = int(os.environ.get("T2_PROJ_K", "30"))+PROJ_MIN_PATHWAY_GENES = int(os.environ.get("T2_PROJ_MIN_PATHWAY_GENES", "3"))+# Which prior gene-set family to read (pathway mode): reactome | go_bp | msigdb.+# go_bp covers ~98% of the panel genes vs ~52% for reactome on this 498-gene+# panel, giving the boost much more reach.+PROJ_SRC = os.environ.get("T2_PROJ_SRC", "go_bp")+# Renormalize Δ' to the original ||Δ|| after projection (direction-only change).+PROJ_NORM = os.environ.get("T2_PROJ_NORM", "1") == "1"+# Support restriction for corr diffusion: only genes with |Δ| >= PROJ_SUPPORT+# receive the diffused value; others keep their measured Δ (0 = off).+PROJ_SUPPORT = float(os.environ.get("T2_PROJ_SUPPORT", "0.15"))+# Correlation kernel for corr mode: "all" pools every bracket cell (program ++# composition structure); "within" centers cells by their own type mean first,+# keeping only within-state co-variation.+PROJ_CORR_KERNEL = os.environ.get("T2_PROJ_CORR_KERNEL", "all")+# Bracket-range clamp: after the displacement/pair steps, clip every gene's+# cell values at BRACKET_CLIP × max(value observed for that gene in either+# bracket stage). Prevents the corr-diffusion magnitude explosion (values up+# to 5× the biological range, which is what collapses mmd_u) while keeping the+# correlation structure the diffusion imprints. 0 = off (parent behavior).+BRACKET_CLIP = float(os.environ.get("T2_BRACKET_CLIP", "0"))+# Pseudobulk-level prior projection (PBPROJ, this node's submitted mechanism).+# Motivation: the stratified (1−t, t) draw weights make the per-cell+# displacement cancel EXACTLY in the output pseudobulk, so de_score /+# de_direction (which only see dp = pb(pred) − pb(ref)) are structurally blind+# to any correction applied to Δ inside the displacement step (verified+# offline: top-50 |dp| set moves ≤ 2 genes for every pathway configuration).+# PBPROJ therefore applies the same pathway-prior boost where the DE metrics+# can see it: after the full expression pipeline, compute the pooled bracket+# change Δ_pb = pb(b) − pb(a) and the pathway-boosted Δ'_pb (same boost_g as+# the pathway mode: top-k active prior gene sets per Δ_pb, boost_g = active/+# total membership), then rescale each gene's output column by+# s_g = 1 + η·boost_g·t·Δ_pb_g/max(pb_out_g, floor), clipped to [0, 1+cap].+# Multiplicative nnz scaling keeps the zero structure, the within-gene relative+# distribution and coordinates untouched; pb_out moves by ≈ η·boost_g·dp_g,+# i.e. genes inside coherently-changing prior sets get their (already+# correctly-signed) change magnitude amplified, isolated genes keep theirs.+# η = 0 disables and reproduces the parent bit-for-bit.+PBPROJ_ENABLE = os.environ.get("T2_PBPROJ_ENABLE", "0") == "1"+PBPROJ_ETA = float(os.environ.get("T2_PBPROJ_ETA", "0.5"))+PBPROJ_K = int(os.environ.get("T2_PBPROJ_K", "30"))+PBPROJ_SRC = os.environ.get("T2_PBPROJ_SRC", "reactome")+PBPROJ_FLOOR = float(os.environ.get("T2_PBPROJ_FLOOR", "0.05"))+PBPROJ_CAP = float(os.environ.get("T2_PBPROJ_CAP", "1.5"))+PBPROJ_BOOST_MODE = os.environ.get("T2_PBPROJ_BOOST_MODE", "graded")+# Weight source for PBPROJ (which genes' dp gets amplified):+#   pathway  — prior gene-set coherence boost (PLAN literal, at pb level)+#   tf       — regulatory-hierarchy re-ranking: genes that are transcription+#              factors (regulators in the view's CollecTRI prior) get w=1,+#              all other genes w=0. Rationale (textbook GRN hierarchy, e.g.+#              Davidson & Erwin 2006, doi:10.1126/science.1121590): in a+#              developmental transition, upstream regulators change before+#              their downstream targets, so at an intermediate target time the+#              regulators have already covered a larger fraction of their+#              bracket change than terminal/downstream genes. Amplifying the+#              regulator dp (and only its sign-consistent magnitude) re-ranks+#              the top-N toward genes that are further along their trajectory.