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

节点 n50

UNIVAL-P：单侧位移的软基因权重加指数 P=3（位移集中到强阶段效应基因），mmd_u/variogram/邻域三项 raw 同时略优、榜分与父 3 种子打平；PLAN 的类型感知 κ 经 5 种解码 + 等均值对照证否（排序无关，只有平均幅度起作用）。

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 9a8c0ad76a: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 2f00188b46。改动的文件：solution/METHOD.md +56 −108、solution/run.py +325 −32

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 531a2d9..8f6aaf6 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,108 +1,56 @@-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。+UNIVAL-P：单侧位移的软基因权重加指数 P=3（位移集中到强阶段效应基因），mmd_u/variogram/邻域三项 raw 同时略优、榜分与父 3 种子打平；PLAN 的类型感知 κ 经 5 种解码 + 等均值对照证否（排序无关，只有平均幅度起作用）。++## 提交内容（family: other，父节点 48）++父 48 全部管线不变（mix + procrustes3d 对齐 + 阻尼 log-RMS + α=5 型内收敛位移 + λ=6 投影加权 + 软阈值 + β=0.2 配对收缩 + 不对称相关扩散 + RECAL + DETR + DETRX-SIDE + NBHDCOH + WIRESET + UNIVAL κ=1）。本节点唯一改动（提交默认）：++- **UNIVAL-P**：UNIVAL 位移的软基因权重从 `w_g = clip(|Δ̂_g|/τ,0,1)`（τ=0.25，P=1 时 480/498 基因全参与）改为 `w_g^P`，**P=3.0**。位移集中到强阶段效应基因，弱基因位移按三次方衰减。方向、截断（cap=2）、乘性衰减、池化伪批量中和（g2）全部不变 → DE 通道结构性不动（de_score .3571 / de_direction .3647 在所有 18 次查分中逐位一致）。+- 本节点新增的 PLAN 机制 UNIVAL-TA 及备选 MATCHDELTA、WIRESET-AC 都保留在代码里，**默认全部关闭**（`T2_UNIVAL_TA=0`、`T2_UNIVAL_MATCH=0`、`T2_WIRESET_ADAPT=0`），证否记录见下。++`--ablate mechanism`：关掉本节点全部新机制（TA、MATCH、ADAPT、P→1、κ_a=κ_b→1），UNIVAL/WIRESET 等父机制保持打开 → 输出与父 48 逐位一致（X digest `56f694d7d8a1a0b2`，本地已验证）。++## PLAN 机制 UNIVAL-TA：证否（步骤 1 诊断 + 5 解码 + 等均值对照）++步骤 1 诊断（不查分）：单侧类型的字面 n_other（对侧同名细胞数）**结构性为 0**（22/22 个类型，>80% 判据触发）→ 按 PLAN 切换到标记基因检出率软权重。诊断还显示持久度信号强不对称：a 侧 7 型的标记在对侧（E8.0）检出率比 d_other/d_in≈1.0–1.5（程序普遍仍在表达），b 侧 15 型只有 0.02–0.10（新出现的特化程序）。++查分记录（proxy A 半，seed 0，父=65.129；每项为 raw）：++| # | 配置 | 榜分 | mmd_u | variogram | nbhd |+|---|---|---:|---|---|---|+| q1 | TA ratio ref=0.1（a 侧 w=1，b 侧分级 0.26–0.97）| 65.009 | .01048 | .007430 | .04900 |+| q3 | 纯 per-side κ_a=1, κ_b=0 | 64.662 | .01065 | .007555 | .05104 |+| q4 | 纯 per-side κ_a=0, κ_b=1 | 65.116 | .01082 | .007362 | .04765 |+| q5 | TA ratio ref=0.1 + κ_b=1.5 | 65.101 | .01080 | .007334 | .04789 |+| q6 | 纯 per-side κ_b=1.5 | 65.084 | .01200 | .007195 | .04606 |+| q7 | 纯 per-side κ_b=2.0 | 64.912 | .01348 | .007031 | .04526 |+| q8 | **等均值对照**：uniform κ_b=0.62（=q1 的 b 侧平均权重）| 65.017 | .01044 | .007458 | .04895 |+| q9 | TA det ref=0.3（PLAN 字面退路信号）| 64.748 | .01053 | .007536 | .05065 |+| q10 | TA inv ref=0.2（反向假设：瞬态型多位移）| 65.064 | .01038 | .007447 | .04872 |++**判决性证据是 q1 vs q8**：分级 κ（按持久度排序）与同平均幅度的统一 κ，三项 raw 差 ≤3e-5、榜分差 0.008——持久度**排序不携带信息**，UNIVAL 的 mmd_u–nbhd 交换只由平均位移幅度驱动。q4/q3 进一步显示 b 侧回退承载全部收益（+0.42 vs UNIVAL 关闭的 64.699），a 侧前移净贡献 ≈0；q6/q7 显示 κ_b 放大时 mmd_u 的点损失约为 nbhd+variogram 收益的 2 倍。PLAN 假设（瞬态型的错误位移拖累 mmd_u）不成立：把瞬态型权重调低（q1/q9/q10）或调高（q10）都只沿同一条交换曲线滑动。++## 备选机制（同一弱项 cell_state/local_spatial）++1. **MATCHDELTA**（类型匹配方向）：单侧型的位移方向改为按表达谱相似度 softmax(β·Pearson) 加权的共有类型 Δ_c 混合（匹配合理：Blood Progenitor→HEM-Endoth、p-EXEM→ExEM-1、PHM/PAM→SOM）。q11（β=3,ρ=1,κ_b=1）65.047：nbhd .04723 为全家族最好，但 mmd_u .01126 更差；q12（+κ_b=1.5）64.959：匹配方向**不保护** mmd_u（.01263 vs 统一方向 .01200）；q14（范数匹配解码）65.049：证明 q11 的 nbhd 收益来自更长步长而非方向。→ 方向也不是 mmd_u 损伤源，证否。+2. **WIRESET-AC**（父 ANALYSIS 建议 3，组大小自适应强度 κ_g=κ·clip(n_g/40,0.25,1)）：q13 = 65.004，mmd_u .01049 改善但 nbhd 同量级损失，证否。+3. **UNIVAL-P**（提交）：见上。P=2（q15）与 P=3（q17）榜分都=65.129；P=3 三项 raw 全部略优于父（mmd_u .01084 vs .01089、variogram .007358 vs .007374、nbhd .04750 vs .04771），P=2 的 nbhd 略差（.04786），故选 P=3。P=2+κ_b=1.3（q16）65.126 亦平。++## 提交配置的机制生效证据++- **改变了哪些细胞**：1815 个单侧型输出细胞（854 a 起源 + 961 b 起源）的位移幅度逐基因改变：|Δ̂_g|≥0.25 的强基因 w_g=1 不变；|Δ̂_g|<0.25 的弱效应基因被三次方压缩（如 |Δ̂|=0.1：w 从 0.4 → 0.064；|Δ̂|=0.05：0.2 → 0.008）。其余 3185 个共有型细胞只受 g2 列缩放影响；坐标逐位不动。X digest 与父不同（`63636e2338b129ae` vs `56f694d7d8a1a0b2`），--ablate 恢复父 digest。+- **四组分变化（A 半 seed 0，vs 父）**：expression_change 62.69=62.69（逐位不变）；shape_scale 77.75=77.75（坐标不动）；cell_state：mmd_u 7.797（+.01）、variogram 6.783（+.005）；local_spatial：nbhd 15.440（+.026）。榜分 65.129 = 65.129。+- **种子稳健性**：seed 1 = 64.295（父 64.294）、seed 2 = 64.445（父 64.439）；机制 raw 与父差 ~1e-5（nbhd .04801/.04730 vs 父 .04803/.04731）。