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

节点 n48

UNIVAL：单侧独有类型细胞的表达值沿共有类型估计的组织阶段效应前移/回退（池化伪批量精确中和保 DE），叠加 WIRESET 型×侧抽样漂移均值回正；PLAN 的 VDEF 方差放气经 5 配置证否。

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 2f00188b46: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 e712a04d58。改动的文件：solution/METHOD.md +108 −86、solution/README.md +28 −0、solution/run.py +397 −8

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex df6850a..531a2d9 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,86 +1,108 @@-NO_CHANGE: occupancy_dice 坐标探针（ANISO2 本征轴比 5 档 α、RDENS 径向密度 4 组全为净负或平手，d2_shape 以 20–40 倍亏损率对冲）--父节点：45（= 42 的输出，A 半 seed 0 = 64.622）。本节点权威输出与父节点**逐位一致**-（X data/indices/indptr + obsm 的 sha256 前缀 `da6d03e8e995e85d`，seed 0；`--ablate mechanism`-输出相同 → mechanism_active=no）。ANISO2 与 RDENS 两段探针代码保留在 run.py 里（环境变量开关，-默认关），供后续节点查证，不影响输出。--## PLAN 机制（ANISO2，family T2EI-06）：实现与诊断--在输出云自身 PCA 框架里做全局对角缩放：目标 = 括号两端云本征值分数（旋转不变，直接从-`stage_a.coords` / `stage_b.coords` 现场算）的 t-对数插值；每轴乘 `(ft_i/fo_i)^α`（clip [0.5,2]），-再均匀缩放恢复探针前的精确 RMS（scale_log_ratio 逐位不动）；只动坐标，表达不动；单输入阶段-（无括号）自动跳过。诊断确认 PLAN 的前提成立：输出 eigfrac `[0.497, 0.317, 0.187]` vs-t-插值目标 `[0.606, 0.251, 0.143]`（log 偏差 `[+0.17, −0.26, −0.29]`）——现有 merged-frame-aniso（node 13，γ=0.5）只对齐了合并框架里的逐轴 spread，输出云自己的本征轴比仍偏离目标。-机制确实改变了坐标（α=0.15 时轴因子 `[1.030, 0.966, 0.961]`，eigfrac → `[0.530, 0.297, 0.173]`；-α=0.70 → `[0.647, 0.225, 0.127]`）。--## 查分记录（proxy A 半，seed 0，共 13 次查询；points = 该指标得分，满分为括号内数字）--基线（父 45）：board **64.622**；de_score 7.81(12.5)、de_direction 7.86(12.5)、mmd_u 7.83(12.5)、-variogram 6.70(12.5)、d2_shape 8.16(8.33)、occupancy_dice 3.51(8.33)、scale_log_ratio 7.77(8.33)、-neighborhood_mmd 14.98(25)。occ_raw=.8066，d2_raw=.0047，nbhd_raw=.05126。--| 配置 | board | d2 | occ | nbhd | de_s/de_d | mmd/var | occ_raw | d2_raw | nbhd_raw |-|---|---:|---:|---:|---:|---|---|---:|---:|---:|-| ANISO2 α=−0.15 | 64.010 | 8.14 | 2.91 | 14.99 | 不变 | 不变 | .7628 | .00487 | .05117 |-| ANISO2 α=0.05 | 64.350 | 7.88 | 3.51 | 14.98 | 不变 | 不变 | .8066 | .00689 | .05121 |-| ANISO2 α=0.15 | 63.773 | 7.26 | 3.56 | 14.98 | 不变 | 不变 | .8094 | .01235 | .05125 |-| ANISO2 α=0.40 | 62.239 | 5.82 | 3.48 | 14.96 | 不变 | 不变 | .8048 | .02961 | .05137 |-| ANISO2 α=0.70 | 61.256 | 4.52 | 3.83 | 14.94 | 不变 | 不变 | .8241 | .05462 | .05160 |-| RDENS 1.03→0.97 | 64.615 | 8.22 | 3.47 | 14.96 | 不变 | 不变 | .8038 | .00423 | .05146 |-| RDENS 1.045→0.955 | 64.129 | 7.81 | 3.40 | 14.94 | 不变 | 不变 | .8000 | .00749 | .05156 |-| RDENS 1.05→0.95 | 63.989 | 7.67 | 3.41 | 14.94 | 不变 | 不变 | .8000 | .00864 | .05159 |-| RDENS 0.97→1.03 | 63.545 | 7.38 | 3.21 | 14.98 | 不变 | 不变 | .7870 | .01130 | .05122 |-| cfgA 跨谱系移植* | 63.594 | 8.22 | 3.47 | 15.22 | 7.81/7.96 | 7.69/5.46 | .8038 | .00423 | .04926 |-| DETRX s=0.85* | 64.666 | 8.22 | 3.47 | 14.89 | 7.81/7.83 | 7.81/6.86 | .8038 | .00423 | .05201 |-| DETRX s=1.0* | 64.642 | 8.22 | 3.47 | 14.82 | 7.81/7.78 | 7.80/6.98 | .8038 | .00423 | .05262 |--\* 这三行跑时 RDENS(1.03→0.97) 默认为开（形状组数字含其 +0.06/−0.05 的 d2/occ 效应）；-DETRX 两行相对 RDENS 单跑的净差 = +0.05 / +0.03（噪声内）。cfgA = 节点 43 配方在本代码库的移植-（RECAL γ=0、DETR s=1.5、DETRX s=0.45 池化、SIDE 关）。--## 结论（为什么判无效）--1. **ANISO2**：occupancy_dice 对轴比修正的响应弱且非单调（raw 最多 +0.0175，+0.32 pts，α=0.70），-   d2_shape 同向恶化 20–40 倍（−3.64 pts）。PLAN 的两条护栏（Δnbhd_raw ≤ +0.0005、-   Δd2_raw ≤ +0.001）在所有 α>0.05 上都被击穿；α=0.05 时 occ 不动、d2 已 −0.28 pts。负 α-   （更等轴）occ 掉 0.6 pts。方向本身与 node 13 的 γ 网格一致（更 a-样各向异性买 occ、卖 d2），-   本征框架不改变这个交换率。-2. **RDENS（PLAN 步骤 4 备选）**：壳层压缩 (1.03→0.97) 是唯一近正配置（d2 +0.063、occ −0.045、-   nbhd −0.022 → 净 −0.007，平手）；更强压缩与反向膨胀都净负。局部 dice 代理（vs 括号云）在-   该方向 +0.007，但真实 occ 不跟（−0.003）——代理不可靠，已弃用。-3. **结构诊断**：occ 地板 ≈ raw .847（线性外推 skill .5），本谱系停在 .807。地板云是**单一真实-   阶段**的连贯切片，而 mix 输出是两阶段嵌合体，其占据体素集合近似两朵云归一化形状的**并集**-   （本地 16³ 统计：out 243 体素 vs a 169 / b 194）。全局仿射/径向形变无法把并集变成连贯中间形，-   occupancy_dice 的缺口对 mix 家族是结构性的；d2_shape (.979) 与 scale (.932) 已近天花板，-   任何大幅形状移动都在净亏。-4. **超出 PLAN 的两条附加测试**：(a) cfgA 移植证明本分支的 RECAL 幅度收缩是承重墙-   （去掉后 variogram .00755→.01107，跌破地板），不能照搬 43 谱系配方；(b) node 42 文档记录的-   DETRX 漂移余量（"s_ext=0.7 还差全漂移 30%"）：s_ext 0.85/1.0 使 variogram 单调改善-   （.00755→.00718→.00693）但 nbhd（.05126→.05201→.05262）与 de_direction（.3647→.3598→.3505）-   单调恶化，净 +0.03/+0.02，噪声内（T2 噪声 ≈1 分，父节点自身 seed 0/1/2 = 64.62/63.75/63.90）。-   检出通道的计数-选择张力再次封顶，与父节点 ANALYSIS 结论一致。--## 验证过的 / 没验证的--- 验证：默认（双探针关）输出与父逐位一致（seed 0）；`--ablate mechanism` 逐位一致；-  seed 0/1/2 均能运行且确定（`np.random.default_rng(seed)`，无全局随机）；vec-check ok；-  伪装视图（时间 +1 平移、文件重命名、manifest 键序打乱）digest 与真实视图逐位一致-  （view-independent）；运行时 ~2 s、内存 <1 GB（limits：28 GB / 30 min）；单输入退路-  （b=None 分支）不经过探针代码。-- 没验证：RDENS/ANISO2 在真实括号（E7.75+E8.25→E8.0，t=1/3）上的行为——两者在代理上即-  净负/平手，未提交，无需验证；occ 地板 raw（.847）是线性外推值，非实测。-- 知识来源：无（本节点全部为从视图输入现场计算的几何/统计探针，未使用任何外部生物学知识、-  文献数值或已发布阶段的测量值；目标 RMS、轴比、径向剖面全部由括号输入与 t 推得）。--## 给后续节点的建议--- 坐标通道对本谱系已封顶：全局各向异性（两个框架）、径向密度、per-type aniso（node 22）、-  体素再分布（node 24）、OT 重定位（方法卡）全部证否。occupancy_dice 的并集缺口只有-  「非嵌合」的坐标云能补，而那与 nbhd 的表达-位置配对不相容。不要再试全局仿射/径向形变。-- 检出通道同样封顶（本节点 s_ext 余量证否 + 父节点 45 分侧证否）。若还要提分，需要-  cell_state 的新信息源（如型内状态连续体重建），而不是现有计数的再校准。-- 树内最优是 43 谱系（65.93 B 半）；本谱系（42/45）与其差异主要在 RECAL 承重 + DETRX-SIDE，-  两者不可互换（cfgA 移植净 −1.03）。合并两谱系的增益需要逐件移植而非整配方照搬。