+PBPROJ_W_MODE = os.environ.get("T2_PBPROJ_W_MODE", "pathway")+# TSHIFT (backup mechanism, expression_change): per-gene trajectory-timing+# correction from within-stage progression slopes. Assumption: cells inside one+# stage are asynchronous — a cell's projection q onto the type's Δ = μ_b − μ_a+# direction is a pseudotemporal progression coordinate (cells with high q are+# further along the a→b transition). A gene whose full a→b rise already happens+# WITHIN stage a (slope within a covers ~Δ over a's q-spread) is an early mover+# (f_g > t: it has covered more than a t-fraction of its change by the target+# time); a gene still rising WITHIN stage b is a late mover (f_g < t). The+# target pseudobulk is then pb_a + f_g·Δ_pb instead of pb_a + t·Δ_pb, realized+# as a per-gene column rescale of the output (nnz-only, sparse structure and+# coordinates untouched). Uses only bracket-stage data (no external knowledge,+# no target information). KAPPA=0 disables and reproduces the parent bit-for-bit.+TSHIFT_ENABLE = os.environ.get("T2_TSHIFT_ENABLE", "0") == "1"+TSHIFT_KAPPA = float(os.environ.get("T2_TSHIFT_KAPPA", "0.5"))+TSHIFT_MIN_CELLS = int(os.environ.get("T2_TSHIFT_MIN_CELLS", "30"))+TSHIFT_MIN_DELTA = float(os.environ.get("T2_TSHIFT_MIN_DELTA", "0.25"))+TSHIFT_FLO = float(os.environ.get("T2_TSHIFT_FLO", "0.5"))   # f >= FLO·t+TSHIFT_FHI = float(os.environ.get("T2_TSHIFT_FHI", "2.0"))   # f <= FHI·t (and <= 0.9)+TSHIFT_CAP = float(os.environ.get("T2_TSHIFT_CAP", "1.0"))+# TSHIFT_MODE: "timing" (slope-asymmetry f_g) or "leakfix" (f_g ≡ t: undo the+# displacement pipeline's nnz/clip leakage so dp equals the clean linear+# estimate t·Δ_pb on reliable genes).+TSHIFT_MODE = os.environ.get("T2_TSHIFT_MODE", "timing")+++def timing_fraction(stage_a, stage_b, shared_labels, kappa: float):+    """Per-gene f_g in [FLO·t, min(FHI·t, 0.9)] from within-stage slope asymmetry."""+    genes_n = len(stage_a.genes)+    num = np.zeros(genes_n)+    den = np.zeros(genes_n)+    for lab in shared_labels:+        A = np.asarray(stage_a.X[stage_a.labels == lab].toarray(), dtype=np.float64)+        B = np.asarray(stage_b.X[stage_b.labels == lab].toarray(), dtype=np.float64)+        if A.shape[0] < TSHIFT_MIN_CELLS or B.shape[0] < TSHIFT_MIN_CELLS:+            continue+        d = B.mean(axis=0) - A.mean(axis=0)+        reliable = np.abs(d) >= TSHIFT_MIN_DELTA+        if reliable.sum() < 5 or np.linalg.norm(d) < 1e-9:+            continue+        slopes = []+        for Xs in (A, B):+            mu = Xs.mean(axis=0)+            q = (Xs - mu) @ d+            sq = q.std()+            if sq < 1e-12:+                slopes.append(np.zeros(genes_n))+                continue+            s = ((Xs - mu) * (q - q.mean())[:, None]).mean(axis=0) / (sq * sq)+            # rise over a ±1 sd progression window, in units of the full change d+            slopes.append(2.0 * sq * s / np.where(np.abs(d) < 1e-9, np.nan, d))+        r_a, r_b = slopes+        timing = 0.5 * (r_a - r_b)  # +: early (done by a), −: late (still going at b)+        timing = np.where(reliable & np.isfinite(timing), np.clip(timing, -1.0, 1.0), 0.0)+        wt = float(min(A.shape[0], B.shape[0])) * np.abs(d)  # reliability weight+        num += wt * timing+        den += wt+    timing = np.where(den > 0, num / np.maximum(den, 1e-12), 0.0)+    return timing+++def tf_indicator(view: str, genes: list[str]) -> np.ndarray | None:+    """1.0 for panel genes that appear as a regulator (tf) in CollecTRI."""