三种子均值 64.623 vs 父 64.621。++## 结论与下一步++本家族在 proxy A 半 seed 0 的天花板 ≈65.13：17 个配置（幅度、类型分级、per-side、方向匹配、基因选择五个自由度）全部落在同一条 mmd_u↔nbhd 交换曲线上，任何再分配净值不变。UNIVAL-P 是曲线上一个 raw 全面不劣于父的点，按规则提交备选机制而非退回父程序；预期 B 半 ≈ 父 ±0.1。**下一步最值得试的不在表达通道**：occupancy_dice skill .4214 仍低于地板（形状组唯一失分项），兄弟谱系节点 47 的 SIDEFRIM（b 侧稀疏边缘体素切除，shape_scale +0.26）值得移植到本谱系；以及 nbhd 与坐标配对的空间通道（表达-位置重配对，而非继续挤压值级位移）。++## 验证与合规++- 完整视图运行 ~11 s / <1 GB 内存（limits 30 min / 28 GB 内）；`vec-check` 通过；`EXECUTION.json {"gpu": false}`（纯 CPU）。+- **视图无关自检**：+1 天时间平移、输入文件改名、manifest 键序打乱的伪装视图上，seed 0 输出 X 与坐标 digest 与真实视图完全一致（`63636e2338b129ae`/`0bc0cb7bdce1aa58`）。程序不读 board/mode/路径/绝对时间，只用括号数据与相对 t。+- 单输入阶段退路：无括号时走父级 copy 路径（本节点未改动）；UNIVAL/TA/MATCH/AC 均只在括号存在时生效。+- 确定性：无全局随机；--seed 传入 mix 抽样 rng，P 机制本身无随机。+- **知识来源：无外部生物学知识**。所有量（标记基因、检出率、相似度、Δ̂、n_g）都在运行时从 view 的括号阶段现场计算；未使用保留阶段/保留基因型的任何测量值，未硬编码任何统计量。+- 查分用量：18/20（q1–q17 + q1 重复捕获全部分项）；剩余 2 次未用。+- 未验证：B 半与正式分（系统执行）；真实 final 括号（E 输入阶段数不同）上 UNIVAL-P 的行为仅由视图无关性与单输入退路保证结构正确，分数未测。diff --git a/solution/run.py b/solution/run.pyindex cc8ced9..f5dca6a 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 (50, family other): PLAN mechanism UNIVAL-TA (type-aware+persistence-weighted κ: w_type from marker-program detection in the opposite+bracket stage, per-side κ_a/κ_b) was implemented and FALSIFIED on the proxy A+half (seed 0, parent = 65.129): ratio ref=0.1 → 65.009, ratio+κ_b=1.5 →+65.101, det ref=0.3 (PLAN literal fallback signal) → 64.748, inverted+(transient-emphasis) ref=0.2 → 65.064, and the decisive control — uniform+κ_b = 0.62 (the MEAN b-side TA weight) → 65.017 with mmd_u/variogram/nbhd+raws identical to graded TA to ≤3e-5 (65.009): the persistence ORDERING+carries no information; only the mean displacement amplitude matters. Per-side+κ grid: κ_b=0 → 64.662, κ_a=0 → 65.116 (b-side regression carries ALL of+UNIVAL's gain; a-side advance is neutral), κ_b=1.5 → 65.084, κ_b=2 → 64.912+(mmd_u degrades ~2x faster in points than nbhd+variogram gain). Backup 2+MATCHDELTA (type-matched stage-effect direction, softmax β=3 over shared-type+profile correlation): κ_b=1 → 65.047 (nbhd .04723 best-in-family but mmd_u+.01126 worse), κ_b=1.5 → 64.959, norm-matched → 65.049 — direction is not the+mmd_u damage source either. Backup 3 WIRESET-AC (group-size-adaptive κ,+nref=40 flo=0.25) → 65.004. SUBMITTED: backup 4 UNIVAL-P — exponent P=3.0 on+the UNIVAL soft gene weight w_g = clip(|Δ̂|/τ,0,1)^P, concentrating the+one-sided displacement on strong stage-effect genes (P=1 spreads it over+480/498 genes). Seed 0: 65.129 (= parent to 3 decimals) with all THREE+cell_state/local_spatial raws slightly better than parent (mmd_u .01084 vs+.01089, variogram .007358 vs .007374, nbhd .04750 vs .04771); seeds 1/2:+64.295/64.445 vs parent 64.294/64.439 — a robust TIE, de_* bit-identical+(pooled-pb neutralisation intact). Interpretation: the UNIVAL gain is+amplitude-driven and gene-selection-neutral; the family's A-half ceiling is+~65.13 at seed 0 and every reallocation between mmd_u and nbhd slides along+one exchange curve. --ablate mechanism (TA/MATCH/ADAPT off, P→1, κ_a=κ_b=1)+reproduces parent node 48 bit-for-bit (X digest 56f694d7d8a1a0b2); disguised+view (+1d shift, renamed inputs, reordered manifest keys) gives the identical+digest (63636e2338b129ae) → view-independent.+ 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@@ -859,12 +890,25 @@ 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")) +# WIRESET-AC (node 50 backup mechanism, parent ANALYSIS suggestion 3):+# group-size-adaptive correction strength. The drawn-group mean of a small+# type×side group is a noisy estimate of its full-stage mean; resetting it+# fully injects resampling noise into the output. Scale the per-group κ by a+# reliability factor clip(n_g / NREF, FLO, 1): large groups (n_g ≥ NREF) keep+# the full correction, small groups are shrunk toward no-op. Bracket data only.+# ADAPT=0 → parent node 48 bit-for-bit.