+UNIVAL：单侧独有类型细胞的表达值沿共有类型估计的组织阶段效应前移/回退（池化伪批量精确中和保 DE），叠加 WIRESET 型×侧抽样漂移均值回正；PLAN 的 VDEF 方差放气经 5 配置证否。++# 节点 48（improve，父 = 46 = 45 = 42 谱系，board T2:embryo:val_interp）++## 提交内容++父管线（mix 混抽 + procrustes3d + 阻尼 log-RMS + α=5 逐类型收敛位移 + λ=6 投影加权 ++软阈值 + β=0.2 配对收缩 + 不对称相关扩散 + RECAL γ=0.25 + DETR s=2.0/NBHDCOH ++DETRX-SIDE s_ext=0.7）全部保留，在表达管线最末（DETR 之后）新增两个机制，坐标逐位不动：++1. **UNIVAL（主机制，`T2_UNIVAL_KAPPA=1.0` 默认开）**：单侧独有类型的**值级**时间对齐。+   输出里约 36%（proxy seed 0 实测 1815 个细胞）属于只在括号一侧出现的类型；父管线只用+   DETRX 校准了它们的**检出率**，表达**值**仍停在原始阶段水平，而真值中对应细胞处于中间+   状态。逐基因从共有类型（两侧各 ≥10 细胞）估计组织级阶段效应+   Δ̂(g) = Σ_c w_c·(pb_b(c,g) − pb_a(c,g)) / Σ_c w_c（w_c = min(两侧细胞数)，只用括号+   数据，无外部知识），软权重 w_g = clip(|Δ̂|/τ, 0, 1)（τ=0.25，与管线 DE 阈一致），+   截断 |Δ̂|≤cap=2.0。a 起源单侧细胞非零条目加 +κ·t·Δ̂·w_g；b 起源细胞负方向用乘性+   衰减 exp(−κ·(1−t)·Δ̂·w_g)（任何条目不会变成 0，支持集与检出计数逐位不变）；随后+   逐基因 g2 = pb_pre/pb_post **精确**复原池化伪批量 → de_score/de_direction 结构性不变+   （所有配置 raw .3571/.3647 逐位一致），只有组间质量分布（mmd_u/variogram/nbhd 可见）+   改变。+2. **WIRESET（次机制，`T2_WIRESET_KAPPA=1.0, CAP=0.4, EPS=0.03` 默认开）**：把节点 44+   的 WITHINP 思想移植到本谱系（该件此前未在本谱系测过；节点 46 证否的是整配方移植）。+   型×侧分组，在**原始阶段**上测得 drawn 均值相对全阶段均值的抽样漂移+   ρ_g = (μ_full+ε)/(μ_drawn+ε)（log 截断 ±cap），乘性作用于输出组非零条目（支持集不+   动），再逐基因精确池化伪批量中和（同样保 DE 通道）。++`--ablate mechanism`（或两个 env κ=0）→ 两机制均恒等，输出与父节点 46 逐位一致+（X digest `c541623c9e07eec5`，实测等于本会话 κ=0 基线运行）。++## PLAN 机制 VDEF 的证否记录（5 配置，6 次查分）++按 PLAN 实现：型×侧×基因，v_pre（输出 drawn 方差，含 DETR 零）对 v_tgt =+v_full,a^(1−t)·v_full,b^t（全阶段方差），仅 r=sqrt(v_tgt/v_pre)<1 时向组均值收缩+（跳过零条目、组均值精确回正、池化 pb 中和）。诊断确认前提成立：v_pre/v_tgt 中位比+1.60、25.2% 基因-组可压缩、f_mean 0.807、pb 漂移 ~1e-8、nnz 逐位不变——但 A 半 seed 0+全部劣于父（64.622）：++| 配置 | board | variogram | mmd_u | nbhd | de_score |+|---|---:|---:|---:|---:|---:|+| κ=1, lo=0.7 | 64.578 | .007559 | .01079 | .05141 | .3571 |+| κ=1, lo=0.7（重复查分，同文件） | 64.578 | 同上 | 同上 | 同上 | .3571 |+| κ=0.5, lo=0.7 | 64.602 | .007554 | .01074 | .05134 | .3571 |+| κ=0.25, lo=0.7 | 64.613 | .007551 | .01072 | .05130 | .3571 |+| κ=1, lo=0.5 | 64.555 | .007562 | .01085 | .05145 | .3571 |+| κ=1, lo=0.9 | 64.607 | .007553 | .01073 | .05132 | .3571 |++mmd_u、nbhd 随 κ 单调恶化，variogram 不降反微升 → 真值中间阶段的型内宽度**不低于**+混合云（PLAN 风险 1 命中），压缩反向。与节点 17/19（膨胀方向证否）合并结论：型内宽度+通道双向封闭。默认 `T2_VDEF_KAPPA=0`，代码保留。++## 查分记录（A 半 proxy，seed 0，父 = 64.622；共 20 次，额度用尽）++| # | 配置 | board | 关键 raw（variogram / mmd_u / nbhd） |+|---|---|---:|---|+| 1–6 | VDEF 5 档（见上表） | 64.555–64.613 | 全部劣化 |+| 7 | WIRESET κ1 cap0.1 eps0.1 | 64.660 | – / .01062 / .05113 |+| 8 | WIRESET κ1 cap0.2 eps0.1 | 64.677 | .007552 / .01059 / .05106 |+| 9 | WIRESET κ0.5 cap0.2 eps0.1 | 64.645 | .007552 / .01066 / .05117 |+| 10 | WIRESET κ1 cap0.4 eps0.1 | 64.691 | .007553 / .01056 / .05101 |+| 11 | WIRESET κ1 cap1.0 eps0.1 | 64.691 | 与 cap0.4 同（漂移已全在界内） |+| 12 | WIRESET κ1 cap0.4 eps0.3 | 64.675 | – / .01060 / .05106 |+| 13 | WIRESET κ1 cap0.4 eps0.03 | 64.699 | .007553 / .01054 / .05099 |+| 14 | **UNIVAL κ1.0** | **65.084** | .007374 / .01089 / **.04771** |+| 15 | UNIVAL κ1.5 | 65.036 | .007213 / .01209 / .04621 |+| 16 | UNIVAL κ2.0 | 64.815 | .007049 / .01372 / .04553 |+| 17 | UNIVAL κ0.7 | 64.997 | .007452 / .01054 / .04890 |+| 18 | **UNIVAL κ1 + WIRESET cap0.4 eps0.03（提交）** | **65.129** | .007357 / .01084 / **.04752** |+| 19 | 提交配置 seed 1 | 64.294 | .007674 / .01046 / .04803 |+| 20 | 提交配置 seed 2 | 64.439 | .007392 / .01057 / .04731 |++κ 扫描显示清晰的交换结构：κ↑ → nbhd/variogram 单调改善、mmd_u 单调恶化（单侧细胞移向+中间态使 15-NN 表达环境更连贯，但把部分细胞推离真值 PCA 里该类型的实际位置），κ=1.0+（线性时间权重，无自由调参）是净峰值。四组分（提交配置 vs 父）：local_spatial+61.75 vs ~60.0（nbhd .04752 vs .05126，本节点主要收益）；cell_state 组内 variogram+改善（.007357 vs .00755）、mmd_u 略差（.01084 vs ~.0101），组分基本持平；+expression_change 与 shape_scale 结构性不变（de_* raw 逐位一致，坐标逐位不动）。++## 机制生效证据（mechanism_active）++- UNIVAL：1815 个单侧细胞（36%）被移动，480 基因 |w_g|>0.05，step 均值 0.357；+  nnz 114241→114241 逐位不变；关掉后 X digest 回到父（`c541623c9e07eec5`）。+- WIRESET：42 组、g2 偏差 ≤0.105；单独 +0.077，与 UNIVAL 叠加再 +0.045。+- 两机制的 g2 中和保证池化 pb 精确不变 → DE 通道对所有配置 raw 逐位 .3571/.3647。++## 验证过 / 没验证++- 验证过：`vec-check` ok（seed 0/1）；seed 0 默认运行与查分配置 digest 逐位一致；+  `--ablate mechanism` 逐位还原父输出；伪装视图（时间 +1 平移、输入重命名、manifest+  键序重排）digest 与真实视图逐位一致 → 视图无关；单输入回退路径（b=None）在两机制+  之前返回，不触发；运行 ~3 s / <1 GB，纯 CPU（EXECUTION.json gpu=false）。+- 没验证：B 半分数（正式分）；seed 1/2 只查了提交配置本身（64.294/64.439），父配置+  在 seed 1/2 的 A 半分没有对照额度——跨 seed 绝对分不可比（混抽随 seed 变），机制+  方向的证据是 seed 0 配对差（+0.51）与三个 seed 上 nbhd/variogram raw 一致优于父+  seed 0 值；真实括号（两输入相邻、共有类型更多）上 Δ̂ 估计会更稳，但单侧类型占比+  可能不同，收益幅度未验证。+- 生物学知识来源：无外部知识——Δ̂、w_g、ρ_g 全部从视图括号输入现场计算（通用机制+  知识仅"组织内细胞随发育时间沿共享转录程序移动"，即谱系/阶段顺序意义上的教科书+  常识，未使用任何保留阶段的测量值）。++## 下一步建议++1. UNIVAL 的 κ-mmd_u 交换提示可做**类型感知幅度**：对真值中大概率仍存在的单侧类型+   （a 侧类型在 b 侧检出率仍高的）用大 κ，对快速消失的类型用小 κ——用 b 侧该类型的+   检出信号做权重，仍是括号内数据。+2. WIRESET 收益小但方向稳定，可试按组大小自适应 cap（小组漂移大、cap 应更小）。+3. nbhd 在 κ=1.5 仍改善而 mmd_u 崩：两指标对单侧细胞位置的偏好相反，值得用+   per-side 独立 κ（κ_a、κ_b）拆一下哪一侧贡献 mmd_u 损失。diff --git a/solution/README.md b/solution/README.mdindex 7c5214d..83bbd3d 100644--- a/solution/README.md+++ b/solution/README.md@@ -1,3 +1,31 @@+# 节点 48：UNIVAL（单侧类型值级时间对齐）+ WIRESET（型×侧抽样漂移回正）++PLAN 机制 VDEF（型×侧单边方差放气）已实现并在 A 半 seed 0 上 5 配置证否+（κ∈{0.25,0.5,1.0}×lo∈{0.5,0.7,0.9} → 64.555–64.613，全部低于父 64.622，+mmd_u/nbhd 随 κ 单调恶化；前提诊断成立：v_pre/v_tgt 中位比 1.60、25% 基因-组可压缩，+但真值型内宽度不低于混合云）。