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

节点 n39 在终选来历上

DETR_EXT：把 DETR 检出率几何插值扩展到单侧独有类型（37% 输出细胞），以共有类型细胞的逐基因检出漂移作缺失侧代理，OFF-only 置零（s=0.35）；PLAN 的 AMPSHRINK 幅度收缩证否（7 配置全劣于父）。

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 2e54d3d851: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 7a44e350fb。改动的文件：solution/METHOD.md +82 −49、solution/README.md +15 −13、solution/run.py +289 −6

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 3b99f0e..82fe83a 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,60 +1,93 @@-DETR：按共有类型把检出率向几何插值目标 p_a^(1−t)·p_b^t 做 OFF 移除（强度1.5），修复地板下 variogram；PLAN 的 SPATRESID 空间门控坍缩证否（8 配置噪声内）。+DETR_EXT：把 DETR 检出率几何插值扩展到单侧独有类型（37% 输出细胞），以共有类型细胞的逐基因检出漂移作缺失侧代理，OFF-only 置零（s=0.35）；PLAN 的 AMPSHRINK 幅度收缩证否（7 配置全劣于父）。 -## 节点 37（improve，父 35，board T2:embryo:val_interp）+- family_id: T2EI-01+- parent: node 37 (DETR, 64.05 official / 63.17 A半 seed0)+- mechanism_off_control: `--ablate mechanism`（或 `T2_DETR_EXT=0`）关闭 DETR_EXT，其余管线不变，输出与父节点 37 逐位一致（sha256 已在 seed 0/1/2 上验证）。 -### 做了什么+## PLAN 机制 AMPSHRINK：已实现并证否（7 次查分、3 种解码） -1. **PLAN 机制 SPATRESID（坐标混居条件化的残余双峰坍缩）——实现、对照通过、证否。**-   - 实现：在最终输出坐标空间（aniso/jitter/RMS 之后）逐共有类型建 15-NN（含自身，与 neighborhood_mmd 同尺度），h_i = 邻域异源占比；只对 h_i ≥ θ 的细胞执行父节点 RESID 坍缩（软权重 w=clip(|s|/0.15,0,1)，nnz-only、clip≥0，保合并均值系数改用门控后计数 c_a=γ·n_bg/(n_ag+n_bg)、c_b=γ·n_ag/(n_ag+n_bg)）。坐标计算被移到表达坍缩之前（不读 xa/xb、不耗 rng，机制关闭时输出与父节点 sha256 逐位一致，已验证）。-   - 关闭对照（PLAN 两条 off 路径均验证）：θ=1.0 输出与父节点默认逐位一致（sha256 79153d…0695 相同）→ 门控无泄漏；--ablate 同样逐位一致。-   - 机制生效证据：混居细胞占比 θ=0.1/0.2/0.3/0.4/0.5 → 20.9%/16.1%/10.3%/8.5%/5.3%（非空转，>PLAN 的 5% 下限）；被门控子集恰是残余分离更大的细胞（gated sep_rel 0.72–0.95 vs 全体 0.63），坍缩后 gated sep_rel 0.58–0.82、全体部分下降；未门控细胞表达改动=0（逐位）；pb_maxshift ≤0.016（DE 通道受保护，de_score raw 0.2857 全程不动）。-   - 查分（A 半 seed 0，基线 62.36）：θ 网格 γ=0.5 → 62.38/62.38/62.37/62.37/62.37；θ=0.2 γ=1.0 → 62.39；θ=0.2 smode=gated γ=0.5/1.0 → 62.38/62.40。**结论：门控把 mmd_u 收益与 nbhd 伤害按被门控比例同比缩小（θ=0.1：mmd −0.00016 / nbhd +0.00023；对比节点 35 无门控：mmd −0.0008 / nbhd +0.0025），交换比改善但总量趋零，净变化 +0.01…+0.04 全部在噪声内。解耦假设不成立。**-2. **备选机制 DETR（检出率几何插值 OFF 校正）——针对同一弱项 cell_state（variogram skill 0.389 地板下）——有效，已提交。**-   - 动机（只用括号数据的事实）：nnz-only 位移把每个细胞的零/非零支持集冻结在源阶段，输出逐基因检出率呈 a/b 双峰。实测：pred 总检出率 0.0571；括号线性插值 0.0711、几何插值 exp((1−t)ln p_a + t ln p_b) ≈ 0.053。表达在 log 空间近乘性增长，检出率的时间插值应取几何而非线性——据此 pred 相对几何目标**过检出**，应做移除而非补入。-   - 实现：逐共有类型（两侧 ≥10 细胞）、逐软权重激活基因（w=clip(|Δ_type|/0.15,0,1)>0.01），目标 p_t = p_a^(1−t)·p_b^t（括号阶段型内检出率），当前过检出的部分按 k=round(strength·n·(p_cur−p_t)) 置零；被选细胞由逐类型抽取、跨基因共享的确定性均匀键排序决定（模拟真实的细胞级深度相关 dropout，而非逐基因独立噪声）。只移除、不补入（OFF_ONLY），坐标与其余管线逐位不动。随机源 np.random.default_rng(seed+777001)，与 mix 的 rng 流隔离，对 seed 确定。-   - 失败的解码（证明"补入"方向结构性有毒）：linear 目标 + ON 补入 q_on=0.25 → **54.57**（variogram 0.0134→0.0310，nbhd 0.0507→0.0844）；q_on=0.05 → **56.69**。注入小值经 variogram/nbhd 的 sqrt 核放大，与任务书"把零值整体抬高有害"一致。linear 目标 OFF-only → 62.57（弱于 geom，因 linear 目标下净方向是补入）。-   - 剂量响应（geom OFF-only，A 半 seed 0）：strength 0.5/1.0/1.5/2.0 → **62.68/62.99/63.17/62.92**。s=1.5：variogram raw 0.01342→0.01046（skill 0.389→0.47 方向修复）、mmd_u 0.01161→0.01139、nbhd 0.05066→0.05083（+0.0002，噪声边缘）、de_direction 0.3852→0.3842（−0.001，噪声内）；s=2.0 时 nbhd 0.05142、de_dir 0.3747 开始受损 → 峰值在 1.5。-   - 种子确认（s=1.5，A 半）：seed 0/1/2 → 63.17/62.67/62.90，均值 62.91，**+0.55 vs 同日基线 62.36**；variogram 在三个种子一致改善（0.0103–0.0109 vs 0.0134，远超 0.0002 噪声阈），mmd_u 一致改善（0.0110–0.0114）。-   - 机制生效证据：s=1.5 移除 18,600 个 nnz 条目（约占全部条目 1%），分布跨 10 个共有类型、约 480 基因/类型；未选中细胞与该类型未激活基因逐位不动；检出率与几何目标的平均差距 0.029→0.008。-   - 关闭对照：DETR_FLIP_ON=0 或 --ablate mechanism → 输出与父节点 35 默认逐位一致（sha256 验证）。+按 PLAN 在 DETR 之后实现：逐共有类型（两侧≥10 细胞）× 软权重激活基因（τ=0.15），+m_t = exp((1−t)·ln m_a + t·ln m_b)（m_a/m_b 为括号阶段实测 nnz 均值），m_cur > m_t 时对该型该基因全部+nnz 条目乘 f = (m_t/m_cur)^strength（只缩不放大）。无随机、保稀疏、保序。 -### 提交配置（run.py 默认）+信号检查（PLAN 步骤 5 / 风险 1）：机制**有信号**——中位相对 gap 9.8%（>5% 门槛），70% 的型×基因对+m_cur > m_t，平均收缩因子 0.91（在预期 0.85–0.98 内），nnz_delta=0（稀疏模式不变）。但查分全部劣于父节点： -父节点 35 全部管线原样（mix + α=5 + soft + corr 扩散 η=1/asym=0.85 + λ=6 + β=0.2 + aniso 1.25），新增 DETR：T2_DETR_FLIP_ON=1.0、MODE=geom、OFF_ONLY=1、STRENGTH=1.5、TAU=0.15、KEY=cell、Q_ON 无效（不补入）。SPATRESID 默认关闭（THETA=0）。