+WIRESET_ADAPT = os.environ.get("T2_WIRESET_ADAPT", "0") == "1"+WIRESET_NREF = float(os.environ.get("T2_WIRESET_NREF", "40"))+WIRESET_FLO = float(os.environ.get("T2_WIRESET_FLO", "0.25"))+  def wireset_reset(expr, out_labels, stage_a, stage_b, ia, ib, n_from_a,-                  kappa, cap, min_cells, eps):+                  kappa, cap, min_cells, eps, adapt=False, nref=40.0, flo=0.25):     """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}+            "wireset_min_cells": int(min_cells), "wireset_applied": False,+            "wireset_adapt": bool(adapt), "wireset_nref": float(nref),+            "wireset_flo": float(flo)}     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@@ -876,6 +920,7 @@ def wireset_reset(expr, out_labels, stage_a, stage_b, ia, ib, n_from_a,     is_a = np.zeros(n, dtype=bool)     is_a[:n_from_a] = True     drift_logs = []+    kg_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:@@ -899,7 +944,9 @@ def wireset_reset(expr, out_labels, stage_a, stage_b, ia, ib, n_from_a,                 continue             n_groups += 1             drift_logs.append(log_rho.tolist())-            f = np.exp(kappa * log_rho)+            kg = kappa * (min(max(ng / max(nref, 1e-9), flo), 1.0) if adapt else 1.0)+            kg_logs.append(float(kg))+            f = np.exp(kg * log_rho)             Xg = E[rows]             nz = Xg != 0.0             E[rows] = np.where(nz, Xg * f[None, :], 0.0)@@ -914,6 +961,8 @@ def wireset_reset(expr, out_labels, stage_a, stage_b, ia, ib, n_from_a,         D = np.abs(np.concatenate([np.asarray(x) for x in drift_logs]))         info.update(             wireset_applied=True, wireset_n_groups=n_groups,+            wireset_kg_min=float(np.min(kg_logs)), wireset_kg_max=float(np.max(kg_logs)),+            wireset_n_adapted=int(sum(1 for x in kg_logs if x < kappa - 1e-9)),             wireset_abslogdrift_median=float(np.median(D)),             wireset_abslogdrift_p90=float(np.quantile(D, 0.90)),             wireset_abslogdrift_max=float(D.max()),@@ -944,13 +993,122 @@ 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"))+# UNIVAL-P (node 50): exponent on the soft gene weight w_g = clip(|Δ̂|/τ,0,1)^P.+# P>1 concentrates the displacement on strong stage-effect genes and spares+# weak ones (480/498 genes are active at P=1). If the mmd_u damage at κ=1+# comes from displacing hundreds of weak/noisy genes while the nbhd gain comes+# from the strong DE genes, P>1 should bend the exchange curve. P=1 → parent.+UNIVAL_P = float(os.environ.get("T2_UNIVAL_P", "3.0"))++# UNIVAL-TA (node 50, PLAN mechanism): type-aware κ. The uniform κ scan of node+# 48 showed a clean exchange — κ↑ improves nbhd/variogram monotonically while+# mmd_u degrades monotonically (net peak κ=1.0). Hypothesis: the exchange comes+# from treating every one-sided type alike. Persistence-weighted κ scales each+# one-sided type's displacement by how strongly its marker program is still (or+# already) detected in the OPPOSITE bracket stage:+#   markers(c) = top-K genes by (det-rate in c − det-rate in rest of origin stage)+#   signal(c)  = "ratio": mean det-rate of markers among opposite-stage cells /+#                          mean det-rate among c's own cells  (program persistence)+#                "det":   mean det-rate of markers among opposite-stage cells+#                "count": PLAN-literal cell count of type c in the opposite stage+#                          (structurally 0 for one-sided types → diagnostic only)+#   w_type     = clip(signal / TA_REF, TA_FLOOR, 1)+#   κ_cell     = κ_side · w_type, with per-side multipliers κ_a / κ_b.+# Bracket data only; no target-stage information; deterministic given the input.+# TA off (T2_UNIVAL_TA=0 or --ablate) → all w_type = 1, κ_a = κ_b = 1 →+# uniform κ = parent node 48 bit-for-bit.