+    path = os.path.join(view, "prior", "tf_regulons", "collectri_mouse.tsv.gz")+    if not os.path.exists(path):+        return None+    import gzip+    gidx = {g: i for i, g in enumerate(genes)}+    out = np.zeros(len(genes))+    with gzip.open(path, "rt") as fh:+        fh.readline()+        for line in fh:+            p = line.split("\t")+            j = gidx.get(p[0])+            if j is not None:+                out[j] = 1.0+    return out # Per-cell heterogeneous displacement (family T2EI-01 refine): each cell's shift # is scaled by w = clip(1 -/+ λ·p, 0.2, 2.5), where p is the cell's deviation # from its type mean projected onto Δ (only genes with Δ≠0). Cells already@@ -178,6 +338,151 @@ VARISO_MIN_CELLS = int(os.environ.get("T2_VARISO_MIN_CELLS", "10")) VARISO_MIN_DE = int(os.environ.get("T2_VARISO_MIN_DE", "10"))  +def load_pathway_sets(view: str, genes: list[str], src: str | None = None):+    """Reactome pathway×gene membership on panel genes, from the view's prior/.++    Returns (M, names): M is a boolean (n_pathways, n_genes) matrix, names the+    pathway ids (sorted, deterministic). Pathways with fewer than+    PROJ_MIN_PATHWAY_GENES panel genes are dropped. Returns (None, None) when+    the prior directory or file is missing/unreadable.+    """+    src_files = {+        "reactome": os.path.join("prior", "reactome", "gene_sets.gmt"),+        "go_bp": os.path.join("prior", "go", "gene_sets_bp.gmt"),+        "msigdb": os.path.join("prior", "msigdb", "hallmark_mouse.gmt"),+    }+    rel = src_files.get(src if src is not None else PROJ_SRC)+    if rel is None:+        return None, None+    gmt = os.path.join(view, rel)+    if not os.path.exists(gmt):+        return None, None+    gene_idx = {g: i for i, g in enumerate(genes)}+    rows, names = [], []+    with open(gmt, "r") as fh:+        for line in fh:+            parts = line.rstrip("\n").split("\t")+            if len(parts) < 3:+                continue+            pid = parts[0]+            member = np.zeros(len(genes), dtype=bool)+            for sym in parts[2:]:+                j = gene_idx.get(sym)+                if j is not None:+                    member[j] = True+            if int(member.sum()) < PROJ_MIN_PATHWAY_GENES:+                continue+            rows.append(member)+            names.append(pid)+    if not rows:+        return None, None+    order = np.argsort(names)+    M = np.vstack(rows)[order]+    names = [names[i] for i in order]+    return M, names+++def corr_kernel(stage_a, stage_b, genes_n: int) -> np.ndarray:+    """Gene-gene Pearson correlation of all bracket-stage cells (dense, panel-sized)."""+    from scipy import sparse as sp+    X = sp.vstack([sp.csr_matrix(stage_a.X), sp.csr_matrix(stage_b.X)]).toarray().astype(np.float64)+    X = X - X.mean(axis=0, keepdims=True)+    sd = X.std(axis=0)+    sd[sd < 1e-12] = np.inf+    C = (X.T @ X) / X.shape[0]+    C = C / np.outer(sd, sd)+    np.fill_diagonal(C, 0.0)+    return C.astype(np.float32)+++def corr_kernel_within(stage_a, stage_b) -> np.ndarray:+    """Gene-gene Pearson correlation of within-type residuals (both stages pooled).++    Centering each cell by its own type mean removes between-type composition+    structure, leaving the within-state gene-gene co-variation — the coordinate+    programs (shared regulation) that the diffusion should align Δ with.+    """+    parts = []+    for st in (stage_a, stage_b):+        X = np.asarray(st.X.toarray(), dtype=np.float64)+        labs = np.asarray(st.labels).astype(str)+        for lab in np.unique(labs):+            m = labs == lab+            if m.sum() >= 10:+                parts.append(X[m] - X[m].mean(axis=0, keepdims=True))+    if not parts:+        return None+    X = np.vstack(parts)+    sd = X.std(axis=0)+    sd[sd < 1e-12] = np.inf+    C = (X.T @ X) / X.shape[0]+    C = C / np.outer(sd, sd)+    np.fill_diagonal(C, 0.0)+    return C.astype(np.float32)+++def pathway_boost(abs_d: np.ndarray, M, k: int):+    """boost_g = (# top-k active sets containing g) / (# all sets containing g)."""