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

节点 n41 在终选来历上

NBHDCOH：DETR_EXT 置零改为优先选「15-NN 邻域内该基因表达最低」的细胞（每基因移除数不变），配合强度 0.35→0.45，nbhd 与 variogram 首次同向改善；PLAN 的 NBHDGATE/cap 均被证否。

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 720fc7fcab: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 2e54d3d851。改动的文件：solution/METHOD.md +61 −88、solution/run.py +285 −26

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 82fe83a..abc54af 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,93 +1,66 @@-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）。+NBHDCOH：DETR_EXT 置零改为优先选「15-NN 邻域内该基因表达最低」的细胞（每基因移除数不变），配合强度 0.35→0.45，nbhd 与 variogram 首次同向改善；PLAN 的 NBHDGATE/cap 均被证否。  - 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）。--## PLAN 机制 AMPSHRINK：已实现并证否（7 次查分、3 种解码）--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。--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：--| 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 |--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`。--## 提交的备选机制 DETR_EXT（针对同一弱项 cell_state/variogram）--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）。--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。--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。--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。+- parent: node 39（DETR_EXT，A 半同期锚点 64.34）+- mechanism (submitted): NBHDCOH — neighborhood-coherent OFF selection+- mechanism_off_control: `--ablate mechanism`（或 `T2_NBHDCOH=0`）：跳过 15-NN 键，回到原随机键 u 选择，其余步骤（含 s=0.45）不变；输出 = 未门控 s=0.45 管线，逐位一致（sha256 验证）。注意：因强度从 0.35 重调到 0.45，ablate 输出不等于父节点 39 的输出，但恰好隔离本节点机制（同强度对照 65.04 vs 64.33 = 机制净贡献 +0.71）。++## PLAN 机制（NBHDGATE）的实现与证否++按 PLAN 实现：输出坐标上 cKDTree 取 15-NN（含自身，与 neighborhood_mmd 同尺度），r_i = 邻域内同型同侧单侧细胞占比；r_i < θ 时该细胞置零预算缩为 round(r_i/θ × k_orig)，按超额概率 (p_cur−p_t) 降序保留（确定性排序，基因索引破平）。++先做了 PLAN 第 6 步的分布检查：1842 个单侧细胞 r_i 中位数 0.20、p90 0.47、仅 9.9% > 0.5、坐标无重复 → 门控有区分度，不触发跳过条件。++A 半 seed 0（同一次评分会话内可比；本会话锚点与 node 39 记录不同，父节点同文件重测 64.337，de_score 0.3448→0.3214，跨会话绝对值不可比）：++| 配置 | 榜分 | nbhd raw | variogram raw |+|---|---:|---:|---:|+| 父（s=0.35，无门控） | 64.34 | 0.05232 | 0.007513 |+| gate θ=0.15 | 64.33 | 0.05225 | 0.007573 |+| gate θ=0.25 | 64.29 | 0.05212 | 0.007731 |+| gate θ=0.35 | 64.25 | 0.05196 | 0.007915 |+| gate θ=0.35 + s=0.45 | 64.34 | 0.05279 | 0.007429 |+| cap K=median⁺=14 | 64.05 | 0.05137 | 0.008662 |++结论：门控方向上 nbhd 确实单调改善（θ↑ → nbhd raw↓，支持"混合邻域中的置零更伤 nbhd"假设），但每 1e-4 nbhd raw 只值 ~0.047 分，而同等置零削减造成的 variogram 回退每 1e-4 值 ~0.125 分——**任何削减置零总量的门控都净负**。PLAN 的 cap 备选按原定义（全体单侧细胞移除数的中位数）退化：68% 单侧细胞零移除 → 中位数 0 → 全部置零被禁；改用正计数中位数 K=14 后削 62% 移除，同样净负（64.05）。NBHDGATE 判无效（6 配置、8 次查分，见上表 + gate035_st045/nogate_st045 归因对）。++## 提交的备选机制：NBHDCOH（针对同一弱项 local_spatial）++证据链：s=0.45 同强度下，门控把混合邻域置零换成同质邻域置零，nbhd 每移除伤害 9.3e-8 → 1.7e-7 之差说明**移除发生的位置**比**移除总量**更关键。于是保持 DETR_EXT 每基因移除数不变（变igram/边际检出修复全部保留），只改选哪些细胞：对基因 g，在其非零细胞中按 15-NN 邻域平均表达 NS（**排除自身**：NS'=(15·NS−x_i)/14，预移除输出云、cKDTree、确定性 lexsort 键 (NS', u, index)）取最低的 k 个置零。生物学依据：表达的空间自协/共环境协变是组织的一般性质（Moran 1950, doi:10.2307/2332142 —— 一般机制知识，不含任何保留阶段测量值；所有量均运行时从视图输入现场计算）。移除一个「微环境本就缺该转录本」的条目，对邻域平均表达场扰动最小——node 39 nbhd 代价的来源正是随机键在混合邻域制造的孤立置零噪声。++强度补偿：无门控 s=0.45/0.55 会把 nbhd 打到 0.0537/0.0553（净负 64.33/64.19）；NBHDCOH 保护下 s=0.45 的 nbhd 仅 0.0493，同时 variogram 收益扩大 → 提交 s=0.45。