+UNIVAL_TA = os.environ.get("T2_UNIVAL_TA", "0") == "1"+UNIVAL_TA_MODE = os.environ.get("T2_UNIVAL_TA_MODE", "ratio")  # ratio | det | count+UNIVAL_TA_REF = float(os.environ.get("T2_UNIVAL_TA_REF", "0.1"))+UNIVAL_TA_FLOOR = float(os.environ.get("T2_UNIVAL_TA_FLOOR", "0.0"))+UNIVAL_TA_K = int(os.environ.get("T2_UNIVAL_TA_K", "25"))+UNIVAL_KA = float(os.environ.get("T2_UNIVAL_KA", "1.0"))+UNIVAL_KB = float(os.environ.get("T2_UNIVAL_KB", "1.0"))++# MATCHDELTA (node 50 backup mechanism, submitted after UNIVAL-TA was falsified):+# per one-sided type, the stage-effect DIRECTION is the softmax(β·sim)-weighted+# blend of per-shared-type deltas Δ_c, where sim = centered Pearson correlation+# of the one-sided type's own-stage pseudobulk profile with each shared type's+# same-stage profile; blended with the pooled Δ̂ by ρ. ρ=0 / MATCH=0 → parent+# node 48 bit-for-bit. Motivation: the uniform-κ scan showed the nbhd gain and+# mmd_u damage scale together with the pooled direction; if the mmd_u damage+# comes from displacing specialized young types along a GENERIC tissue axis,+# type-matched directions should let a larger κ_b buy nbhd without mmd_u loss.+UNIVAL_MATCH = os.environ.get("T2_UNIVAL_MATCH", "0") == "1"+UNIVAL_MATCH_BETA = float(os.environ.get("T2_UNIVAL_MATCH_BETA", "3.0"))+UNIVAL_MATCH_RHO = float(os.environ.get("T2_UNIVAL_MATCH_RHO", "1.0"))+UNIVAL_MATCH_TOPK = int(os.environ.get("T2_UNIVAL_MATCH_TOPK", "0"))+UNIVAL_MATCH_NORM = os.environ.get("T2_UNIVAL_MATCH_NORM", "1") == "1"+++def _det_rate(X, rows=None):+    """Per-gene detection rate (>0 fraction) of a CSR matrix (optionally row subset)."""+    if rows is None:+        nnz = np.asarray((X != 0).sum(axis=0)).ravel()+        return nnz / max(X.shape[0], 1)+    Xs = X[rows]+    nnz = np.asarray((Xs != 0).sum(axis=0)).ravel()+    return nnz / max(Xs.shape[0], 1)+++def unival_ta_weights(stage_a, stage_b, one_sided_types, mode, ref, floor, k):+    """Persistence weight w_type ∈ [floor, 1] for each one-sided type (dict)."""+    la = np.asarray(stage_a.labels).astype(str)+    lb = np.asarray(stage_b.labels).astype(str)+    set_a, set_b = set(la.tolist()), set(lb.tolist())+    det_a = _det_rate(stage_a.X)+    det_b = _det_rate(stage_b.X)+    ws = {}+    for c in one_sided_types:+        in_a = c in set_a+        labs = la if in_a else lb+        X_own = stage_a.X if in_a else stage_b.X+        X_oth = stage_b.X if in_a else stage_a.X+        det_own_all = det_a if in_a else det_b+        det_oth = det_b if in_a else det_a+        m = labs == c+        n_c = int(m.sum())+        if n_c == 0:+            ws[c] = float(floor)+            continue+        if mode == "count":+            ws[c] = float(np.clip(0.0 / max(ref, 1e-9), floor, 1.0))+            continue+        d_in_c = _det_rate(X_own, m)+        d_rest = (det_own_all * len(labs) - d_in_c * n_c) / max(len(labs) - n_c, 1)+        spec = d_in_c - d_rest+        kk = min(int(k), len(spec))+        idx = np.argsort(-spec, kind="stable")[:kk]+        d_other = float(det_oth[idx].mean())+        if mode == "det":+            sig = d_other+        elif mode == "inv":+            # opposite hypothesis probe: TRANSIENT types (signal≈0) get w≈1,+            # persistent types (signal≥ref) get w→floor.+            ratio = d_other / max(float(d_in_c[idx].mean()), 1e-6)+            ws[c] = float(np.clip(1.0 - ratio / max(ref, 1e-9), floor, 1.0))+            continue+        else:  # ratio+            sig = d_other / max(float(d_in_c[idx].mean()), 1e-6)+        ws[c] = float(np.clip(sig / max(ref, 1e-9), floor, 1.0))+    return ws   def unival_align(expr, out_labels, stage_a, stage_b, t, n_from_a,-                 kappa, tau, cap, min_cells):+                 kappa, tau, cap, min_cells,+                 ta_enable=False, ta_mode="ratio", ta_ref=0.1, ta_floor=0.0,+                 ta_k=25, kappa_a=1.0, kappa_b=1.0,+                 match_enable=False, match_beta=3.0, match_rho=1.0, match_topk=0,+                 match_norm=False, wp=1.0):     """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}+            "unival_cap": float(cap), "unival_applied": False,+            "unival_ta": bool(ta_enable), "unival_ta_mode": str(ta_mode),+            "unival_ta_ref": float(ta_ref), "unival_ta_floor": float(ta_floor),+            "unival_kappa_a": float(kappa_a), "unival_kappa_b": float(kappa_b),+            "unival_match": bool(match_enable), "unival_match_beta": float(match_beta),+            "unival_match_rho": float(match_rho), "unival_match_topk": int(match_topk),+            "unival_match_norm": bool(match_norm)}     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@@ -968,6 +1126,8 @@ def unival_align(expr, out_labels, stage_a, stage_b, t, n_from_a,     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)+    if float(wp) != 1.0:+        wg = wg ** float(wp)     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)@@ -975,6 +1135,63 @@ def unival_align(expr, out_labels, stage_a, stage_b, t, n_from_a,     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])+    ta_ws = {}+    w_cell = None+    if ta_enable:+        os_types = sorted({c for c, o in zip(labs_out.tolist(), one_sided.tolist()) if o})+        ta_ws = unival_ta_weights(stage_a, stage_b, os_types, ta_mode, ta_ref,+                                  ta_floor, ta_k)+        w_cell = np.array([ta_ws.get(c, 1.0) for c in labs_out.tolist()],+                          dtype=np.float64)+    # MATCHDELTA (node 50 backup mechanism): type-matched stage-effect+    # direction. The pooled Δ̂ is the count-weighted mean over shared types —+    # one generic direction for every one-sided type. Instead, per one-sided+    # type u, match u's own-stage pseudobulk profile against the same-stage+    # shared-type profiles (centered Pearson over panel genes), softmax-weight+    # the per-shared-type deltas Δ_c by exp(β·sim), and blend with the pooled+    # Δ̂: step_u = clip((1−ρ)Δ̂ + ρ·Σ softmax·Δ_c) · w_g(u). ρ = 0 (or match+    # off) → pooled step → parent bit-for-bit. Bracket data only.+    step_by_type = {}+    if match_enable and one_sided.any():+        def _sim(p, q):+            pc = p - p.mean()+            qc = q - q.mean()+            den = np.linalg.norm(pc) * np.linalg.norm(qc)+            return float(pc @ qc / den) if den > 1e-12 else 0.0+        os_types = sorted({c for c, o in zip(labs_out.tolist(), one_sided.tolist()) if o})+        sims_log = {}+        for u in os_types:+            in_a = u in set_a+            own_stage = stage_a if in_a else stage_b+            own_labs = la if in_a else lb+            own_means = ma_a if in_a else ma_b+            p_u = np.asarray(own_stage.X[own_labs == u].mean(axis=0)).ravel().astype(np.float64)+            sims = np.array([_sim(p_u, own_means[c]) for c in shared])+            order = np.argsort(-sims, kind="stable")+            kk = len(shared) if int(match_topk) <= 0 else min(int(match_topk), len(shared))+            order = order[:kk]+            sm = sims[order]+            ex = np.exp(float(match_beta) * (sm - sm.max()))+            ww = ex / ex.sum()+            matched = (ww[:, None] * D[order]).sum(axis=0)+            blend = (1.0 - float(match_rho)) * dhat + float(match_rho) * matched+            wg_u = np.clip(np.abs(blend) / max(tau, 1e-9), 0.0, 1.0)+            step_u = np.clip(blend, -cap, cap) * wg_u+            if match_norm:+                # direction-only change: restore the pooled step's L2 norm so+                # per-type steps displace cells by the SAME total amplitude as+                # the parent (matched Δ_c blends are less cancellation-averaged+                # than the pooled Δ̂ and hence longer; the extra length, not the+                # direction, is what pushes cells off the truth manifold).