+| 配置 | board (A半 seed0) | variogram | mmd_u | nbhd | de_score |+|---|---:|---:|---:|---:|---:|+| 父节点 37 | **63.17** | 0.010463 | 0.01139 | 0.05083 | 0.2857 |+| pooled s=0.25 | 63.04 | 0.010353 | 0.01130 | 0.05270 | 0.2857 |+| pooled s=0.5 | 62.42 | 0.010276 | 0.01185 | 0.05556 | 0.2500 |+| pooled s=1.0 | 61.00 | 0.010207 | 0.01463 | 0.06350 | 0.2500 |+| sided s=0.5 | 61.30 | 0.010151 | 0.01318 | 0.06183 | 0.2143 |+| sided s=1.0 | 57.88 | 0.010085 | 0.02098 | 0.07994 | 0.2143 |+| gapped(>10%) s=0.5 | 62.49 | 0.010281 | 0.01177 | 0.05527 | 0.2500 |+| gapped(>10%) s=1.0 | 61.16 | 0.010212 | 0.01430 | 0.06270 | 0.2500 | -### 验证过 / 没验证+结论（判无效，满足 ≥3 解码/幅度、≥6 次查分）：variogram 收益始终 <0.0004（PLAN 自己的阈值），而+mmd_u / neighborhood_mmd / de_score 随 strength 单调恶化——乘性值缩放即便保序、保稀疏，也把细胞推离+真值拟合的 PCA 流形并破坏表达-位置配对；pb_maxshift（0.14–0.82）远超 PLAN 的 <0.02 预期，"伪批量近似+对消"假设不成立。"nnz 幅度冻结在源阶段"这一动机假设本身为假：α=5 位移后幅度已接近目标（gap 中位仅+9.8%，且收缩方向系统性地偏离真值）。与 node 17/19/37 的教训一致：**值通道修正一律伤 nbhd/mmd，只有+支持集（零/非零）通道可安全修**。AMPSHRINK 代码保留，默认 `T2_AMPSHRINK_ON=0`。 -- 验证：vec-check ok；seed 0 重跑逐位确定；--ablate 逐位还原父节点；θ=1.0 逐位还原（门控无泄漏）；A 半 19 次查分（额度 20 内，剩 1 未用）；单输入回退路径未受影响（代码路径在 bracket 分支内）。-- 没验证：B 半（正式分）；heart 两榜（本 run 只挂 embryo 视图）；strength 1.25/1.75 细扫；KEY=gene（逐基因独立键）变体；伪装视图重跑（程序只读视图数据与 seed，不读绝对时间/路径，t 为平移不变量，预期通过）。-- 耗时 ~30 s、峰值内存 <1 GB，limits（30 min / 28 GB）内；EXECUTION.json {"gpu": false}。+## 提交的备选机制 DETR_EXT（针对同一弱项 cell_state/variogram） -### 知识来源+诊断：父节点 DETR 只覆盖共有类型（两侧≥10 细胞，10 个），但 mix 输出按全类型分层抽样，**37% 的输出+细胞（a 侧 857 + b 侧 985，共 24 个单侧类型）的检出支持集完全未校正**，仍冻结在源阶段水平。b 侧细胞+相对合并几何目标在 95% 基因上过检出（0.134 vs 0.045）。 -- 未使用任何保留阶段（E7.25/E7.5 窗口）或保留基因型数据；DETR 的目标检出率完全由视图内括号阶段（E6.75、E8.0）现场计算。-- 方法学依据（非阶段特异性测量）：log 空间表达近乘性 → 检出率几何插值；真实 scRNA/空间数据的 dropout 是细胞级深度相关的（逐细胞键跨基因共享），不是逐基因独立噪声。无外部文献数值。-- 评分规则事实来自任务书"T2 评分规则简报"（variogram/nbhd 的 sqrt 核对小值敏感、地板定义、噪声 ~1 分）。+机制：对每个单侧类型的输出细胞，用**逐基因全局检出漂移**作缺失侧代理：+p_t(g) = p_own(g)·(pa_sh(g)/pb_sh(g))^((1−t) 对 b 侧, −t 对 a 侧)，其中 pa_sh/pb_sh 只在**共有类型+细胞**上计算（`DETR_EXT_MODE=shared`）——去掉类型组成变化对漂移估计的混杂；两侧任一未检出的基因被+门掉（保护单侧类型的身份基因不被清零）。过检出部分按与 DETR 相同的规则 OFF-only 置零：+k = round((p_cur−p_t)·n·strength)，逐细胞确定性键跨基因共享（模拟深度相关 dropout）。无新随机源+（独立 rng = seed+777002，确定性）；不注入任何值。 -### 查分明细（A 半，seed 0，除注明外）+生物学依据：发育中检出率随时间的乘性（log 线性）漂移与 node 37 已验证的几何插值目标同源；缺失侧用+群体水平漂移近似是"谱系去向/发育事件顺序"级别的通用程序知识，不含任何保留阶段的测量值。所有量均从+view 输入现场计算，无硬编码。 -| 配置 | board | mmd_u | variogram | nbhd | de_dir |-|---|---:|---:|---:|---:|---:|-| 基线（父 35 默认） | 62.36 | 0.01161 | 0.01342 | 0.05066 | 0.3852 |-| SPATRESID θ=0.1 γ0.5 | 62.38 | 0.01145 | 0.01341 | 0.05089 | 0.3861 |-| SPATRESID θ=0.2 γ0.5 | 62.38 | 0.01149 | 0.01341 | 0.05083 | 0.3859 |-| SPATRESID θ=0.3 γ0.5 | 62.37 | 0.01154 | 0.01342 | 0.05072 | 0.3853 |-| SPATRESID θ=0.4 γ0.5 | 62.37 | 0.01155 | 0.01342 | 0.05070 | 0.3852 |-| SPATRESID θ=0.5 γ0.5 | 62.37 | 0.01158 | 0.01342 | 0.05065 | 0.3853 |-| SPATRESID θ=0.2 γ1.0 | 62.39 | 0.01141 | 0.01339 | 0.05095 | 0.3865 |-| SPATRESID θ=0.2 gated γ0.5 | 62.38 | 0.01149 | 0.01340 | 0.05083 | 0.3862 |-| SPATRESID θ=0.2 gated γ1.0 | 62.40 | 0.01142 | 0.01338 | 0.05094 | 0.3868 |-| DETR linear q0.25 | 54.57 | 0.01744 | 0.03104 | 0.08444 | 0.3613 |-| DETR linear q0.05 | 56.69 | 0.01609 | 0.02827 | 0.07426 | 0.3642 |-| DETR linear off-only | 62.57 | 0.01154 | 0.01243 | 0.05073 | 0.3783 |-| DETR geom off-only s1.0 | 62.99 | 0.01145 | 0.01109 | 0.05062 | 0.3833 |-| DETR geom off-only s0.5 | 62.68 | 0.01149 | 0.01224 | 0.05052 | 0.3839 |-| **DETR geom off-only s1.5（提交）** | **63.17** | 0.01139 | 0.01046 | 0.05083 | 0.3842 |-| DETR geom off-only s2.0 | 62.92 | 0.01141 | 0.01030 | 0.05142 | 0.3747 |-| 提交配置 seed 1 | 62.67 | 0.01101 | 0.01087 | 0.05155 | 0.3738 |-| 提交配置 seed 2 | 62.90 | 0.01112 | 0.01030 | 0.05027 | 0.3780 |--（基线被查分两次——首次输出未落盘，重查一次作锚点；两次同预测。）+机制证据（seed 0）：覆盖 24 个类型、1842/5000 细胞（37%）、445 基因；移除 13,678 个 nnz 条目+（这些细胞 nnz 的 ~17%）；逐型检出率相对目标的平均绝对 gap 0.0763 → 0.0552；pb_maxshift=0.299+（与已验证安全的 DETR 0.211 同量级）；共享类型细胞的输出不受影响（rng 独立，逐位不变）。++## 查分记录（A 半 proxy，共 19 次，余 1）++DETR_EXT 剂量响应（seed 0）：++| 配置 | board | variogram | mmd_u | nbhd | de_score | de_dir |+|---|---:|---:|---:|---:|---:|---:|+| mode=all s=0.3 | 64.11 | 0.007848 | 0.01106 | 0.05248 | 0.3214 | 0.3724 |+| mode=all s=0.4 | 64.13 | 0.007258 | 0.01121 | 0.05404 | 0.3214 | 0.3701 |+| mode=all s=0.5 | 64.16 | 0.006922 | 0.01154 | 0.05582 | 0.3571 | 0.3651 |+| mode=all s=0.75 | 63.45 | 0.006364 | 0.01255 | 0.06220 | 0.3571 | 0.3548 |+| mode=all s=1.5 | 54.55 | 0.009063 | 0.01973 | 0.09083 | 0.1429 | 0.0912 |+| **mode=shared s=0.35（提交）** | **64.34** | 0.007513 | 0.01094 | 0.05232 | 0.3214 | 0.3786 |+| mode=shared s=0.5 | 64.27 | 0.006872 | 0.01120 | 0.05444 | 0.3214 | 0.3707 |+| mode=shared s=0.75 | 63.73 | 0.006245 | 0.01209 | 0.05985 | 0.3214 | 0.3657 |++种子确认（mode=shared s=0.35）：seed 0/1/2 = 64.34 / 63.89 / 63.97，均值 **64.07**（父节点 seed 0+= 63.17，官方 rank3 = 63.71）。