++A 半查分（同会话）：++| 配置 | 榜分 | nbhd | variogram | mmd_u |+|---|---:|---:|---:|---:|+| coh self s=0.35 | 64.84 | 0.04880 | 0.00772 | 0.01035 |+| coh noself s=0.35 | 64.85 | 0.04882 | 0.00768 | 0.01034 |+| coh P=0.5 blend s=0.35 | 64.69 | 0.05017 | 0.00765 | 0.01040 |+| coh self s=0.45 | 65.02 | 0.04928 | 0.00717 | 0.01027 |+| **coh noself s=0.45（提交）** | **65.04** | 0.04929 | 0.00711 | 0.01028 |+| coh noself s=0.55 | 65.04 | 0.05022 | 0.00672 | 0.01043 |+| coh self s=0.55 | 65.05 | 0.05017 | 0.00675 | 0.01037 |+| 无 coh s=0.45（=ablate） | 64.33 | 0.05365 | 0.00704 | 0.01111 |+| 无 coh s=0.55 | 64.19 | 0.05531 | 0.00670 | 0.01139 |++3 种子配对（A 半，final vs parent 同种子）：65.04/64.64/64.68 vs 64.34/63.89/63.97 → **+0.71/+0.75/+0.71，均值 +0.72 ≥ 0.3**；nbhd raw 0.0493–0.0500 vs 0.0521–0.0535，variogram 0.0070–0.0072 vs 0.0075–0.0076，两组首次同向改善（node 39 是 variogram↑ nbhd↓ 交换）。++## 机制生效证据（PLAN mechanism_evidence 对应项）++1. 改变的细胞：s=0.45 时每基因移除数与无 coh 逐位相同（rate_gap_after 一致），但 74% 的被选细胞与随机键不同（`nbhdcoh_sel_move_frac`≈0.74，stderr info）；同一类型两个细胞因邻域不同获得不同移除集合（非型级常数位移/全局重加权，见 PLAN vs_constant_shift）。+2. nbhd raw 0.05365→0.04929（s=0.45 同强度对照），variogram 保持 ≤0.00711：两项同时满足（优于 PLAN 快筛线的相对形式）。+3. 四组分（seed 0）：local_spatial 58.80→60.87（+2.07），cell_state 58.62→59.39（noself s=0.45，mmd_u 0.01111→0.01028），expression_change 62.15→62.16（持平），shape_scale 77.75 不变（坐标逐位不动，d2/occupancy/scale 三项 raw 相同）。+4. 移除数与 r_i 的相关在 gate 实验中为 +0.039（弱正）；NBHDCOH 不依赖 r_i，依赖 NS'（更直接的邻域场扰动度量）。++## 验证过 / 未验证++- 验证过：seed 0 双跑逐位一致；`--ablate` = 无 coh s=0.45 逐位一致；`T2_NBHDCOH_P=1.0` = 父管线逐位一致；θ=0 gate = 父逐位一致；vec-check 全部通过；运行时 ~2s / <1GB（限 28GB/30min）；单输入阶段退路（b is None → 分层照抄，机制自然不激活）；两阶段重构（先收集后写回）对父输出逐位中性。+- 未验证：B 半与伪装视图（harness 执行）；final 真实括号上单侧类型占比若低于本视图 37%，机制收益同比缩小（s=0.45 的选择基于本视图剂量-响应，真实括号未测）；nbhd 的评分器抽样噪声（同会话内查分是确定性的，跨会话锚点不同）。+- 查分使用：20/20。  ## 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）。+- 表达的空间自协变/邻域共变是组织一般性质：Moran (1950) Biometrika 37:17–23, doi:10.2307/2332142（空间自相关度量的一般定义；仅用概念，不用任何数值）。+- neighborhood_mmd 的 15-NN 尺度：任务书评分简报（公开规则），机制与度量尺度对齐是设计选择，不读取评分器。+- 未使用任何保留阶段/保留基因型的测量值；全部统计量（检出率、漂移比、邻域均值、类型集合）运行时从视图 manifest 输入现场计算。diff --git a/solution/run.py b/solution/run.pyindex 370b6bf..c26db10 100644--- a/solution/run.py+++ b/solution/run.py@@ -8,7 +8,39 @@ 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 (39, DETR_EXT, family T2EI-01): the PLAN mechanism (AMPSHRINK:+This node (41, NBHDCOH, family T2EI-01): the PLAN mechanism (NBHDGATE: shrink+each one-side cell's DETR_EXT removal budget by its 15-NN same-type same-side+fraction r_i when r_i < theta) was implemented and FALSIFIED — the r_i+distribution has good discrimination (median 0.20, only 10% > 0.5) and gating+does improve neighborhood_mmd monotonically (raw 0.05232 -> 0.05196 over+theta 0.15/0.25/0.35), but every budget cut trades variogram loss (~0.125+pts per 1e-4 raw) against a ~3x smaller nbhd gain (~0.047 pts per 1e-4 raw),+so all configs land <= parent (A half seed 0, same-session anchors: parent+64.34; theta 64.33/64.29/64.25). The PLAN fallback (per-cell cap at the+median removal count) is degenerate as specified (median over ALL one-side+cells is 0; 68% get no removal) and its positive-median version (K=14,+-62% removals) also loses: 64.05. Attribution runs isolate the mechanism:+at matched strength the coherent selection is worth +0.71 (s=0.45: 65.02 vs+64.33) and +0.85 (s=0.55: 65.04 vs 64.19). The SUBMITTED alternative is+NBHDCOH (neighborhood-coherent OFF selection): keep DETR_EXT's per-gene+removal COUNTS (the variogram/marginal-detection repair) and change only+WHICH cells are zeroed — rank a gene's nonzero cells by their 15-NN+neighborhood-average expression of that gene (self excluded, cKDTree on the+final output coordinates, deterministic lexsort key (NS_excl_self, u, index))+and remove the k lowest — combined with strength 0.35 -> 0.45, which the+coherence now protects (ungated s=0.45/0.55 tank nbhd to 0.0537/0.0553).