+    s = (M.astype(np.float64) @ abs_d) / np.maximum(M.sum(axis=1), 1)+    k = min(int(k), M.shape[0])+    act = np.argsort(-s, kind="stable")[:k]+    A = M[act]  # active set membership+    n_active = A.astype(np.float64).sum(axis=0)+    n_total = M.astype(np.float64).sum(axis=0)+    boost = np.where(n_total > 0, n_active / np.maximum(n_total, 1.0), 0.0)+    return boost, act+++def pathway_boost_graded(abs_d: np.ndarray, M, k: int):+    """Graded set-activity boost in [0, 1] with full dynamic range.++    Set activity s_p = mean(|Δ_g|, g∈p). Sets are soft-weighted by the rank of+    s_p: the top-k sets get weight linearly decreasing from 1 (rank 0) to 0+    (rank k), the rest 0. boost_g = Σ_p w_p·M_pg / Σ_p M_pg — the mean active+    weight of the sets gene g belongs to. Genes in no set get 0.+    """+    s = (M.astype(np.float64) @ abs_d) / np.maximum(M.sum(axis=1), 1)+    k = min(int(k), M.shape[0])+    order = np.argsort(-s, kind="stable")+    w = np.zeros(M.shape[0])+    w[order[:k]] = np.linspace(1.0, 0.0, k, endpoint=False)+    Mf = M.astype(np.float64)+    if PBPROJ_BOOST_MODE == "graded_max":+        # best-set attribution: boost_g = max weight among sets containing g+        boost = np.where(Mf > 0, w[:, None], 0.0).max(axis=0)+    else:+        num = Mf.T @ w+        den = Mf.sum(axis=0)+        boost = np.where(den > 0, num / np.maximum(den, 1.0), 0.0)+    return boost, order[:k]+++def project_delta(d: np.ndarray, M, names, k: int, eta: float, corr: np.ndarray | None):+    """Direction-correct one type's Δ with the prior. Returns (Δ', info)."""+    ad_ = np.abs(d).astype(np.float64)+    if M is not None:+        boost, _ = pathway_boost(ad_, M, k)+        dp = (d.astype(np.float64) * (1.0 + eta * boost)).astype(np.float32)+        cos = float(d @ dp / max(np.linalg.norm(d.astype(np.float64)) * np.linalg.norm(dp.astype(np.float64)), 1e-12))+        info = {"boost_mean": float(boost.mean()), "boost_max": float(boost.max()),+                "cos": cos, "mode": "pathway"}+        return dp, info+    if corr is not None:+        d64 = d.astype(np.float64)+        dp = d64 + eta * (corr.astype(np.float64) @ d64)+        if PROJ_SUPPORT > 0:+            # keep the diffused value only on reliable DE genes; genes with+            # |Δ| < support keep their measured Δ (no recruited noise genes,+            # no support/sign changes outside the DE set)+            dp = np.where(np.abs(d64) >= PROJ_SUPPORT, dp, d64)+        if PROJ_NORM:+            dp = dp * (np.linalg.norm(d64) / max(np.linalg.norm(dp), 1e-12))+        dp = dp.astype(np.float32)+        cos = float(d @ dp / max(np.linalg.norm(d.astype(np.float64)) * np.linalg.norm(dp.astype(np.float64)), 1e-12))+        return dp, {"cos": cos, "mode": "corr"}+    return d, {"mode": "none"}++ def aniso_reshape(coords, aligned_a, aligned_b, t: float, damp: float):     """Rescale per-axis spreads of `coords` toward the log-interpolated bracket spreads."""     a = np.asarray(aligned_a, dtype=np.float64)@@ -265,7 +570,8 @@ def type_aniso_reshape(coords, labs_out, ca, cb, la_all, lb_all, t: float):     return c, info  -def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float):+def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: str | None = None,+                 proj_eta: float | None = None):     t = float(t)     damp = float(params.get("scale_damp", 1.0))     align = str(params.get("align", "procrustes"))@@ -297,8 +603,33 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float):         shared = sorted(k for k in set(means_a) & set(means_b) if cnt_a[k] >= 10 and cnt_b[k] >= 10)         delta = {}         n_part, n_hard = [], []+        # Prior-projection setup (family T2EI-01, this node): direction-correct+        # each type's raw Δ with Reactome pathway membership (or gene-gene+        # correlation fallback) before the soft weights and displacement.