默认 `T2_VDEF_KAPPA=0`，代码保留。++提交备选机制（接在 DETR 之后、只动表达、坐标逐位不动、DE 通道经池化伪批量精确中和+结构性不变，de_score/de_direction raw 全配置逐位一致）：++1. **UNIVAL**（`T2_UNIVAL_KAPPA=1.0`）：36% 输出细胞属于单侧独有类型，父管线只校准了+   检出率（DETRX）、值停在原始阶段。沿共有类型细胞数加权的逐基因阶段效应 Δ̂(g)+   （软权重 clip(|Δ̂|/0.25,0,1)，截断 2.0），a 起源细胞非零条目 +t·Δ̂·w_g、b 起源+   负方向乘性衰减（支持集不变）。A 半 seed 0：65.084（nbhd .05126→.04771）；+   κ=0.7/1.5/2.0 → 64.997/65.036/64.815，κ=1 为峰值（κ↑ 时 nbhd/variogram 继续改善+   但 mmd_u 单调恶化）。+2. **WIRESET**（`T2_WIRESET_KAPPA=1.0, CAP=0.4, EPS=0.03`）：节点 44 WITHINP 思想移植——+   型×侧 drawn 均值（原始阶段上测）相对全阶段均值的抽样漂移用有界乘性因子回正。+   单独 64.699（cap/eps/κ 7 档网格全部 ≥64.64），与 UNIVAL 叠加 **65.129**（提交，+   父 +0.51）。seed 1/2：64.294/64.439（混抽 seed 方差，机制侧 raw 一致：nbhd+   .04803/.04731，variogram .007674/.007392）。++`--ablate mechanism` → 两机制恒等，输出与父节点 46 逐位一致（X digest+`c541623c9e07eec5`）；伪装视图（时间 +1、重命名、键序重排）digest 与真实视图逐位一致+（`9d822d386cf5a8d4`）。查分 20/20 用尽。详见 METHOD.md。++---+ # 节点 46：NO_CHANGE — occupancy_dice 坐标通道探针（ANISO2 + RDENS）已证否  PLAN 机制 ANISO2（输出云自身 PCA 框架里的全局各向异性比值修正，向括号本征值分数的 t-对数插值diff --git a/solution/run.py b/solution/run.pyindex 0ff5d32..cc8ced9 100644--- a/solution/run.py+++ b/solution/run.py@@ -8,6 +8,37 @@ 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 (48, family other): PLAN mechanism VDEF (one-sided per type×side+variance deflation toward the t-geometric full-stage variance target) was+implemented and FALSIFIED on the proxy A half (seed 0) across 5 amplitude/clip+configs: κ∈{0.25,0.5,1.0}×lo∈{0.5,0.7,0.9} → 64.555–64.613, ALL below parent+64.622, with variogram/mmd_u/neighborhood_mmd all monotonically worse in κ+(the premise held — v_pre/v_tgt median 1.60, 25% of gene-groups compressible —+but the truth's within-type width is not below the mixture's; DE channel+structurally untouched, de_score raw .3571 bit-identical everywhere). Default+T2_VDEF_KAPPA=0. SUBMITTED backup mechanisms (both new, both target the same+cell_state/local_spatial weakness):+1) UNIVAL (main gain): one-sided-type VALUE-level temporal alignment — the 36%+   of output cells whose type exists in only ONE bracket stage kept raw-stage+   expression values (only their detection rates were calibrated, by DETRX).+   Each such cell is advanced (a-origin, +κ·t·Δ̂·w_g) or regressed (b-origin,+   −κ·(1−t)·Δ̂·w_g, multiplicative decay on negatives) along the shared-type+   cell-count-weighted stage effect Δ̂(g), soft-weighted by clip(|Δ̂|/0.25,0,1),+   then a per-gene pooled-pseudobulk neutralisation restores pb EXACTLY (de_*+   invariant, support untouched). κ=1.0: board 65.084 (nbhd .05126→.04771,+   variogram .00755→.007374; κ=1.5/2.0 trade mmd_u against nbhd, net lower).+2) WIRESET (ports node 44's WITHINP idea to this lineage): type×side+   resampling-drift reset — drawn group means measured on the RAW stages are+   nudged back to the full-stage type means by bounded multiplicative factors+   (log-clip ±cap=0.4, eps=0.03, κ=1), pooled-pb neutralised. Alone +0.077+   (64.699); stacked with UNIVAL: 65.129 (best, submitted default).+Seeds 1/2 of the submitted config: 64.294/64.439 (mix-draw seed variance moves+all configs; mechanism-side raws stay improved: nbhd .04803/.04731 vs parent+.05126, variogram .007674/.007392). --ablate mechanism (or both env κ=0)+reproduces parent node 46 bit-for-bit (X digest c541623c9e07eec5); disguised+view (+1d time shift, renamed inputs, reordered manifest keys) gives the+identical digest (9d822d386cf5a8d4) → view-independent.+ This node (46, family T2EI-06): PLAN mechanism ANISO2 — global anisotropy-ratio probe in the output cloud's OWN PCA frame (scale each principal axis by (target_eigfrac_i/current_eigfrac_i)^alpha toward the t-log-interpolated bracket@@ -663,6 +694,322 @@ DETRSHRINK_MODE = os.environ.get("T2_DETRSHRINK_MODE", "plan") # bit-for-bit unchanged; only neighborhood_mmd can move. 0 = off (parent). SIDECLUSTER = int(os.environ.get("T2_SIDECLUSTER", "0")) +# VDEF (node 48, PLAN mechanism, family "other"): one-sided within-type+# variance deflation, applied at the very END of the pipeline (after DETR/+# DETRX). The mix output is, per type×side, a binary mixture of two real+# stages' cells plus DETR's zero-heavy detection thinning; both inflate the+# per-gene within-group variance above the t-interpolated target. For each+# (type × side) group and gene:+#   v_pre  = variance of the drawn OUTPUT values (zeros included — same+#            accounting as the DETR budget),+#   v_tgt  = v_full,a^(1−t) · v_full,b^t with v_full,side the per-gene+#            variance of that type in the FULL (unsampled) input stage of+#            that side (one-sided-only types: the single available v_full;+#            single-input views: skipped),+#   r_g    = sqrt(v_tgt / v_pre), applied ONLY where r_g < 1 (one-sided:+#            never inflate), clipped to [VDEF_LO, 1], factor+#            f = 1 − κ·(1 − r_clip),+#   x'     = μ_pre + f·(x − μ_pre) on NONZERO entries only (zeros skipped →+#            DETR support and detection counts untouched), then a per+#            group×gene multiplicative mean-fix over the nonzero entries+#            restores the group mean EXACTLY (VDEF_MEANFIX=1) so the pooled+#            pseudobulk — hence the whole DE channel — is structurally+#            neutral; a final per-gene g2 = clip(pb0/pb1, 0.98, 1.02) only+#            absorbs float32 drift.