+Spatial-autocorrelation rationale: dropping a transcript whose micro-+environment already lacks it perturbs the local average-expression field+least (isolated-dropout noise was what nbhd penalized in node 39).+A half seeds 0/1/2: 65.04/64.64/64.68 vs parent 64.34/63.89/63.97 (mean++0.72, consistent); nbhd raw 0.0493-0.0500 vs 0.0521-0.0535 AND variogram+0.0070-0.0072 vs 0.0075-0.0076 — both move together for the first time.+NBHDCOH_SELF=1 (include own value in the key) ties: 65.02. P-blend 0.5:+64.69. --ablate (or T2_NBHDCOH=0) reproduces the ungated s=0.45 pipeline+bit-for-bit (sha256-verified; board 64.33), isolating the mechanism.+See METHOD.md.++Parent 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@@ -337,7 +369,59 @@ DETR_EXT_TAU = float(os.environ.get("T2_DETR_EXT_TAU", "0.15")) # 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"))+DETR_EXT_STRENGTH = float(os.environ.get("T2_DETR_EXT_STRENGTH", "0.45"))+# NBHDGATE (mechanism of this node 41, family T2EI-01): DETR_EXT's OFF-zeroing+# is decided per cell INDEPENDENTLY of space; neighborhood_mmd reads the+# 15-NN average-expression field, so isolated removals inside mixed+# neighborhoods break the expression-position pairing there (node 39: nbhd+# raw 0.04775->0.04996, the only group that got worse). NBHDGATE couples the+# removal budget to the local composition at the metric's own scale: build a+# KD-tree on the FINAL output coordinates, take the K=15 nearest neighbors+# (incl. self, same scale as neighborhood_mmd) of every one-side-only output+# cell, and let r_i = fraction of those neighbors that are same-type+# same-side one-side-only cells. Mode "gate": if r_i < theta the cell's+# removal budget shrinks to k_i = round(r_i/theta * k_orig_i), keeping the+# k_i entries with the LARGEST excess detection count (-dpr), ties by gene+# index (deterministic); r_i >= theta keeps the full DETR_EXT removal.+# Mode "cap" (PLAN fallback): every one-side cell's removals are capped at+# K = cap_frac * median(k_orig over ALL one-side cells), keeping the largest+# excess entries. Cells that keep their full removal are bit-identical to+# DETR_EXT; only gated cells differ.+# T2_NBHDGATE=0 (or --ablate) disables and reproduces parent node 39+# bit-for-bit.+NBHDGATE = os.environ.get("T2_NBHDGATE", "0") == "1"+NBHDGATE_THETA = float(os.environ.get("T2_NBHDGATE_THETA", "0.25"))+NBHDGATE_K = int(os.environ.get("T2_NBHDGATE_K", "15"))+NBHDGATE_MODE = os.environ.get("T2_NBHDGATE_MODE", "gate")+NBHDGATE_CAP = float(os.environ.get("T2_NBHDGATE_CAP", "1.0"))+# absolute override for the cap (0 = use cap_frac × median-of-positive)+NBHDGATE_CAP_K = int(os.environ.get("T2_NBHDGATE_CAP_K", "0"))+NBHDGATE_DIAG = os.environ.get("T2_NBHDGATE_DIAG", "0") == "1"+# NBHDCOH (alternative mechanism of this node 41, after NBHDGATE/cap+# falsification): the gate/cap experiments established the exchange rate —+# every removal-count reduction trades variogram loss (~0.125 pts/1e-4 raw)+# against a smaller nbhd gain (~0.047 pts/1e-4 raw), so ANY budget cut is net+# negative; but at matched budgets, removals in spatially MIXED neighborhoods+# cost ~2x more nbhd per removal than in homogeneous ones (s=0.45 gated vs+# ungated: dnbhd/dremoval 1.7e-7 vs 9.3e-8). NBHDCOH therefore keeps the+# DETR_EXT per-gene removal COUNTS unchanged (variogram/marginal detection+# targets untouched) and changes only WHICH cells are zeroed: among a gene's+# nonzero cells, prefer those whose 15-NN neighborhood-average expression of+# that gene is LOWEST (deterministic sort key (NS, u, index); NS computed once+# from the pre-removal output cloud, same 15-NN scale as neighborhood_mmd).