+        eta = PROJ_ETA if proj_eta is None else float(proj_eta)+        proj_M, proj_names, proj_corr = None, None, None+        proj_cos, proj_boost = [], []+        if eta != 0.0 and shared and view is not None:+            def _kernel():+                if PROJ_CORR_KERNEL == "within":+                    k = corr_kernel_within(stage_a, stage_b)+                    if k is not None:+                        return k+                return corr_kernel(stage_a, stage_b, len(stage_a.genes))+            if PROJ_MODE == "pathway":+                proj_M, proj_names = load_pathway_sets(view, list(stage_a.genes))+                if proj_M is None:+                    proj_corr = _kernel()+            elif PROJ_MODE == "corr":+                proj_corr = _kernel()         for lab in shared:             d = (means_b[lab] - means_a[lab]).astype(np.float32)+            if eta != 0.0 and (proj_M is not None or proj_corr is not None):+                d, pinfo = project_delta(d, proj_M, proj_names, PROJ_K, eta, proj_corr)+                if "cos" in pinfo:+                    proj_cos.append(pinfo["cos"])+                if "boost_mean" in pinfo:+                    proj_boost.append(pinfo["boost_mean"])             if SOFT_ENABLE and SOFT_P > 0:                 w = np.clip(np.abs(d) / max(SOFT_TAU, 1e-12), 0.0, 1.0) ** SOFT_P                 n_part.append(int((w > 0.01).sum()))@@ -307,6 +638,14 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float):             elif CONVERGE_MIN_DELTA > 0:                 d = np.where(np.abs(d) >= CONVERGE_MIN_DELTA, d, np.float32(0.0)).astype(np.float32)             delta[lab] = d+        if eta != 0.0 and (proj_M is not None or proj_corr is not None):+            conv["proj_eta"] = eta+            conv["proj_mode"] = "pathway" if proj_M is not None else "corr"+            conv["proj_k"] = PROJ_K if proj_M is not None else None+            conv["proj_n_pathways"] = int(proj_M.shape[0]) if proj_M is not None else None+            conv["proj_cos_mean"] = float(np.mean(proj_cos)) if proj_cos else None+            conv["proj_cos_min"] = float(np.min(proj_cos)) if proj_cos else None+            conv["proj_boost_mean"] = float(np.mean(proj_boost)) if proj_boost else None         if n_part:             conv["soft_enable"] = True             conv["soft_tau"] = SOFT_TAU@@ -538,6 +877,89 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float):             )      expr = np.clip(np.vstack([xa, xb]), 0.0, None).astype(np.float32) if (ia.size or ib.size) else np.zeros((0, len(stage_a.genes)), np.float32)+    clamp_info = {"bracket_clip": BRACKET_CLIP, "n_clamped_cells": 0}+    if BRACKET_CLIP > 0 and expr.shape[0]:+        mx_a = np.asarray(stage_a.X.max(axis=0).todense()).ravel().astype(np.float64)+        mx_b = np.asarray(stage_b.X.max(axis=0).todense()).ravel().astype(np.float64)+        cap = (BRACKET_CLIP * np.maximum(mx_a, mx_b)).astype(np.float32)+        n_over = int((expr > cap[None, :]).any(axis=1).sum())+        expr = np.minimum(expr, cap[None, :]).astype(np.float32)+        clamp_info["n_clamped_cells"] = n_over+    tshift_info = {"tshift_enable": bool(TSHIFT_ENABLE), "tshift_kappa": TSHIFT_KAPPA,+                   "tshift_f_mean": None, "tshift_f_min": None, "tshift_f_max": None,+                   "tshift_n_moved": 0, "tshift_dp_top50_changed": None,+                   "tshift_timing_absmean": None}+    if TSHIFT_ENABLE and expr.shape[0] and ia.size and ib.size:+        if TSHIFT_MODE == "leakfix":+            # No timing model: f_g = t for every reliable gene — removes the+            # nnz-mask/clip leakage of the displacement pipeline so that dp+            # equals the clean linear-interpolation estimate t·Δ_pb.