+# κ = 0 (default / --ablate mechanism) → identity, parent node 46 bit-for-bit.+VDEF_KAPPA = float(os.environ.get("T2_VDEF_KAPPA", "0"))+VDEF_LO = float(os.environ.get("T2_VDEF_LO", "0.7"))+VDEF_MIN_CELLS = int(os.environ.get("T2_VDEF_MIN_CELLS", "10"))+VDEF_MEANFIX = os.environ.get("T2_VDEF_MEANFIX", "1") == "1"+++def _full_type_vars(stage, min_cells=2):+    """Per-type per-gene population variance of a FULL input stage (csr)."""+    from scipy import sparse as sp+    X = stage.X.tocsr()+    X2 = X.copy()+    X2.data = X2.data.astype(np.float64) ** 2+    Xd = X.astype(np.float64)+    labs = np.asarray(stage.labels).astype(str)+    out = {}+    for c in np.unique(labs):+        m = labs == c+        k = int(m.sum())+        if k < min_cells:+            continue+        mean = np.asarray(Xd[m].mean(axis=0)).ravel()+        sq = np.asarray(X2[m].mean(axis=0)).ravel()+        out[c] = np.maximum(sq - mean * mean, 0.0)+    return out+++def vdef_deflate(expr, out_labels, stage_a, stage_b, t, n_from_a,+                 kappa, lo, min_cells, meanfix):+    """One-sided per (type×side) variance deflation toward the t-geometric+    full-stage target. See the VDEF_* config docs. κ ≤ 0 → identity."""+    info = {"vdef_kappa": float(kappa), "vdef_lo": float(lo),+            "vdef_min_cells": int(min_cells), "vdef_meanfix": bool(meanfix),+            "vdef_applied": False}+    n = expr.shape[0]+    if kappa <= 0.0 or n == 0 or out_labels is None or out_labels.shape[0] != n:+        return expr, info+    if n_from_a is None or n_from_a <= 0 or n_from_a >= n:+        return expr, info  # no bracket mixing → no binary-mixture inflation+    VFLOOR = 1e-8+    E = np.asarray(expr, dtype=np.float64)+    G = E.shape[1]+    is_a = np.zeros(n, dtype=bool)+    is_a[:n_from_a] = True+    var_a = _full_type_vars(stage_a)+    var_b = _full_type_vars(stage_b)+    n_groups = 0+    n_active = 0+    n_genes_tot = 0+    ratios = []+    fs = []+    post_ratios = []+    for c in np.unique(np.asarray(out_labels).astype(str)):+        in_a = var_a.get(c)+        in_b = var_b.get(c)+        if in_a is None and in_b is None:+            continue+        if in_a is not None and in_b is not None:+            v_tgt = np.exp((1.0 - t) * np.log(np.maximum(in_a, VFLOOR))+                           + t * np.log(np.maximum(in_b, VFLOOR)))+        else:+            v_tgt = in_a if in_a is not None else in_b+        for side_mask in (is_a, ~is_a):+            rows = (np.asarray(out_labels).astype(str) == c) & side_mask+            ng = int(rows.sum())+            if ng < min_cells:+                continue+            Xg = E[rows]+            mu = Xg.mean(axis=0)+            v_pre = Xg.var(axis=0)+            n_groups += 1+            n_genes_tot += G+            r = np.sqrt(np.maximum(v_tgt, 0.0) / np.maximum(v_pre, VFLOOR))+            active = (r < 1.0) & (v_pre > VFLOOR)+            if not active.any():+                continue+            rc = np.clip(r, lo, 1.0)+            f = np.where(active, 1.0 - kappa * (1.0 - rc), 1.0)+            nz = Xg != 0.0+            Y = np.where(nz, mu[None, :] + f[None, :] * (Xg - mu[None, :]), 0.0)+            np.maximum(Y, 0.0, out=Y)+            if meanfix:+                s_pre = Xg.sum(axis=0)+                s_post = Y.sum(axis=0)+                sc = np.ones(G)+                ok = s_post > 0.0+                sc[ok] = s_pre[ok] / s_post[ok]+                Y = np.where(nz, Y * sc[None, :], 0.0)+            E[rows] = Y+            ra = ratios if len(ratios) < 400 else None+            n_active += int(active.sum())+            fs.append(float(f[active].mean()))+            if ra is not None:+                ra.append((v_pre[active] / np.maximum(v_tgt[active], VFLOOR)).tolist())+            vp = Y.var(axis=0)+            post_ratios.append((vp[active] / np.maximum(v_tgt[active], VFLOOR)).tolist())+    out = E.astype(np.float32)+    # pooled pseudobulk safety net (near-identity by construction with meanfix)+    pb0 = np.asarray(expr, dtype=np.float64).mean(axis=0) if n else None+    pb1 = out.astype(np.float64).mean(axis=0)+    if pb0 is not None:+        g2 = np.ones_like(pb1)+        ok = pb1 > 1e-12+        g2[ok] = np.clip(pb0[ok] / pb1[ok], 0.98, 1.02)+        out = (out.astype(np.float64) * g2[None, :]).astype(np.float32)+        info["vdef_g2_dev_max"] = float(np.max(np.abs(g2 - 1.0)))+    allr = np.concatenate([np.asarray(x) for x in ratios]) if ratios else np.zeros(0)+    allp = np.concatenate([np.asarray(x) for x in post_ratios[:200]]) if post_ratios else np.zeros(0)+    info.update(+        vdef_applied=True, vdef_n_groups=n_groups, vdef_n_gene_groups=n_genes_tot,+        vdef_active_frac=float(n_active / max(n_genes_tot, 1)),+        vdef_vpre_over_vtgt_median=float(np.median(allr)) if allr.size else None,+        vdef_vpre_over_vtgt_p90=float(np.quantile(allr, 0.90)) if allr.size else None,+        vdef_f_mean=float(np.mean(fs)) if fs else None,+        vdef_f_min=float(np.min(fs)) if fs else