+# Rationale (spatial autocorrelation of expression): removing a transcript+# from a cell whose microenvironment already lacks it perturbs the local+# average-expression field least — the isolated-dropout noise that+# neighborhood_mmd penalizes is removed while the support-channel repair+# stays intact. Blend: P = fraction of genes selected purely by the legacy+# random key u (P=1 → parent bit-for-bit; P=0 → fully coherent).+NBHDCOH = os.environ.get("T2_NBHDCOH", "1") == "1"+NBHDCOH_P = float(os.environ.get("T2_NBHDCOH_P", "0.0"))+# NBHDCOH_SELF=0 excludes the cell's own value from its neighborhood-average+# key (rank by the microenvironment alone, not by own expression), so a+# high-expression cell in a gene-off neighborhood stays selectable. A-half+# seed 0: noself 65.043 vs self 65.022 at s=0.45 (default: noself).+NBHDCOH_SELF = os.environ.get("T2_NBHDCOH_SELF", "0") == "1" # 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@@ -1590,13 +1674,82 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s         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)+        # NBHDGATE (this node 41): 15-NN same-type same-side fraction r_i on+        # the FINAL output coordinates, measured at the neighborhood_mmd+        # scale; see the config block for the mechanism rationale.+        ng = {"nbhdgate_enable": bool(NBHDGATE), "nbhdgate_mode": NBHDGATE_MODE,+              "nbhdgate_theta": NBHDGATE_THETA, "nbhdgate_k": NBHDGATE_K,+              "nbhdgate_cap_frac": NBHDGATE_CAP,+              "nbhdgate_r_mean": None, "nbhdgate_r_p10": None, "nbhdgate_r_p50": None,+              "nbhdgate_r_p90": None, "nbhdgate_frac_r_gt_half": None,+              "nbhdgate_coord_dups": None, "nbhdgate_cap_median": None,+              "nbhdgate_cells_total": 0, "nbhdgate_cells_gated": 0,+              "nbhdgate_n_off_orig": 0, "nbhdgate_n_off_kept": 0,+              "nbhdgate_removal_r_corr": None}+        r_all = None+        coh_info = {"nbhdcoh_enable": bool(NBHDCOH), "nbhdcoh_p": NBHDCOH_P,+                    "nbhdcoh_n_genes_coh": 0, "nbhdcoh_n_genes_leg": 0,+                    "nbhdcoh_ns_zero_frac": None, "nbhdcoh_sel_move_frac": None}+        NS = None+        if NBHDGATE or NBHDCOH:+            coord_out = np.asarray(coords[:ia.size + ib.size], dtype=np.float64)+            n_out = coord_out.shape[0]+            kk_n = int(min(NBHDGATE_K, n_out))+            labs_full = np.concatenate([la_o, lb_o])+            side_full = np.concatenate([np.zeros(ia.size, dtype=np.int8),+                                        np.ones(ib.size, dtype=np.int8)])+            # deterministic KNN at the neighborhood_mmd scale (15-NN incl+            # self): cKDTree, ties broken by stable sort on (dist, index).+            from scipy.spatial import cKDTree+            tree = cKDTree(coord_out)+            dist_n, nbr = tree.query(coord_out, k=kk_n)+            if nbr.ndim == 1:+                nbr = nbr[None, :]+                dist_n = dist_n[None, :]+            nbr = np.ascontiguousarray(nbr, dtype=np.int64)+            # deterministic order among the kk_n nearest: primary distance,+            # secondary cell index (guards against equal-distance ties).