+                nu = float(np.linalg.norm(step_u))+                if nu > 1e-12:+                    step_u = step_u * (float(np.linalg.norm(step)) / nu)+            step_by_type[u] = step_u+            sims_log[u] = {"side": "a" if in_a else "b",+                           "top": [(shared[i], round(float(sims[i]), 3), round(float(wi), 3))+                                   for i, wi in zip(order[:3], ww[:3])],+                           "cos_with_pooled": _sim(blend, dhat)}+        info["unival_match_sim"] = sims_log     n_one = 0     for side in (True, False):         rows = one_sided & (is_a == side)@@ -982,14 +1199,44 @@ def unival_align(expr, out_labels, stage_a, stage_b, t, n_from_a,         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+        kside = float(kappa_a) if side else float(kappa_b)+        amp = kappa * kside * t if side else -kappa * kside * (1.0 - t)+        Xg_all = E[rows]+        if step_by_type:+            labs_rows = labs_out[rows]+            w_rows = w_cell[rows] if ta_enable else None+            Yall = np.zeros_like(Xg_all)+            for u in np.unique(labs_rows):+                m_u = labs_rows == u+                s = amp * step_by_type.get(u, step)+                Xg = Xg_all[m_u]+                nz = Xg != 0.0+                if w_rows is not None:+                    S = w_rows[m_u][:, None] * s[None, :]+                    Y = np.where(nz, np.where(S >= 0.0, Xg + np.maximum(S, 0.0),+                                              Xg * np.exp(np.minimum(S, 0.0))), 0.0)+                else:+                    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)+                Yall[m_u] = Y+            np.maximum(Yall, 0.0, out=Yall)+            E[rows] = Yall+        else:+            s = amp * step+            Xg = Xg_all+            nz = Xg != 0.0+            if ta_enable:+                S = w_cell[rows][:, None] * s[None, :]+                pos = np.maximum(S, 0.0)+                dec = np.exp(np.minimum(S, 0.0))+                Y = np.where(nz, np.where(S >= 0.0, Xg + pos, Xg * dec), 0.0)+            else:+                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)@@ -1008,6 +1255,22 @@ def unival_align(expr, out_labels, stage_a, stage_b, t, n_from_a,         unival_nnz_before=int((np.asarray(expr) != 0).sum()),         unival_nnz_after=int((out != 0).sum()),     )+    if ta_enable:+        wv = np.array(list(ta_ws.values()), dtype=np.float64) if ta_ws else np.zeros(0)+        wc = np.array([ta_ws.get(c, 1.0) for c in labs_out[one_sided].tolist()]) \+            if one_sided.any() else np.zeros(0)+        info.update(+            unival_ta_applied=bool(len(ta_ws) > 0),+            unival_ta_n_types=int(len(ta_ws)),+            unival_ta_n_zero_types=int((wv <= 1e-9).sum()) if wv.size else 0,+            unival_ta_n_clipped_types=int((wv >= 1.0 - 1e-9).sum()) if wv.size else 0,+            unival_ta_w_min=float(wv.min()) if wv.size else None,+            unival_ta_w_med=float(np.median(wv)) if wv.size else None,+            unival_ta_w_max=float(wv.max()) if wv.size else None,+            unival_ta_cellw_mean=float(wc.mean()) if wc.size else None,+            unival_ta_cellw_zero_frac=float((wc <= 1e-9).mean()) if wc.size else None,+            unival_ta_weights={k: round(v, 4) for k, v in sorted(ta_ws.items())},+        )     return out, info  @@ -2608,13 +2871,13 @@ 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 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")+                        help="mechanism-off control (node 50): 'mechanism' (or any name) disables "+                             "ONLY UNIVAL-TA (type-aware persistence-weighted κ): all w_type = 1, "+                             "κ_a = κ_b = 1, so UNIVAL keeps its uniform κ = 1.0. Inherited "+                             "mechanisms (UNIVAL, WIRESET, NBHDCOH, DETRX-SIDE, DETR, RECAL, ...) "+                             "stay ON in both runs, so the ablated output reproduces parent node 48 "+                             "bit-for-bit. ANISO2/RDENS/VDEF coordinate/variance probes stay off "+                             "(falsified in nodes 46/48)")     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@@ -2733,28 +2996,45 @@ def main() -> None:                 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+    # at the very end of the expression pipeline (after DETR/DETRX). Default+    # κ = 0 (falsified in node 48).