None,+        vdef_vpost_over_vtgt_median=float(np.median(allp)) if allp.size else None,+        vdef_pb_drift_max=float(np.max(np.abs(pb1 - pb0) / np.maximum(pb0, 1e-9))) if pb0 is not None else None,+        vdef_nnz_before=int((np.asarray(expr) != 0).sum()),+        vdef_nnz_after=int((out != 0).sum()),+    )+    return out, info+++# WIRESET (node 48 backup mechanism; ports the WITHINP idea of node 44 to this+# lineage): the stratified mix DRAWS a subset of each stage's cells per type, so+# each (type × side) group's drawn per-gene mean carries resampling noise+# around the full-stage type mean. On the sibling branch, resetting that drift+# (node 44 WITHINP) was the only cell_state-positive change left untested on+# THIS branch. For each (type × side) group and gene:+#   ρ_g = (μ_full,g + ε)/(μ_drawn,g + ε) with μ_full the mean over ALL cells of+#   that type in that side's full input stage and μ_drawn the mean over the+#   drawn cells measured on the RAW stage (pre-pipeline, so only the sampling+#   component is targeted, never the temporal displacement), log-clipped to+#   ±cap; the group's OUTPUT nonzero entries are multiplied by ρ_g^κ (support+#   and coordinates untouched), then a per-gene pooled-pseudobulk+#   neutralisation g2 = pb_pre/pb_post restores the pooled pb EXACTLY, so the+#   whole DE channel (de_score, de_direction) is structurally invariant.+# κ = 0 (default / --ablate) → identity, parent node 46 bit-for-bit.+WIRESET_KAPPA = float(os.environ.get("T2_WIRESET_KAPPA", "1.0"))+WIRESET_CAP = float(os.environ.get("T2_WIRESET_CAP", "0.4"))+WIRESET_MIN_CELLS = int(os.environ.get("T2_WIRESET_MIN_CELLS", "10"))+WIRESET_EPS = float(os.environ.get("T2_WIRESET_EPS", "0.03"))+++def wireset_reset(expr, out_labels, stage_a, stage_b, ia, ib, n_from_a,+                  kappa, cap, min_cells, eps):+    """Type×side resampling-drift reset with exact pooled-pb neutralisation."""+    info = {"wireset_kappa": float(kappa), "wireset_cap": float(cap),+            "wireset_min_cells": int(min_cells), "wireset_applied": False}+    n = expr.shape[0]+    if kappa <= 0.0 or n == 0 or out_labels is None or out_labels.shape[0] != n:+        return expr, info+    if n_from_a is None or n_from_a <= 0 or n_from_a >= n:+        return expr, info+    E = np.asarray(expr, dtype=np.float64)+    G = E.shape[1]+    labs_out = np.asarray(out_labels).astype(str)+    is_a = np.zeros(n, dtype=bool)+    is_a[:n_from_a] = True+    drift_logs = []+    n_groups = 0+    for stage, rows_idx, side_mask in ((stage_a, ia, is_a), (stage_b, ib, ~is_a)):+        if rows_idx is None or len(rows_idx) == 0:+            continue+        Xs = stage.X.tocsr().astype(np.float64)+        Xd = Xs[rows_idx]                      # drawn cells, RAW stage+        labs_full = np.asarray(stage.labels).astype(str)+        labs_drawn = labs_full[rows_idx]+        for c in np.unique(labs_drawn):+            m_full = labs_full == c+            m_dr = labs_drawn == c+            if int(m_dr.sum()) < min_cells or int(m_full.sum()) < 2:+                continue+            mu_full = np.asarray(Xs[m_full].mean(axis=0)).ravel()+            mu_dr = np.asarray(Xd[m_dr].mean(axis=0)).ravel()+            log_rho = np.log((mu_full + eps) / (mu_dr + eps))+            log_rho = np.clip(log_rho, -cap, cap)+            rows = (labs_out == c) & side_mask+            ng = int(rows.sum())+            if ng < min_cells:+                continue+            n_groups += 1+            drift_logs.append(log_rho.tolist())+            f = np.exp(kappa * log_rho)+            Xg = E[rows]+            nz = Xg != 0.0+            E[rows] = np.where(nz, Xg * f[None, :], 0.0)+    out = E.astype(np.float32)+    pb0 = np.asarray(expr, dtype=np.float64).mean(axis=0)+    pb1 = out.astype(np.float64).mean(axis=0)+    g2 = np.ones_like(pb1)+    ok = pb1 > 1e-12+    g2[ok] = pb0[ok] / pb1[ok]+    out = (out.astype(np.float64) * g2[None, :]).astype(np.float32)+    if drift_logs:+        D = np.abs(np.concatenate([np.asarray(x) for x in drift_logs]))+        info.update(+            wireset_applied=True, wireset_n_groups=n_groups,+            wireset_abslogdrift_median=float(np.median(D)),+            wireset_abslogdrift_p90=float(np.quantile(D, 0.90)),+            wireset_abslogdrift_max=float(D.max()),+            wireset_g2_dev_max=float(np.max(np.abs(g2 - 1.0))),+            wireset_nnz_before=int((np.asarray(expr) != 0).sum()),+            wireset_nnz_after=int((out != 0).sum()),+        )+    return out, info+++# UNIVAL (node 48 backup mechanism 2, cell_state): one-sided-type VALUE-level+# temporal alignment. 36–37% of the output cells belong to types present in+# only ONE bracket stage; the pipeline calibrates their detection rates (DETRX)+# but leaves their expression VALUES at the raw stage level, while the truth's+# corresponding cells sit at the intermediate stage. For each gene, estimate+# the tissue-level stage effect Δ̂(g) as the cell-count-weighted mean of the+# shared-type deltas pb_b(c,g) − pb_a(c,g) (only bracket data; no external+# knowledge), soft-weighted w_g = clip(|Δ̂|/τ, 0, 1) (τ = 0.25, the pipeline's+# DE threshold). One-sided a-origin cells are advanced by +κ·t·Δ̂·w_g+# (additive on nonzero entries), b-origin cells regressed by −κ·(1−t)·Δ̂·w_g+# (multiplicative decay exp(s) on negative genes so no entry ever hits 0 —+# support and detection counts untouched), then a per-gene pooled-pseudobulk+# neutralisation g2 = pb_pre/pb_post restores the pooled pb EXACTLY (DE+# channel structurally invariant; only the between-group mass distribution+# moves). Shared-type cells are untouched except by the g2 column rescale.