四组变化（seed 0 vs 父）：cell_state 53.17→57.94（variogram raw+0.01046→0.00751，mmd_u 0.01139→0.01094），expression_change 61.64→62.24（de_score 0.286→0.321，+de_dir 0.384→0.379），local_spatial 60.12→59.41（nbhd 0.0508→0.0523，代价 −0.7），shape_scale 不变+（坐标逐位未动）。净 +1.17（seed 0）/ +0.9（三种子均值口径）。++选择 shared s=0.35 而非 all s=0.5 的理由：board 更高且 nbhd 伤害减半（0.0523 vs 0.0558）、+de_dir 保留更多；shared 漂移估计去掉组成混杂，更可能迁移到真实括号（两端相邻、共有类型多）。++## 验证与未验证++- 已验证：`vec-check` 通过（seed 0/1/2）；`--ablate mechanism` 输出与父节点 37 逐位一致（sha256，+  3 个种子）；重跑同 seed 输出逐位一致（确定性）；nnz 条目只减不增（OFF-only）；单输入回退路径+  （b is None 提前返回，DETR_EXT 需 bracket）不受影响；程序只依赖时间差 t，不依赖绝对时间/视图字段。+- 未验证：真实括号（final 视图）上的效果——真实两端相邻、单侧类型构成不同，漂移代理的偏差未知；+  B 半分数；s 在 0.35–0.5 之间的更细网格（差异在噪声内，未花额度）。+- 已知风险：nbhd 小幅受损（−0.7 分）是该机制的系统性代价；若正式评分的 nbhd 地板/天花板更紧，+  损失可能放大。s≥0.75 时 nbhd 伤害陡增，不要上调。++## 知识来源++无外部文献/数据库条目。全部量（检出率、漂移比、目标、细胞数、t）从 view 的 manifest 与输入 h5ad+现场计算；所用生物学假设仅为"表达检出随发育时间近似乘性漂移"（与父节点已验证的几何插值目标同源，+属通用机制知识，不含保留阶段测量值）。diff --git a/solution/README.md b/solution/README.mdindex 81cabb3..c0786dc 100644--- a/solution/README.md+++ b/solution/README.md@@ -1,16 +1,18 @@-# mix + 表达收敛管线 + DETR 检出率几何插值 OFF 校正（T2:embryo:val_interp）+# mix + 表达收敛管线 + DETR / DETR_EXT 检出率几何插值（T2:embryo:val_interp） -父节点 35 管线原样保留（mix 混抽 + α=5 类型级收敛位移 + 软阈值基因权重 + 相关扩散-η=1/asym=0.85 + λ=6 投影加权 + β=0.2 型内配对收缩 + aniso 坐标整形 damp=1.25；-RESID/EIGPROJ/ITERDIFF/VARISO/TYPE_ANISO/PBPROJ/TSHIFT/SPATRESID 默认关闭）。+父节点 37 管线原样保留（mix 混抽 + α=5 类型级收敛位移 + 软阈值基因权重 + 相关扩散+η=1/asym=0.85 + λ=6 投影加权 + β=0.2 型内配对收缩 + aniso 坐标整形 damp=1.25 + DETR+共有类型检出率 OFF 校正 s=1.5；RESID/EIGPROJ/ITERDIFF/VARISO/TYPE_ANISO/PBPROJ/+TSHIFT/SPATRESID/AMPSHRINK 默认关闭）。 -本节点（37）提交机制 DETR：nnz-only 位移把细胞的零/非零支持集冻结在源阶段，逐基因-检出率相对几何插值目标 p_t = p_a^(1−t)·p_b^t 系统性偏高（log 空间乘性时间插值）。-DETR 按共有类型/软权重基因，把过检出的 nnz 条目按确定性逐细胞键排序置零-（T2_DETR_STRENGTH=1.5，只移除、不补入；ON 补入在 proxy 上结构性有害：54.6/56.7）。-A 半：62.36 → 63.17（seed 0），seeds 0/1/2 均值 62.91；variogram raw 0.01342→0.01046-（地板下弱项），nbhd/mmd_u 不劣。--ablate（或 T2_DETR_FLIP_ON=0）逐位还原父节点 35。+本节点（39）提交机制 DETR_EXT：DETR 只覆盖共有类型，37% 输出细胞（单侧独有类型）的+检出支持集仍冻结在源阶段。DETR_EXT 用共有类型细胞测得的逐基因检出漂移+(pa_sh/pb_sh) 作缺失侧代理，对这些细胞做同一几何插值目标的 OFF-only 置零+（T2_DETR_EXT_MODE=shared，T2_DETR_EXT_STRENGTH=0.35，逐细胞确定性键）。+A 半 seed 0：63.17 → 64.34；seeds 0/1/2 均值 64.07；variogram 0.01046→0.00751，+de_score 0.286→0.321，nbhd 0.0508→0.0523（代价 −0.7 分，已计入净收益）。+--ablate（或 T2_DETR_EXT=0）逐位还原父节点 37（sha256 已验证）。 -PLAN 机制 SPATRESID（坐标混居门控的残余双峰坍缩）已实现并证否：8 配置（θ 0.1–0.5 ×-γ 0.5/1.0 × s-mode）全在基线 +0.04 内，门控把 mmd_u 收益与 nbhd 伤害同比缩小，-解耦失败。T2_SPATRESID_THETA=0 默认关闭。详见 METHOD.md。+PLAN 机制 AMPSHRINK（nnz 幅度几何收缩）已实现并证否：3 解码 × 7 配置全部劣于父节点+（57.88–63.04），值通道乘性缩放单调伤害 mmd_u/nbhd/de。默认 T2_AMPSHRINK_ON=0。+详见 METHOD.md。diff --git a/solution/run.py b/solution/run.pyindex e279333..370b6bf 100644--- a/solution/run.py+++ b/solution/run.py@@ -8,7 +8,31 @@ 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 (37, DETR, family T2EI-01): the PLAN mechanism (SPATRESID: gate the+This node (39, DETR_EXT, family T2EI-01): the PLAN mechanism (AMPSHRINK:+shrink-only geometric interpolation of per-type per-gene nnz amplitude means+toward m_t = m_a^(1−t)·m_b^t) was implemented and FALSIFIED: 7 configs+(3 decodes pooled/sided/gapped × strengths 0.25..1.5) all land BELOW the+parent (A half seed 0: 63.17 → 57.88..63.04) with monotone dose-response+damage to mmd_u, neighborhood_mmd and de_score — multiplicative value scaling+moves cells off the truth-fitted PCA manifold even though it preserves rank+order and sparsity; the frozen-amplitude hypothesis is false (the α=5+displacement already lands the nnz means near target). AMPSHRINK stays in the+code behind T2_AMPSHRINK_ON=0 (default off). The submitted alternative+mechanism is DETR_EXT: parent DETR only covers shared types (>=10 cells at+both bracket stages), leaving ~37% of output cells (one-side-only types, e.g.+Forebrain/EXE-Ectoderm here) with detection support frozen at the source+stage. DETR_EXT extends the geometric detection interpolation to those types+using the per-gene detection drift measured on SHARED-type cells only+(composition-confounding removed; genes undetected in shared cells at either+stage are gated out, protecting one-side-only identity genes from erasure):+p_t = p_own·(pa_sh/pb_sh)^(1−t) for b-only cells (OFF-only flips, per-cell+deterministic key, strength 0.35). A half seed 0: 64.34 (parent 63.17),+seeds 0/1/2 mean 64.07; variogram 0.01046→0.00751 (biggest below-floor+weak item), de_score 0.286→0.321, nbhd damage contained (0.0508→0.0523).