+    vdef_kappa = 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+    # WIRESET (node 48 mechanism, inherited) + WIRESET-AC (node 50 backup+    # mechanism): group-size-adaptive strength. --ablate turns off the node-50+    # mechanisms (TA, MATCH, ADAPT) → parent node 48 bit-for-bit.+    wireset_kappa = WIRESET_KAPPA+    ws_adapt = WIRESET_ADAPT and not args.ablate     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)+                                  WIRESET_EPS, adapt=ws_adapt, nref=WIRESET_NREF,+                                  flo=WIRESET_FLO)     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+    # UNIVAL (node 48 mechanism, inherited) + node 50 mechanisms: UNIVAL-TA+    # (type-aware persistence-weighted κ — FALSIFIED on the proxy A half,+    # default off) and MATCHDELTA (type-matched stage-effect direction).+    # --ablate turns off ONLY the node-50 mechanisms (TA + MATCH) → parent+    # node 48 bit-for-bit.+    unival_kappa = UNIVAL_KAPPA+    ta_enable = UNIVAL_TA and not args.ablate+    ta_ka = 1.0 if args.ablate else UNIVAL_KA+    ta_kb = 1.0 if args.ablate else UNIVAL_KB+    match_enable = UNIVAL_MATCH and not args.ablate     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)+                                 unival_kappa, UNIVAL_TAU, UNIVAL_CAP, UNIVAL_MIN_CELLS,+                                 ta_enable=ta_enable, ta_mode=UNIVAL_TA_MODE,+                                 ta_ref=UNIVAL_TA_REF, ta_floor=UNIVAL_TA_FLOOR,+                                 ta_k=UNIVAL_TA_K, kappa_a=ta_ka, kappa_b=ta_kb,+                                 match_enable=match_enable, match_beta=UNIVAL_MATCH_BETA,+                                 match_rho=UNIVAL_MATCH_RHO, match_topk=UNIVAL_MATCH_TOPK,+                                 match_norm=UNIVAL_MATCH_NORM,+                                 wp=(1.0 if args.ablate else UNIVAL_P))     info.update(uv_info)     sc_info = {"sidecluster": sidecluster}     if sidecluster and expr.shape[0]:@@ -2867,6 +3147,8 @@ def main() -> None:                                          "vdef_g2_dev_max", "vdef_nnz_before", "vdef_nnz_after",                                          "wireset_kappa", "wireset_cap", "wireset_min_cells",                                          "wireset_applied", "wireset_n_groups",+                                         "wireset_adapt", "wireset_nref", "wireset_flo",+                                         "wireset_kg_min", "wireset_kg_max", "wireset_n_adapted",                                          "wireset_abslogdrift_median", "wireset_abslogdrift_p90",                                          "wireset_abslogdrift_max", "wireset_g2_dev_max",                                          "wireset_nnz_before", "wireset_nnz_after",@@ -2874,7 +3156,18 @@ def main() -> None:                                          "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")}+                                         "unival_g2_dev_max", "unival_nnz_before", "unival_nnz_after",+                                         "unival_ta", "unival_ta_mode", "unival_ta_ref",+                                         "unival_ta_floor", "unival_kappa_a", "unival_kappa_b",+                                         "unival_ta_applied", "unival_ta_n_types",+                                         "unival_ta_n_zero_types", "unival_ta_n_clipped_types",+                                         "unival_ta_w_min", "unival_ta_w_med", "unival_ta_w_max",+                                         "unival_ta_cellw_mean", "unival_ta_cellw_zero_frac",+                                         "unival_ta_weights",+                                         "unival_match", "unival_match_beta",+                                         "unival_match_rho", "unival_match_topk",+                                         "unival_match_norm",+                                         "unival_match_sim")}     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 state verified: deterministic, `--ablate` reproduces parent bit-for-bit, disguised-view identical, vec-check ok.