+            f_g = np.full(len(stage_a.genes), float(t))+        else:+            la_all = np.asarray(stage_a.labels).astype(str)+            lb_all = np.asarray(stage_b.labels).astype(str)+            shared_lab = sorted(set(la_all.tolist()) & set(lb_all.tolist()))+            timing = timing_fraction(stage_a, stage_b, shared_lab, TSHIFT_KAPPA)+            f_g = np.clip(t + TSHIFT_KAPPA * timing, TSHIFT_FLO * t, min(TSHIFT_FHI * t, 0.9))+        pb_a = np.asarray(stage_a.X.mean(axis=0)).ravel().astype(np.float64)+        pb_b = np.asarray(stage_b.X.mean(axis=0)).ravel().astype(np.float64)+        dpb = pb_b - pb_a+        pb_out = expr.mean(axis=0).astype(np.float64)+        top_before = set(np.argsort(-np.abs(pb_out - pb_a))[:50].tolist())+        target_pb = pb_a + f_g * dpb+        s = np.clip(target_pb / np.maximum(pb_out, PBPROJ_FLOOR), 1.0 - TSHIFT_CAP, 1.0 + TSHIFT_CAP)+        s = np.where(np.abs(dpb) >= TSHIFT_MIN_DELTA, s, 1.0)+        expr = (expr * s.astype(np.float32)[None, :]).astype(np.float32)+        dp_after = expr.mean(axis=0).astype(np.float64) - pb_a+        top_after = set(np.argsort(-np.abs(dp_after))[:50].tolist())+        moved = np.abs(dpb) >= TSHIFT_MIN_DELTA+        tshift_info.update(+            tshift_f_mean=float(f_g[moved].mean()) if moved.any() else None,+            tshift_f_min=float(f_g[moved].min()) if moved.any() else None,+            tshift_f_max=float(f_g[moved].max()) if moved.any() else None,+            tshift_n_moved=int(moved.sum()),+            tshift_mode=TSHIFT_MODE,+            tshift_timing_absmean=(float(np.abs(timing[moved]).mean())+                                   if (moved.any() and TSHIFT_MODE != "leakfix") else None),+            tshift_dp_top50_changed=int(50 - len(top_before & top_after)),+        )+    pbproj_info = {"pbproj_enable": bool(PBPROJ_ENABLE), "pbproj_eta": PBPROJ_ETA,+                   "pbproj_k": PBPROJ_K, "pbproj_src": PBPROJ_SRC,+                   "pbproj_n_boosted": 0, "pbproj_s_min": None, "pbproj_s_max": None,+                   "pbproj_s_mean": None, "pbproj_dp_top50_changed": None}+    if PBPROJ_ENABLE and PBPROJ_ETA != 0.0 and expr.shape[0] and view is not None:+        pb_a = np.asarray(stage_a.X.mean(axis=0)).ravel().astype(np.float64)+        pb_out = expr.mean(axis=0).astype(np.float64)+        dpb_out = pb_out - pb_a+        top_before = set(np.argsort(-np.abs(dpb_out))[:50].tolist())+        w = None+        if PBPROJ_W_MODE == "tf":+            w = tf_indicator(view, list(stage_a.genes))+            if w is not None:+                pbproj_info["pbproj_w_mode"] = "tf"+        else:+            M, names = load_pathway_sets(view, list(stage_a.genes), PBPROJ_SRC)+            if M is not None:+                pb_b = np.asarray(stage_b.X.mean(axis=0)).ravel().astype(np.float64)+                dpb = pb_b - pb_a+                if PBPROJ_BOOST_MODE in ("graded", "graded_max"):+                    w, _ = pathway_boost_graded(np.abs(dpb), M, PBPROJ_K)+                else:+                    w, _ = pathway_boost(np.abs(dpb), M, PBPROJ_K)+                pbproj_info["pbproj_w_mode"] = "pathway"+                pbproj_info["pbproj_n_pathways"] = int(M.shape[0])+        if w is not None:+            corr = PBPROJ_ETA * w * dpb_out / np.maximum(pb_out, PBPROJ_FLOOR)+            s = np.clip(1.0 + corr, 0.0, 1.0 + PBPROJ_CAP)+            expr = (expr * s.astype(np.float32)[None, :]).astype(np.float32)+            top_after = set(np.argsort(-np.abs(expr.mean(axis=0).astype(np.float64) - pb_a))[:50].tolist())+            pbproj_info.update(+                pbproj_n_boosted=int((w > 0).sum()),+                