+--ablate (or T2_DETR_EXT=0) reproduces parent node 37 bit-for-bit+(sha256-verified on seeds 0/1/2). See METHOD.md.++Parent node (37, DETR, family T2EI-01): the PLAN mechanism (SPATRESID: gate the node-35 RESID residual collapse on coordinate-mixed cells, 15-NN hetero-source fraction h_i >= θ) was implemented and falsified — 8 configs (θ 0.1..0.5 × γ 0.5/1.0 × s-mode all/gated) all land within +0.04 of the parent (A half seed 0:@@ -283,6 +307,62 @@ DETR_OFF_ONLY = os.environ.get("T2_DETR_OFF_ONLY", "1") == "1" # Flip magnitude: k = round(STRENGTH · n · (p_cur − p_t)_+). STRENGTH=1 closes # the full gap to the target rate; <1 partial (dose-response); >1 overshoots. DETR_STRENGTH = float(os.environ.get("T2_DETR_STRENGTH", "1.5"))+# DETR_EXT (backup mechanism of this node 39, after AMPSHRINK falsification):+# DETR only covers shared types (>=10 cells at BOTH bracket stages), which+# leaves ~36% of output cells (one-side-only types) with their detection+# support completely frozen at the source stage. DETR_EXT extends the same+# geometric-interpolation logic to those types using the GLOBAL per-gene+# detection drift between the bracket stages as the missing side:+#   b-only type: p_t(g) = p_own(g) · (pa_all(g)/pb_all(g))^(1−t)+#   a-only type: p_t(g) = p_own(g) · (pb_all(g)/pa_all(g))^t   (growth; a+#                no-op under OFF-only removal except where drift is negative)+# Rationale: a type absent from one bracket stage has no measurable own-type+# future/past detection, but the stage-level per-gene detection drift+# (sequencing depth + global biology) is observable and is the dominant+# component of the a→b rate change; scaling the type's own rate by the global+# drift raised to the time fraction is the geometric-interpolation target+# under that approximation. Genes not detected at BOTH stages are skipped+# (no drift estimate — protects stage-b-specific identity genes from+# erasure). OFF-only always (ON injection is structurally toxic, node-37+# lesson); same per-cell deterministic key shared across genes.+# T2_DETR_EXT=0 (or --ablate) disables and reproduces parent node 37+# bit-for-bit.+DETR_EXT = os.environ.get("T2_DETR_EXT", "1") == "1"+DETR_EXT_TAU = float(os.environ.get("T2_DETR_EXT_TAU", "0.15"))+# DETR_EXT_MODE: "all" — drift ratio from ALL cells of each bracket stage+# (confounded by type-composition change: genes marking b-only types get a+# huge negative logratio and are nearly erased); "shared" — drift estimated+# only on cells of shared types (present at both stages), so it measures+# technical/generic developmental drift, and genes undetected in shared+# cells at either stage are gated out (protects one-side-only identity+# genes from erasure).+DETR_EXT_MODE = os.environ.get("T2_DETR_EXT_MODE", "shared")+DETR_EXT_STRENGTH = float(os.environ.get("T2_DETR_EXT_STRENGTH", "0.35"))+# AMPSHRINK (mechanism of this node 39, family T2EI-01): DETR repaired the+# zero/nonzero SUPPORT channel; the nnz AMPLITUDE channel is still frozen at+# the source stage (a-side cells keep the E_a nnz-value distribution, b-side+# the E_b one) because nnz-only displacement adds a uniform Δ without changing+# the amplitude distribution shape. AMPSHRINK generalizes the geometric+# interpolation target to nnz amplitudes: per shared type × activated gene+# (same soft-weight decode as DETR, τ=0.15), compute measured nnz means m_a,+# m_b at the bracket stages, the target m_t = exp((1−t)·ln m_a + t·ln m_b),+# and the current output nnz mean m_cur; if m_cur > m_t, multiply every nnz+# entry of that type×gene by f = (m_t/m_cur)^STRENGTH (shrink-only; the ON /+# amplify direction is structurally toxic per node-37 lessons). Deterministic+# proportional scaling: sparsity pattern and within-gene cell order preserved.+# T2_AMPSHRINK_ON=0 (or --ablate mechanism) disables and reproduces parent+# node 37 bit-for-bit.+AMPS_ENABLE = os.environ.get("T2_AMPSHRINK_ON", "0") == "1"+AMPS_STRENGTH = float(os.environ.get("T2_AMPSHRINK_STRENGTH", "0.5"))+AMPS_TAU = float(os.environ.get("T2_AMPSHRINK_TAU", "0.15"))+AMPS_MIN_CELLS = int(os.environ.get("T2_AMPSHRINK_MIN_CELLS", "10"))+# AMPS_DECODE: "pooled" — one f per type×gene from the merged a+b output nnz+# mean (PLAN spec); "sided" — separate f per side (a-side and b-side nnz means+# can straddle m_t in opposite directions, pooled hides that); "gapped" —+# pooled but only for pairs whose relative gap exceeds AMPS_GAP_MIN (spares+# near-correct pairs from any perturbation).