**做了什么**：实现 PLAN 的 UNIVAL-TA（持久度加权 κ，含 det/inv 退路解码）并证否——判决性对照显示分级 κ 与同均值统一 κ_b（0.62）三项 raw 差 ≤3e-5，持久度排序不携带信息，交换只由平均幅度驱动；per-side 网格证明 UNIVAL 收益全部来自 b 侧回退（κ_b=0→64.66，κ_a=0→65.12）。随后依次证否三个备选（MATCHDELTA 类型匹配方向 65.05/64.96、范数匹配 65.05、WIRESET-AC 自适应强度 65.00），最终提交 UNIVAL-P（软基因权重指数 P=3，位移集中到强阶段效应基因）。

**查分结果**（A 半，18/20 次）：提交配置 seed 0 = 65.129（= 父），但 mmd_u .01084/.01089、variogram .007358/.007374、nbhd .04750/.04771 三项 raw 同时略优；seed 1/2 = 64.295/64.445 vs 父 64.294/64.439，稳健打平；DE 通道逐位不变（池化中和完好）；--ablate 逐位还原父 48。本家族在 seed 0 的天花板 ≈65.13——17 个配置全部落在同一条 mmd_u↔nbhd 交换曲线上。

**下一步最值得试**：离开表达值通道——(1) 移植兄弟谱系节点 47 的 SIDEFRIM（occupancy_dice skill .421 仍低于地板，是形状组唯一失分项）；(2) 表达-位置重配对的空间通道（nbhd 权重 25+门，值级位移已榨干）。

调研员

80%的单侧类型n_other=0，说明持久度信号太稀疏，退路改用该类型标记基因在对侧的平均检出率（>0的基因占比）作为软持久度权重。\n\n步骤2（核心机制）：对每个单侧类型计算持久度权重 w_type = clip(n_other / n_thresh, 0, 1)。a起源细胞的位移幅度乘 κ_a × w_type，b起源乘 κ_b × w_type。n_other=0的类型κ=0（不做值级位移，保留原始阶段表达），避免把已消失/未出现类型的细胞推向不存在的中间态。位移方向、Δ̂估计、软权重w_g、截断、乘性衰减、池化伪批量中和全部不变（DE通道结构性不动）。\n\n步骤3（参数扫描，用vec-score A半seed 0筛选）：\n- 第一轮（4次查分）：κ_a=κ_b=1.0，n_thresh∈{3,5,10}，加一组n_thresh=3+per-side κ_a=1.0/κ_b=0.7。目标：确认类型感知κ是否优于统一κ=1.0（65.084）。\n- 第二轮（6次查分）：取最佳n_thresh，κ_a×κ_b∈{0.7,1.0,1.5}×{0.7,1.0,1.5}，共9组取最佳6组。目标：找到mmd_u中性而nbhd保持或改善的组合。\n- 第三轮（2-3次查分）：最佳配置重复查分确认（差距<1分需2次以上同向），再跑seed 1验证。\n\n单输入阶段退路：本机制只用括号内数据（两侧细胞计数），若只有一个输入阶段则无括号、无单侧类型，机制自动退化为恒等（κ×0=0），不影响输出。\n\n时间预算：实现约30行（在UNIVAL循环内加类型权重查找），运行时间与父相同（~2s），20次查分额度足够。",
  "expected_groups": ["cell_state", "local_spatial"],
  "risks": "1）多数单侧类型n_other=0（持久度信号稀疏），导致UNIVAL-TA≈关闭UNIVAL，分数回落到父46水平（64.62）。Engineer应在步骤1先统计n_other分布，若>80%为0则立即切换到标记基因检出率的软权重退路，不要浪费查分额度。2）持久度权重引入额外离散噪声（小样本类型的n_other不稳定），可能使κ在种子间波动。缓解：n_thresh≥3，且用min(两侧细胞数)加权（与Δ̂估计一致）。3）改善幅度可能<1分（T2噪声），需重复查分确认；若两轮最佳配置差<0.5分且方向不一致，提交父配置（NO_CHANGE）。",
  "family_id": "other",
  "mechanism": "UNIVAL的κ从统一常数改为按类型持久度加权：κ_cell = κ_side × clip(n_other/n_thresh, 0, 1)，其中n_other是该单侧类型在对侧括号阶段的细胞数。持久型（n_other高）获得接近κ_side的位移，瞬态型（n_other=0）不做值级位移，从而在保持nbhd收益的同时保护mmd_u不被瞬态型的错误位移拖累。可选per-side非对称κ_a≠κ_b进一步拆解两侧贡献。",
  "vs_constant_shift": "不是全局常数位移：κ按类型×起源侧变化，由括号内实测细胞计数决定（生物学证据驱动），位移方向是逐基因阶段效应Δ̂（非统一方向），幅度受软权重w_g和截断约束，且池化伪批量精确中和保证DE通道不变。常数位移对所有细胞施加相同向量，本机制对不同持久度的类型施加不同幅度甚至零幅度。",
  "mechanism_evidence": "1）κ_type分布应为非均匀（部分类型κ=0，部分接近κ_side），若全部κ相同则机制未生效；2）对比统一κ=1.0，瞬态型（κ=0）细胞的表达应不变、持久型细胞的表达应被位移，可用X digest或逐细胞L2范数差验证；3）mmd_u raw应从.01089下降（瞬态型不再被错误位移），nbhd raw应保持≤.04771或进一步改善；4）de_score/de_direction raw应逐位不变（.3448/.3845），确认池化中和完好；5）四组分各自变化：cell_state和local_spatial应改善，expression_change和shape_scale应不变。",
  "mechanism_off_control": "设T2_UNIVAL_TA=0（或env κ_a=κ_b=1.0, n_thresh=∞），使所有w_type=1，κ_cell=κ_side×1=统一κ，退化为父48的UNIVAL。预期输出与父48逐位一致（X digest相同），榜分差=0。若关闭后输出仍不同，说明实现有bug。",
  "sources": []
}
```

审查员

REVIEW.json written