pbproj_boost_max=float(w.max()),+                pbproj_s_min=float(s.min()), pbproj_s_max=float(s.max()),+                pbproj_s_mean=float(s.mean()),+                pbproj_dp_top50_changed=int(50 - len(top_before & top_after)),+            )     coord_parts = []     if ia.size:         coord_parts.append(np.asarray(ca[ia], dtype=np.float64))@@ -562,6 +984,9 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float):     info.update(aniso_info)     info.update(ta_info)     info.update(variso_info)+    info.update(pbproj_info)+    info.update(tshift_info)+    info.update(clamp_info)     info.update(         t=t, n=int(expr.shape[0]), rms_a=rms_a, rms_b=rms_b, target_rms=target_rms,         out_rms=rms_radius(coords), n_from_a=int(ia.size), n_from_b=int(ib.size), **conv, **pair_info,@@ -574,7 +999,11 @@ 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("--ablate", default=None,+                        help="mechanism-off control: 'mechanism' (or any name) sets PROJ_ETA=0, "+                             "reproducing the parent node bit-for-bit")     args = parser.parse_args()+    proj_eta = 0.0 if args.ablate else None      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)@@ -588,13 +1017,24 @@ def main() -> None:     stage_a = read_stage(args.data, a, genes)     stage_b = read_stage(args.data, b, genes)     params = board_params(manifest, "mix", PARAMS, args.seed)-    expr, coords, info = mix_converge(stage_a, stage_b, t, params, CONVERGE_ALPHA)+    expr, coords, info = mix_converge(stage_a, stage_b, t, params, CONVERGE_ALPHA,+                                      view=args.data, proj_eta=proj_eta)     keep = {k: info.get(k) for k in ("t", "n", "rms_a", "rms_b", "out_rms", "n_shared_types", "z_dot", "z_flipped",                                      "align", "alpha", "n_shared_types_converged", "within_type_std_rel",                                      "between_type_mean_dist_rel", "n_from_a", "n_from_b",                                        "lambda", "weight_mean", "weight_std",-                                        "soft_enable", "soft_tau", "soft_p",-                                        "soft_n_participating_mean", "soft_n_hard_mean",+                                         "soft_enable", "soft_tau", "soft_p",+                                         "soft_n_participating_mean", "soft_n_hard_mean",+                                        "proj_eta", "proj_mode", "proj_k", "proj_n_pathways",+                                        "proj_cos_mean", "proj_cos_min", "proj_boost_mean",+                                        "pbproj_enable", "pbproj_eta", "pbproj_k", "pbproj_src",+                                        "pbproj_n_pathways", "pbproj_n_boosted", "pbproj_boost_max",+                                        "pbproj_s_min", "pbproj_s_max", "pbproj_s_mean",+                                        "pbproj_dp_top50_changed",+                                        "tshift_enable", "tshift_kappa", "tshift_f_mean",+                                        "tshift_f_min", "tshift_f_max", "tshift_n_moved",+                                        "tshift_mode", "tshift_timing_absmean", "tshift_dp_top50_changed",+                                        "bracket_clip", "n_clamped_cells",                                        "pair_beta", "n_paired_types", "n_pairs", "pair_within_std_rel",                                        "pair_orth_var_rel", "pair_dist_over_diam",                                        "aniso_enable", "aniso_damp", "aniso_gamma", "aniso_spread_a",

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: defaults reproduce the scored 62.06 config, `--ablate mechanism` reproduces parent bit-for-bit, deterministic, view-independent (disguised-view test passed), vec-check ok on seeds 0/1/2.