+AMPS_DECODE = os.environ.get("T2_AMPSHRINK_DECODE", "pooled")+AMPS_GAP_MIN = float(os.environ.get("T2_AMPSHRINK_GAP_MIN", "0.10")) # per-cell λ-style weighting inside the residual collapse (0 = constant shift) RESID_LAMBDA = float(os.environ.get("T2_RESID_LAMBDA", "0")) # which side moves in the residual collapse: both | a | b (node-34 lesson:@@ -1474,6 +1554,194 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s             pb_after_d = np.vstack([xa, xb]).mean(axis=0).astype(np.float64)             det_info["detr_dp_max"] = float(np.abs(pb_after_d - pb_before_d).max()) +    # DETR_EXT (backup mechanism of this node 39): extend the detection-rate+    # geometric interpolation to one-side-only output types using the global+    # per-gene detection drift as the missing side. See the config block for+    # the full rationale. OFF-only; per-cell deterministic key; no value+    # injection; genes undetected at either bracket stage are skipped.+    ext_info = {"detr_ext_enable": bool(DETR_EXT), "detr_ext_tau": DETR_EXT_TAU, "detr_ext_mode": DETR_EXT_MODE,+                "detr_ext_strength": DETR_EXT_STRENGTH, "detr_ext_n_types_a": 0,+                "detr_ext_n_types_b": 0, "detr_ext_cells_a": 0, "detr_ext_cells_b": 0,+                "detr_ext_n_genes": 0, "detr_ext_n_off": 0,+                "detr_ext_rate_gap_before": None, "detr_ext_rate_gap_after": None,+                "detr_ext_pb_maxshift": None}+    if DETR_FLIP_ON > 0.0 and DETR_EXT and ia.size and ib.size:+        la_e = np.asarray(stage_a.labels).astype(str)+        lb_e = np.asarray(stage_b.labels).astype(str)+        la_o = la_e[ia]+        lb_o = lb_e[ib]+        cnt_a_e = {k: int((la_e == k).sum()) for k in set(la_e.tolist())}+        cnt_b_e = {k: int((lb_e == k).sum()) for k in set(lb_e.tolist())}+        shared10 = {k for k in (set(cnt_a_e) & set(cnt_b_e))+                    if cnt_a_e[k] >= 10 and cnt_b_e[k] >= 10}+        if DETR_EXT_MODE == "shared" and shared10:+            ma_sh = np.isin(la_e, list(shared10))+            mb_sh = np.isin(lb_e, list(shared10))+        else:+            ma_sh = np.ones(la_e.shape[0], dtype=bool)+            mb_sh = np.ones(lb_e.shape[0], dtype=bool)+        pa_all = (np.asarray(stage_a.X[ma_sh].todense(), dtype=np.float64) > 0).mean(axis=0)+        pb_all = (np.asarray(stage_b.X[mb_sh].todense(), dtype=np.float64) > 0).mean(axis=0)+        eps_e = 1e-4+        logratio = np.log((pa_all + eps_e) / (pb_all + eps_e))+        gate = (pa_all > 0) & (pb_all > 0)+        gpb_a = np.asarray(stage_a.X[ma_sh].mean(axis=0)).ravel().astype(np.float64)+        gpb_b = np.asarray(stage_b.X[mb_sh].mean(axis=0)).ravel().astype(np.float64)+        w_e = np.clip(np.abs(gpb_b - gpb_a) / max(DETR_EXT_TAU, 1e-12), 0.0, 1.0)+        act_e = np.where((w_e > 0.01) & gate)[0]+        ext_info["detr_ext_n_genes"] = int(act_e.size)+        if act_e.size:+            rng_e = np.random.default_rng(int(params.get("seed", 0)) + 777002)+            pb_before_e = np.vstack([xa, xb]).mean(axis=0).astype(np.float64)+            gaps_b_e, gaps_a_e = [], []+            for side in ("a", "b"):+                lo = la_o if side == "a" else lb_o+                xo = xa if side == "a" else xb+                expo = -t if side == "a" else (1.0 - t)  # p_t = p_own·exp(expo·logratio)+                for lab in sorted(set(lo.tolist()) - shared10):+                    m = lo == lab+                    n_c = int(m.sum())+                    if n_c == 0:+                        continue+                    Xc = xo[m]+                    p_cur = (Xc > 0).mean(axis=0)+                    p_t = p_cur * np.exp(expo * logratio)+                    u = rng_e.random(n_c)+                    gap_b_side = float(np.abs(p_cur[act_e] - p_t[act_e]).mean())+                    n_off_t = 0+                    Xn = Xc.copy()+                    for g in act_e:+                        dpr = (p_t[g] - p_cur[g]) * n_c * DETR_EXT_STRENGTH+                        if dpr <= -1.0:+                            nz = np.where(Xn[:, g] > 0)[0]+                            k = min(int(round(-dpr)), nz.size)+                            