+# κ = 0 (default / --ablate) → identity, parent node 46 bit-for-bit.+UNIVAL_KAPPA = float(os.environ.get("T2_UNIVAL_KAPPA", "1.0"))+UNIVAL_TAU = float(os.environ.get("T2_UNIVAL_TAU", "0.25"))+UNIVAL_CAP = float(os.environ.get("T2_UNIVAL_CAP", "2.0"))+UNIVAL_MIN_CELLS = int(os.environ.get("T2_UNIVAL_MIN_CELLS", "10"))+++def unival_align(expr, out_labels, stage_a, stage_b, t, n_from_a,+                 kappa, tau, cap, min_cells):+    """Advance/regress one-sided-type cells along the shared-type stage effect."""+    info = {"unival_kappa": float(kappa), "unival_tau": float(tau),+            "unival_cap": float(cap), "unival_applied": False}+    n = expr.shape[0]+    if kappa <= 0.0 or n == 0 or out_labels is None or out_labels.shape[0] != n:+        return expr, info+    if n_from_a is None or n_from_a <= 0 or n_from_a >= n:+        return expr, info+    la = np.asarray(stage_a.labels).astype(str)+    lb = np.asarray(stage_b.labels).astype(str)+    shared = sorted(c for c in set(la.tolist()) & set(lb.tolist())+                    if int((la == c).sum()) >= min_cells and int((lb == c).sum()) >= min_cells)+    if not shared:+        return expr, info+    ma_a = {c: np.asarray(stage_a.X[la == c].mean(axis=0)).ravel().astype(np.float64) for c in shared}+    ma_b = {c: np.asarray(stage_b.X[lb == c].mean(axis=0)).ravel().astype(np.float64) for c in shared}+    w = np.array([float(min(int((la == c).sum()), int((lb == c).sum()))) for c in shared])+    D = np.stack([ma_b[c] - ma_a[c] for c in shared])           # (n_shared, G)+    dhat = (w[:, None] * D).sum(axis=0) / w.sum()               # (G,)+    wg = np.clip(np.abs(dhat) / max(tau, 1e-9), 0.0, 1.0)+    step = np.clip(dhat, -cap, cap) * wg                        # bounded stage effect+    E = np.asarray(expr, dtype=np.float64)+    labs_out = np.asarray(out_labels).astype(str)+    is_a = np.zeros(n, dtype=bool)+    is_a[:n_from_a] = True+    set_a, set_b = set(la.tolist()), set(lb.tolist())+    one_sided = np.array([c not in set_a or c not in set_b for c in labs_out])+    n_one = 0+    for side in (True, False):+        rows = one_sided & (is_a == side)+        k = int(rows.sum())+        if k == 0:+            continue+        n_one += k+        s = (kappa * t if side else -kappa * (1.0 - t)) * step+        Xg = E[rows]+        nz = Xg != 0.0+        pos = np.maximum(s, 0.0)+        dec = np.exp(np.minimum(s, 0.0))+        Y = np.where(nz, np.where(s[None, :] >= 0.0, Xg + pos[None, :], Xg * dec[None, :]), 0.0)+        np.maximum(Y, 0.0, out=Y)+        E[rows] = Y+    if n_one == 0:+        return expr, info+    out = E.astype(np.float32)+    pb0 = np.asarray(expr, dtype=np.float64).mean(axis=0)+    pb1 = out.astype(np.float64).mean(axis=0)+    g2 = np.ones_like(pb1)+    ok = pb1 > 1e-12+    g2[ok] = pb0[ok] / pb1[ok]+    out = (out.astype(np.float64) * g2[None, :]).astype(np.float32)+    info.update(+        unival_applied=True, unival_n_shared=len(shared), unival_n_one_sided_cells=n_one,+        unival_step_absmean=float(np.abs(step).mean()),+        unival_step_absmax=float(np.abs(step).max()),+        unival_n_genes_active=int((wg > 0.05).sum()),+        unival_g2_dev_max=float(np.max(np.abs(g2 - 1.0))),+        unival_nnz_before=int((np.asarray(expr) != 0).sum()),+        unival_nnz_after=int((out != 0).sum()),+    )+    return out, info+  def side_cluster_repair(coords, n_a):     """Deterministically re-pair expression rows to coordinate slots so that@@ -2225,6 +2572,8 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s     out_labels = np.concatenate([_la[ia] if ia.size else _la[:0],                                  _lb[ib] if ib.size else _lb[:0]])     info["out_labels"] = out_labels+    info["mix_ia"] = ia+    info["mix_ib"] = ib     aniso_info = {"aniso_enable": bool(ANISO_ENABLE), "aniso_damp": ANISO_DAMP}     if ANISO_ENABLE and coords.shape[0] >= 10:         coords, ai = aniso_reshape(coords, aligned_a, aligned_b, t, ANISO_DAMP)@@ -2259,13 +2608,13 @@ def main() -> None:     parser.add_argument("--out", required=True)     parser.add_argument("--seed", type=int, default=0)     parser.add_argument("--ablate", default=None,-                        help="mechanism-off control (node 46): 'mechanism' (or any name) disables "-                             "ANISO2 (global anisotropy-ratio probe in the output cloud's own PCA "-                             "frame) and RDENS (radial shell-density probe) — coordinates are "-                             "written exactly as the parent pipeline produced them, so the ablated "-                             "output reproduces parent node 45 bit-for-bit (both probes are also "-                             "off by default after falsification). Inherited mechanisms (NBHDCOH, "-                             "DETRX-SIDE, DETR, ...) stay ON in both runs")+                        help="mechanism-off control (node 48): 'mechanism' (or any name) disables "+                             "VDEF (one-sided per type×side variance deflation, κ→0 = identity), "+                             "WIRESET (type×side resampling-drift mean reset, κ→0 = identity) "+                             "and UNIVAL (one-sided-type value-level temporal alignment, κ→0 = identity); "+                             "ANISO2/RDENS coordinate probes stay off (falsified in node 46), so "+                             "the ablated output reproduces parent node 46 bit-for-bit. Inherited "+                             "mechanisms (NBHDCOH, DETRX-SIDE, DETR, RECAL, ...) stay ON in both runs")     args = parser.parse_args()     # Mechanism-off control (node 45): DETR-SIDE-SHARED (origin-split shared-type     # removal counts). This is the ONLY mechanism this node adds on top of parent@@ -2383,6 +2732,30 @@ def main() -> None:                 side_strength_a=DETR_SIDE_STRENGTH_A,                 side_strength_b=DETR_SIDE_STRENGTH_B)             info.update(detr_info)+    # VDEF (node 48, PLAN mechanism): one-sided within-type variance deflation+    # at the very end of the expression pipeline (after DETR/DETRX). --ablate+    # mechanism forces κ = 0 → identity → parent node 46 bit-for-bit.+    vdef_kappa = 0.0 if args.ablate else VDEF_KAPPA+    expr, vdef_info = vdef_deflate(expr, info.get("out_labels"), stage_a, stage_b,+                                   float(t), int(info.get("n_from_a", 0) or 0),+                                   vdef_kappa, VDEF_LO, VDEF_MIN_CELLS, VDEF_MEANFIX)+    info.update(vdef_info)+    # WIRESET (node 48 backup mechanism): type×side resampling-drift reset.+    # --ablate mechanism forces κ = 0 → identity → parent node 46 bit-for-bit.+    wireset_kappa = 0.0 if args.ablate else WIRESET_KAPPA+    expr, ws_info = wireset_reset(expr, info.get("out_labels"), stage_a, stage_b,+                                  info.get("mix_ia"), info.get("mix_ib"),+                                  int(info.get("n_from_a", 0) or 0),+                                  wireset_kappa, WIRESET_CAP, WIRESET_MIN_CELLS,+                                  WIRESET_EPS)+    info.update(ws_info)+    # UNIVAL (node 48 backup mechanism 2): one-sided-type value-level temporal+    # alignment. --ablate mechanism forces κ = 0 → identity → parent bit-for-bit.+    unival_kappa = 0.0 if args.ablate else UNIVAL_KAPPA+    expr, uv_info = unival_align(expr, info.get("out_labels"), stage_a, stage_b,+                                 float(t), int(info.get("n_from_a", 0) or 0),+                                 unival_kappa, UNIVAL_TAU, UNIVAL_CAP, UNIVAL_MIN_CELLS)+    info.update(uv_info)     sc_info = {"sidecluster": sidecluster}     if sidecluster and expr.shape[0]:         coords, sci = side_cluster_repair(coords, int(info.get("n_from_a", 0)))@@ -2485,7 +2858,23 @@ def main() -> None:                                           "detr_side_shared", "detr_ss_n_types",                                           "detr_ss_rem_a", "detr_ss_rem_b",                                           "detr_ss_rem_pool", "detr_ss_n_genes_gt_pool",-                                         "sidecluster", "sidecluster_n_a", "sidecluster_applied")}+                                         "sidecluster", "sidecluster_n_a", "sidecluster_applied",+                                         "vdef_kappa", "vdef_lo", "vdef_min_cells", "vdef_meanfix",+                                         "vdef_applied", "vdef_n_groups", "vdef_n_gene_groups",+                                         "vdef_active_frac", "vdef_vpre_over_vtgt_median",+                                         "vdef_vpre_over_vtgt_p90", "vdef_f_mean", "vdef_f_min",+                                         "vdef_vpost_over_vtgt_median", "vdef_pb_drift_max",+                                         "vdef_g2_dev_max", "vdef_nnz_before", "vdef_nnz_after",+                                         "wireset_kappa", "wireset_cap", "wireset_min_cells",+                                         "wireset_applied", "wireset_n_groups",+                                         "wireset_abslogdrift_median", "wireset_abslogdrift_p90",+                                         "wireset_abslogdrift_max", "wireset_g2_dev_max",+                                         "wireset_nnz_before", "wireset_nnz_after",+                                         "unival_kappa", "unival_tau", "unival_cap",+                                         "unival_applied", "unival_n_shared",+                                         "unival_n_one_sided_cells", "unival_step_absmean",+                                         "unival_step_absmax", "unival_n_genes_active",+                                         "unival_g2_dev_max", "unival_nnz_before", "unival_nnz_after")}     print(json.dumps({"bracket": [a["stage"], b["stage"]], **keep}, default=float), file=sys.stderr)     write_t2(args.out, expr, coords, genes, seed=args.seed) 