**做了什么**：实现 PLAN 的通路/相关先验投影后发现结构性对消——分层混抽的 (1−t,t) 权重使 Δ 位移在伪批量中精确抵消，de_score/de_direction 只看 dp，故通路 boost（8 个离线配置，top-50 |dp| 变化 ≤2 基因）按 PLAN 自己的筛选线无效（存档查分 60.86，劣于父节点）。改交同族备选解码：范数保持 + 支持集限制（|Δ|≥0.15）的基因相关扩散 Δ'=renorm(Δ+C·Δ)，只转方向、不爆幅度、不招募噪声基因。

**查分结果**（A 半 proxy，19/20 次）：提交配置 seed 0 = **62.06**（父 61.48，+0.58），seed 1 配对 = 61.55 vs 60.93（+0.63）。收益：neighborhood_mmd +0.40、de_score raw 0.25→0.2857、variogram +0.12~0.19；微损 mmd_u −0.08、de_dir −0.07。中间共证否 7 个变体（TF 重排两向、tshift 时序、leakfix、pb 级通路放大、无范数 corr、clamp 组合），全部记录在 METHOD.md §5。

**下一步最值得试**：(1) 相关扩散的不对称解码（只对 b 侧或用位移后残差核迭代一步），neighborhood 是最大收益来源；(2) clip 感知扩散以挽回 mmd_u 微损；(3) de_score 0.2857 疑似局部上限，勿再投 dp 重排。

调研员

```json
{
  "name": "通路/相关先验投影校正逐类型位移方向攻 de_score top-N 排序",
  "motivation": "父节点 27 的 ANALYSIS 明确指出 de_score（raw 0.3103，skill 0.600）缺口不在幅度而在位移方向的基因级错位：α·Δ_type 与真值 dt 的基因排序不一致，导致 |dp| top-N 集合与 |dt| top-N 重合率低。软阈值（节点 27）把参与基因从 166 扩到 478 但 de_score 纹丝不动，证实问题在方向而非阈值。expression_change 组 62.16 是仅次于 cell_state（50.36）的第二弱组，de_score 是该组内最弱指标（7.50/12.5）。",
  "approach": "步骤：1) Engineer 先检查视图 prior/ 目录是否有通路/基因集文件（Reactome、CollecTRI、GO 等）；若有，构建通路×基因二元矩阵 M；若无，用括号两阶段全部细胞计算基因-基因 Pearson 相关矩阵 C（~498×498）作为平滑核。2) 对每个共有类型的 Δ=μ_b−μ_a：(a) 通路版：计算通路活性分 s_p=mean(|Δ_g|, g∈p)，取 top-k（k=30）活性通路，对每个基因 g 计算 boost_g = (g 所属活性通路数)/(g 所属总通路数)，Δ'_g = Δ_g·(1+η·boost_g)；(b) 相关版（fallback）：Δ' = Δ + η·C·Δ（一步扩散），然后对 Δ' 做 clip 保持 nnz-only 语义（只移已表达项）。3) 用 Δ' 替换原 Δ 进入既有位移步（α=5、λ=6、β=0.2、软权重均保留不变）。4) η 初值 1.0，搜索 {0.5, 1.0, 2.0}；通路版 k 搜索 {20, 30, 50}。5) 快速筛选：先在 seed 0 上比较 off/on 的 |dp| top-50 基因集合重合率变化（若重合率变化 <2 个基因则方向无效，尽早放弃）；再 vec-score 查分确认。6) 单输入阶段（无括号）走既有 copy_last 退路，校正代码不执行。7) 不新增随机源，确定性输出。",
  "expected_groups": ["expression_change"],
  "risks": "1) prior/ 为空或无可用通路文件 → Engineer 第一步检查，立即切换到相关矩阵 fallback；2) 通路/相关校正改变不了 top-N 集合（Δ 的排序本身已接近最优）→ 用 top-50 重合率快速判断，若 <2 个基因变化则停止；3) η 过大导致位移方向过度偏离 Δ、mmd_u 恶化 → 否决线 mmd_u raw 增幅 <0.001，η=2 时若破线则放弃；4) 相关矩阵在小样本类型上噪声大 → 只用 ≥10 细胞的类型，相关矩阵用全类型合并计算；5) 效应量在噪声（~1 分）内 → 若 A 半提升 <0.5 分且 top-N 重合率无变化，如实报告失败。",
  "family_id": "T2EI-01",
  "mechanism": "用通路成员关系（或基因-基因相关矩阵）对逐类型伪时间差 Δ 做方向投影/平滑，使位移方向与生物学上协调变化的基因簇对齐，从而改变 |dp| 的 top-N 排序集合。",
  "vs_constant_shift": "常数位移对所有基因施加同一方向同一幅度；本机制按通路/相关结构对每个基因施加不同的方向校正权重（boost_g 因基因而异），且校正方向取决于该基因所属通路的集体活性，不是全局均匀缩放或组成重加权。",
  "mechanism_evidence": "1) 比较 off/on 下 |dp| top-50 基因集合：报告进入/退出 top-50 的基因数及身份；2) de_score raw 变化（目标 ≥+0.02）；3) 逐类型 Δ 与 Δ' 的余弦相似度分布（应 <1 但 >0.5，证明方向改变但不过度）；4) 四组分各自变化，确认 cell_state/local_spatial/shape_scale 无恶化。",
  "mechanism_off_control": "η=0（环境变量 T2_PROJ_ETA=0）时 Δ'=Δ，输出与父节点 27 逐位一致。Engineer 用 seed 0 验证 off 输出与父节点 A 半分数（62.55 或 A 半 61.48）完全匹配。",
  "sources": []
}
```

审查员

REVIEW.json written