if k > 0:+                                sel = nz[np.argsort(u[nz], kind="stable")[:k]]+                                Xn[sel, g] = 0.0+                                n_off_t += k+                    if n_off_t:+                        xo[m] = Xn+                    p_post = (Xn > 0).mean(axis=0)+                    gaps_b_e.append(gap_b_side)+                    gaps_a_e.append(float(np.abs(p_post[act_e] - p_t[act_e]).mean()))+                    ext_info["detr_ext_n_off"] += n_off_t+                    if side == "a":+                        ext_info["detr_ext_n_types_a"] += 1+                        ext_info["detr_ext_cells_a"] += n_c+                    else:+                        ext_info["detr_ext_n_types_b"] += 1+                        ext_info["detr_ext_cells_b"] += n_c+            if gaps_b_e:+                ext_info["detr_ext_rate_gap_before"] = float(np.mean(gaps_b_e))+                ext_info["detr_ext_rate_gap_after"] = float(np.mean(gaps_a_e))+            pb_after_e = np.vstack([xa, xb]).mean(axis=0).astype(np.float64)+            ext_info["detr_ext_pb_maxshift"] = float(np.abs(pb_after_e - pb_before_e).max())++    # AMPSHRINK (this node 39): shrink-only geometric interpolation of nnz+    # amplitude means, applied AFTER DETR (order matters: DETR removed entries+    # change m_cur). See the config block for the mechanism rationale. No rng:+    # fully deterministic proportional scaling; zeros untouched (sparsity+    # pattern and within-gene cell ranking preserved, so the expression-position+    # pairing that neighborhood_mmd reads is undisturbed).+    amp_info = {"ampshrink_enable": bool(AMPS_ENABLE), "ampshrink_strength": AMPS_STRENGTH,+                "ampshrink_tau": AMPS_TAU, "ampshrink_decode": AMPS_DECODE,+                "ampshrink_n_types": 0, "ampshrink_n_pairs": 0,+                "ampshrink_n_shrunk": 0, "ampshrink_frac_shrunk": None,+                "ampshrink_f_mean": None, "ampshrink_f_min": None, "ampshrink_f_max": None,+                "ampshrink_gap_med_before": None, "ampshrink_gap_med_after": None,+                "ampshrink_gap_frac_over5_before": None, "ampshrink_gap_frac_over5_after": None,+                "ampshrink_pb_maxshift": None, "ampshrink_nnz_delta": None}+    if AMPS_ENABLE and ia.size and ib.size:+        la_s = np.asarray(stage_a.labels).astype(str)+        lb_s = np.asarray(stage_b.labels).astype(str)+        la_o = la_s[ia]+        lb_o = lb_s[ib]+        shared_s = sorted(set(la_o.tolist()) & set(lb_o.tolist()))+        means_a_s = type_means(stage_a.X, stage_a.labels)+        means_b_s = type_means(stage_b.X, stage_b.labels)+        cnt_a_s = {k: int((la_s == k).sum()) for k in means_a_s}+        cnt_b_s = {k: int((lb_s == k).sum()) for k in means_b_s}+        shared_s = [k for k in shared_s+                    if cnt_a_s.get(k, 0) >= AMPS_MIN_CELLS and cnt_b_s.get(k, 0) >= AMPS_MIN_CELLS]+        pb_before_s = np.vstack([xa, xb]).mean(axis=0).astype(np.float64)+        nnz_before = int((xa > 0).sum() + (xb > 0).sum())+        gaps_b, gaps_a, fs, n_pairs, n_shrunk = [], [], [], 0, 0++        def _nnz_mean(Xs):+            s = np.asarray(Xs.sum(axis=0)).ravel().astype(np.float64)+            c = np.asarray((Xs > 0).sum(axis=0)).ravel().astype(np.float64)+            return np.where(c > 0, s / np.maximum(c, 1.0), 0.0), c++        for lab in shared_s:+            ma = la_o == lab+            mb = lb_o == lab+            A = xa[ma]+            B = xb[mb]+            d0 = (means_b_s[lab] - means_a_s[lab]).astype(np.float64)+            w = np.clip(np.abs(d0) / max(AMPS_TAU, 1e-12), 0.0, 1.0)+            act = w > 0.01+            if int(act.sum()) < 5:+                continue+            m_a, _ = _nnz_mean(stage_a.X[la_s == lab])+            m_b, _ = _nnz_mean(stage_b.X[lb_s == lab])+            sA = np.where(A > 0, A, 0.0).sum(axis=0).astype(np.float64)+            cA = (A > 0).sum(axis=0).astype(np.float64)+            sB = np.where(B > 0, B, 0.0).sum(axis=0).astype(np.float64)+            cB = (B > 0).sum(axis=0).astype(np.float64)+            m_cur = np.where(cA + cB > 0, (sA + sB) / np.maximum(cA + cB, 1.0), 0.0)+            ok = act & (m_a > 0) & (m_b > 0) & (cA + cB > 0) & (m_cur > 0)+            if not ok.any():+                continue+            m_t = np.exp((1.0 - t) * np.log(np.maximum(m_a, 1e-9))+                         + t * np.log(np.maximum(m_b, 1e-9)))+            f = np.ones(len(m_t))+            fB = f.copy()+            rel_gap_b = (m_cur - m_t) / np.maximum(m_t, 1e-12)+            if AMPS_DECODE == "sided":+                mA_cur = np.where(cA > 0, sA / np.maximum(cA, 1.0), 0.0)+                mB_cur = np.where(cB > 0, sB / np.maximum(cB, 1.0), 0.0)+                okA = ok & (cA > 0) & (mA_cur > m_t)+                okB = ok & (cB > 0) & (mB_cur > m_t)+                f[okA] = (m_t[okA] / mA_cur[okA]) ** AMPS_STRENGTH+                fB[okB] = (m_t[okB] / mB_cur[okB]) ** AMPS_STRENGTH+                over = okA | okB+            else:+                over = ok & (m_cur > m_t)+                if AMPS_DECODE == "gapped":+                    over &= rel_gap_b > AMPS_GAP_MIN+                f[over] = (m_t[over] / m_cur[over]) ** AMPS_STRENGTH+                fB = f.copy()+            rel_gap_a = (m_cur * np.where(cA + cB > 0, (cA * f + cB * fB) / np.maximum(cA + cB, 1.0), 1.0)+                         - m_t) / np.maximum(m_t, 1e-12)+            if over.any():+                An = np.where(A > 0, A * f.astype(np.float32)[None, :], A).astype(np.float32)+                Bn = np.where(B > 0, B * fB.astype(np.float32)[None, :], B).astype(np.float32)+                xa[ma] = An+                xb[mb] = Bn+            n_pairs += int(ok.sum())+            n_shrunk += int(over.sum())+            gaps_b.extend(rel_gap_b[ok].tolist())+            gaps_a.extend(rel_gap_a[ok].tolist())+            fs.extend(f[over].tolist())+            amp_info["ampshrink_n_types"] += 1+        if n_pairs:+            amp_info["ampshrink_n_pairs"] = n_pairs+            amp_info["ampshrink_n_shrunk"] = n_shrunk+            amp_info["ampshrink_frac_shrunk"] = float(n_shrunk / n_pairs)+            gb = np.asarray(gaps_b)+            ga = np.asarray(gaps_a)+            amp_info["ampshrink_gap_med_before"] = float(np.median(gb))+            amp_info["ampshrink_gap_med_after"] = float(np.median(ga))+            amp_info["ampshrink_gap_frac_over5_before"] = float(np.mean(np.abs(gb) > 0.05))+            amp_info["ampshrink_gap_frac_over5_after"] = float(np.mean(np.abs(ga) > 0.05))+            if fs:+                amp_info["ampshrink_f_mean"] = float(np.mean(fs))+                amp_info["ampshrink_f_min"] = float(np.min(fs))+                amp_info["ampshrink_f_max"] = float(np.max(fs))+            pb_after_s = np.vstack([xa, xb]).mean(axis=0).astype(np.float64)+            amp_info["ampshrink_pb_maxshift"] = float(np.abs(pb_after_s - pb_before_s).max())+            amp_info["ampshrink_nnz_delta"] = int((xa > 0).sum() + (xb > 0).sum()) - nnz_before+     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]:@@ -1567,6 +1835,8 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s     info.update(resid_info)     info.update(spat_info)     info.update(det_info)+    info.update(ext_info)+    info.update(amp_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,@@ -1581,14 +1851,15 @@ def main() -> None:     parser.add_argument("--seed", type=int, default=0)     parser.add_argument("--ablate", default=None,                         help="mechanism-off control: 'mechanism' (or any name) disables this "-                             "node's mechanisms (RESID_GAMMA=0, EIGPROJ_K=0), reproducing "-                             "parent node 34 bit-for-bit")+                             "node's mechanisms (DETR_EXT coverage extension; AMPSHRINK stays "+                             "off), reproducing parent node 37 bit-for-bit")     args = parser.parse_args()-    global RESID_GAMMA, SPATRESID_THETA, DETR_FLIP_ON+    global RESID_GAMMA, SPATRESID_THETA, DETR_FLIP_ON, AMPS_ENABLE, DETR_EXT     if args.ablate:         RESID_GAMMA = 0.0         SPATRESID_THETA = 0.0-        DETR_FLIP_ON = 0.0+        AMPS_ENABLE = False+        DETR_EXT = False     proj_asym = None     eigproj_k = 0 if args.ablate else None @@ -1660,7 +1931,19 @@ def main() -> None:                                           "detr_enable", "detr_flip_on", "detr_q_on", "detr_mode",                                           "detr_key", "detr_off_only", "detr_strength", "detr_n_types",                                           "detr_n_genes", "detr_n_on", "detr_n_off", "detr_dp_max",-                                          "detr_rate_gap_before", "detr_rate_gap_after",+                                           "detr_rate_gap_before", "detr_rate_gap_after",+                                           "detr_ext_enable", "detr_ext_tau", "detr_ext_mode", "detr_ext_strength",+                                           "detr_ext_n_types_a", "detr_ext_n_types_b",+                                           "detr_ext_cells_a", "detr_ext_cells_b", "detr_ext_n_genes",+                                           "detr_ext_n_off", "detr_ext_rate_gap_before",+                                           "detr_ext_rate_gap_after", "detr_ext_pb_maxshift",+                                           "ampshrink_enable", "ampshrink_strength", "ampshrink_tau", "ampshrink_decode",+                                           "ampshrink_n_types", "ampshrink_n_pairs", "ampshrink_n_shrunk",+                                           "ampshrink_frac_shrunk", "ampshrink_f_mean", "ampshrink_f_min",+                                           "ampshrink_f_max", "ampshrink_gap_med_before",+                                           "ampshrink_gap_med_after", "ampshrink_gap_frac_over5_before",+                                           "ampshrink_gap_frac_over5_after", "ampshrink_pb_maxshift",+                                           "ampshrink_nnz_delta",                                          "iterdiff_enable", "iterdiff_eta2", "iterdiff_kernel_cos_vs_C",                                          "asympt_zeta", "asympt_cap", "asympt_asym_mean",                                          "asympt_asym_min", "asympt_asym_max",

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 summary:

**做了什么**：PLAN 的 AMPSHRINK（nnz 幅度几何收缩）已完整实现（pooled/sided/gapped 3 种解码 × 0.25–1.5 强度，7 次查分），信号检查通过（gap 中位 9.8%、f_mean 0.91、稀疏模式不变）但**全部劣于父节点**（57.88–63.04 vs 63.17），mmd_u/nbhd/de_score 随强度单调恶化——值通道乘性修正被证否，默认关闭。改交备选机制 **DETR_EXT**：诊断发现父节点 DETR 只覆盖共有类型，37% 输出细胞（24 个单侧独有类型）的检出支持集完全未校正；DETR_EXT 用共有类型细胞测得的逐基因检出漂移作缺失侧代理，对这些细胞做几何插值目标的 OFF-only 置零（shared 模式、s=0.35、逐细胞确定性键）。

**查分结果**（A 半 proxy，19/20 次）：seed 0/1/2 = 64.34 / 63.89 / 63.97，均值 64.07（父 63.17）；variogram 0.01046→0.00751（地板下最弱项）、de_score 0.286→0.321，代价 nbhd 0.0508→0.0523。`--ablate` 与父节点 37 逐位一致（3 种子 sha256 验证）；伪装视图（+1 天平移、乱序 manifest、换路径）输出逐位一致；vec-check 通过；确定性重跑一致。

**下一步最值得试**：① DETR_EXT 的 nbhd 代价或可用"移除条目按 15-NN 空间平滑门控"缓解（只对空间邻域内同类细胞多的做移除）；② shared/all 模式漂移估计在真实括号（两端相邻、共有类型更多）上的迁移性需在 final 前抽查；③ DETR 主线 strength 1.75 细扫仍未做（s=2.0 已知伤 nbhd）。

调研员

标整形均劣化它，暂不触碰。",
  "approach": "步骤：\n1. 在 DETR 之后新增 AMPSHRINK 模块。对每个共有类型（两侧≥10 细胞）× 每个软权重激活基因（与 DETR 相同的 τ=0.15 激活）：\n   a) 从括号阶段实测数据计算 m_a = mean(nnz values at stage a)，m_b = mean(nnz at stage b)；\n   b) 几何目标 m_t = exp((1−t)·ln m_a + t·ln m_b)；\n   c) 当前输出中该型×基因的 nnz 均值 m_cur；\n   d) 若 m_cur > m_t（过表达）：对该型该基因所有 nnz 条目乘收缩因子 f = (m_t / m_cur)^strength；\n   e) 若 m_cur ≤ m_t：不动（只收缩不放大，ON 方向已证明结构性有毒）。\n2. 参数：strength 初值 0.5，网格 [0.25, 0.5, 1.0, 1.5]；t 为插值分数（与 DETR 相同）。\n3. 实现约束：\n   - 随机源：无随机性（确定性比例缩放），无需额外 rng；\n   - 保稀疏：不改变零/非零模式（只做乘法）；\n   - 保序：同一基因内的细胞间相对大小不变（比例缩放），保护空间配对；\n   - pb 保护：逐型逐基因的收缩在伪批量上近似对消（a 侧略缩、b 侧略缩），de_score/de_direction 预期不动。\n4. 单输入回退：无括号时不执行（与 DETR 相同，在 bracket 分支内）。\n5. 快速筛选：先跑 strength=0.5，检查 variogram raw 是否下降≥0.0003（当前 0.01016）；若不动则检查 m_cur 与 m_t 的实际差距（若差距<5%则机制无信号，直接放弃）；若 nbhd raw 上升>0.001 则放弃。\n6. 种子确认：最佳 strength 用 seed 0/1/2 确认，要求三种子 variogram raw 一致低于父节点。\n7. 诊断输出（机制证据）：打印被修正的型×基因数、平均收缩因子、m_cur−m_t 差距分布、pb_maxshift。",
  "expected_groups": ["cell_state"],
  "risks": "1) 位移 α=5 已把 nnz 均值推到接近目标（m_cur≈m_t），则无信号可修——Engineer 应先测量 gap 分布，若中位 gap<5% 立即放弃，不浪费查分；2) 比例缩放改变绝对表达水平，可能轻微影响 mmd_u 的核带宽匹配——监控 mmd_u raw 若上升>0.0002 则回退；3) 若收缩幅度太大（strength>1），可能把正常异质性压扁，损害 variogram 的共变信号——以 variogram raw 非单调为停止信号；4) 与 DETR 的交互：DETR 移除了部分条目，剩余条目的均值已变——AMPSHRINK 必须在 DETR 之后计算 m_cur（顺序依赖）。",
  "family_id": "T2EI-01",
  "mechanism": "对每个共有类型×激活基因的 nnz 均值，向几何时间插值目标 m_t=exp((1−t)ln m_a + t ln m_b) 做单侧比例收缩（只缩不放大），修复位移管线冻结在源阶段的幅度双峰。",
  "vs_constant_shift": "常数位移对每型所有基因加同一 Δ；AMPSHRINK 是逐型×逐基因×逐条目的比例缩放（x·f），缩放因子取决于该基因当前 nnz 均值与几何目标的比值，方向单侧（只缩），且只作用于 nnz 条目。它是乘性、条件化、基因特异的，不是加性全局平移。",
  "mechanism_evidence": "Engineer 应报告：(1) 被修正的型×基因对数及占比（应>10% 否则空转）；(2) 平均收缩因子（预期 0.85–0.98）；(3) 修正前后逐型 nnz 均值与几何目标的差距（应从>10% 降到<5%）；(4) variogram raw 变化≥0.0003；(5) pb_maxshift<0.02（DE 通道保护）；(6) nnz 条目数不变（不改变稀疏模式）。",
  "mechanism_off_control": "T2_AMPSHRINK_ON=0（或 --ablate mechanism）时，模块不执行，输出与父节点 37 逐位一致（sha256 验证）。预期差别：关闭后 variogram raw 回到 0.01016，cell_state 组回到 54.70。",
  "sources": []
}

审查员

REVIEW.json written