调研来源?调研员查到并用到的知识条目和文献检索结果（只列标题和编号）。

用到的知识库条目

分析结果?分析员写的 ANALYSIS.json：改了什么、各组分数怎么变、假设是否成立、经验和下一步建议。

对话摘要?每个角色和大模型对话的统计：轮数、工具调用、用时、token 数和最后的回答摘录；原始记录只给路径。

只给统计和最后回答的摘录；完整对话请到原始记录位置里列出的文件看。

分析员

ANALYSIS.json written

工程师

All checks pass. Final summary:

**做了什么**：按 PLAN 实现了 VDEF（型×侧单边方差放气），在 A 半 seed 0 上 5 配置全部证否（64.555–64.613 < 父 64.622，前提诊断成立但 mmd_u/nbhd 随 κ 单调恶化——真值型内宽度不低于混合云，与节点 17/19 合并成"宽度通道双向封闭"）。按规则提交备选机制：**UNIVAL**（单侧独有类型细胞的表达值沿共有类型估计的阶段效应 Δ̂ 前移 t·Δ̂ / 回退 (1−t)·Δ̂，支持集与池化伪批量精确不变）+ **WIRESET**（节点 44 WITHINP 思想移植：型×侧抽样漂移均值回正），均接在 DETR 之后，坐标不动。

**查分结果**（A 半 proxy，20/20 用尽）：UNIVAL κ=1 → 65.084（nbhd .05126→.04771）；κ 扫描 0.7/1.5/2.0 → 64.997/65.036/64.815，κ=1 峰值；WIRESET 单独 64.699；**叠加提交配置 65.129（父 +0.51）**；seed 1/2 = 64.294/64.439（混抽 seed 方差，机制侧 raw 一致改善）。de_score/de_direction 全配置逐位不变；`--ablate` 逐位还原父输出；伪装视图 digest 一致（视图无关）；vec-check 通过；~3 s / <1 GB 纯 CPU。

**下一步最值得试**：(1) UNIVAL 类型感知幅度——用 b 侧检出信号给"大概率仍存在"的单侧类型更大 κ，缓解 mmd_u 损失；(2) per-side 独立 κ_a/κ_b 拆解 mmd_u 与 nbhd 的相反偏好；(3) WIRESET cap 按组大小自适应。

调研员

标族）。机制 VDEF 接在父管线最末（WITHINP 之后），默认环境变量 T2_VDEF_KAPPA=0（关闭）。步骤：(1) 按型×侧分组，在输出值矩阵上算每组每基因的 drawn 均值 μ_pre 与方差 v_pre（含 DETR 零，与 DETR 预算同口径）。(2) 目标方差 v_tgt = (v_full,a)^(1−t)·(v_full,b)^t，v_full,side 用该型在该侧完整输入阶段（全部细胞、未抽样）的每基因方差——与 WITHINP 用全阶段均值回正同构，自动消掉抽样偏差；单输入阶段（无括号）直接跳过。(3) 单边因子 r_g = sqrt(v_tgt/v_pre)，仅当 r_g<1 时应用：κ=1 时 r 截断在 [0.7, 1.0]；r_g≥1 的基因不动（绝不膨胀）。(4) 应用 x' = μ_pre + κ-混合后的因子 × (x − μ_pre)，且对 DETR/DETRX 已置零的 (细胞,基因) 条目跳过——防止把零抬回正值、改变支持集与检出计数；因子由数据现算，视图无关。(5) 安全网：逐基因整列乘 g2_g = clip(pb0/pb1, 0.98, 1.02) 复原池化伪批量（组均值按构造不变，g2 应近恒等，仅作浮点漂移保险）。查分流程：先算诊断（输出 r_g<1 的基因-组占比、v_pre/v_tgt 中位比）确认前提；首轮只跑 κ=1.0 看方向（预期 variogram raw .007372 与 nbhd raw .04988 同向下降、de_score raw .3448 逐位不变）；方向正确再网格 κ∈{0.25, 0.5} 与截断下限 {0.7, 0.5}（≤6 配置，预算 20 次查分内）；若 mmd_u raw .01008 随 κ 变差而 variogram/nbhd 变好，取 mmd_u 恶化 < 0.3 分的折中 κ，否则弃。A 半领先 ≥1 分后用 seed 0/1/2 配对确认（T2 噪声约 1 分）。",
  "expected_groups": ["cell_state", "local_spatial"],
  "risks": "(1) 真值中间阶段的型内宽度可能不低于混合云（分化中间态更宽），则压缩反向伤 mmd_u——首轮 κ=1.0 立即看 mmd_u raw 方向，变差即弃，不要恋战；(2) 邻域项对型内方差操作历史上单调敏感（节点17），理论上方差收缩使 15-NN 均值分布更窄应利邻域，但设否决线 Δnbhd_raw ≤ +0.0005，击穿即弃；(3) 收益可能 <1 分噪声——必须 3 种子配对差同号才算数；(4) DE 通道被破坏——每配置核对 de_score raw 是否仍为 .3448、支持集基数是否逐位不变；(5) 收益被 WITHINP/RECAL 覆盖——若诊断显示 v_pre/v_tgt 中位比 ≈1（无可压缩量），直接弃，不做坐标或计数的补偿动作。",
  "family_id": "other",
  "mechanism": "对每个型×侧×基因，仅当混合云的 drawn 方差高于 t 几何插值的全阶段目标方差时，把细胞表达向该组 drawn 均值收缩（单边、有界、跳过 DETR 零），压缩二元混合与检出置零造成的型内宽度膨胀；组均值与支持集不动。",
  "vs_constant_shift": "常数位移给每个型加同一个逐基因偏移、只动伪批量均值；VDEF 的因子乘在 (x − μ) 上，每个细胞按其偏离组均值的程度被不同幅度地缩放，型均值与池化伪批量按构造不变（DE 通道结构性中性），改变的是分布的二阶结构（宽度、重尾）与基因间共变幅度，这是任何每型常数位移或全局组成重加权做不到的。",
  "mechanism_evidence": "Engineer 报告：(a) r_g<1 实际应用因子 ≠1 的基因-组占比与收缩量分布（占比过低说明前提不成立）；(b) 前后型内每基因方差比应整体移向 1；(c) 池化伪批量逐基因相对漂移 ~0、de_score raw 逐位 .3448、支持集基数不变；(d) 四组分分解：variogram raw 与 nbhd raw 是否随 κ 单调下降、mmd_u raw 的方向；(e) seed 0/1/2 配对差。",
  "mechanism_off_control": "T2_VDEF_KAPPA=0（或 --ablate mechanism）时因子 ≡1，代码路径退化为父管线，输出应与父节点 46（=42）逐位一致（digest 比对）；开启后输出改变且上述 (a)–(d) 量随之单调变化，证明机制在运行。",
  "sources": []
}
```

审查员

REVIEW.json written