+            order = np.lexsort((nbr, dist_n), axis=1)+            nbr = np.take_along_axis(nbr, order, axis=1)+            del tree, dist_n, order+            if NBHDGATE:+                same = (labs_full[nbr] == labs_full[:, None]) & (side_full[nbr] == side_full[:, None])+                r_all = same.mean(axis=1).astype(np.float64)+                ng["nbhdgate_coord_dups"] = int(n_out - np.unique(np.round(coord_out, 6), axis=0).shape[0])+                one_m = ~np.isin(labs_full, list(shared10))+                rr = r_all[one_m]+                if rr.size:+                    ng["nbhdgate_r_mean"] = float(rr.mean())+                    ng["nbhdgate_r_p10"] = float(np.quantile(rr, 0.10))+                    ng["nbhdgate_r_p50"] = float(np.quantile(rr, 0.50))+                    ng["nbhdgate_r_p90"] = float(np.quantile(rr, 0.90))+                    ng["nbhdgate_frac_r_gt_half"] = float((rr > 0.5).mean())+                del same+            if NBHDCOH:+                # neighborhood-average expression field of the PRE-removal+                # output cloud at the neighborhood_mmd scale (15-NN incl self).+                from scipy import sparse as sps+                rows_w = np.repeat(np.arange(n_out, dtype=np.int32), kk_n)+                W = sps.csr_matrix((np.full(rows_w.size, 1.0 / kk_n, dtype=np.float32),+                                    (rows_w, nbr.ravel())), shape=(n_out, n_out))+                Xfull = np.vstack([xa, xb]).astype(np.float32)+                NS = np.asarray(W @ Xfull, dtype=np.float32)+                del W, Xfull, rows_w+            del nbr         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 = [], []+            # Phase 1: collect candidate removals per type group. Identical+            # rng consumption order and identical per-gene selection as the+            # ungated parent (column-wise zeroing never touches other+            # columns, so deferring the writes is bit-for-bit neutral).+            groups_e = []+            move_fr, nsz_fr = [], []             for side in ("a", "b"):                 lo = la_o if side == "a" else lb_o                 xo = xa if side == "a" else xb+                off_g = 0 if side == "a" else int(ia.size)                 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@@ -1608,34 +1761,131 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s                     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()+                    cand = []+                    gidx_e = (np.where(m)[0] + off_g) if (NBHDCOH and NS is not None) else None                     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]+                            nz = np.where(Xc[:, 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+                                coh = (NBHDCOH and NS is not None and kk_n >= 2+                                       and (int(g) * 2654435761) % 1000000 >= NBHDCOH_P * 1e6)+                                if coh:+                                    ns_g = NS[gidx_e[nz], int(g)].astype(np.float64)+                                    if not NBHDCOH_SELF:+                                        ns_g = (ns_g * kk_n - Xc[nz, g].astype(np.float64)) / (kk_n - 1.0)+                                    ord_ = np.lexsort((nz, u[nz], ns_g))[:k]+                                    sel = nz[ord_]+                                    leg = nz[np.argsort(u[nz], kind="stable")[:k]]+                                    move_fr.append(float(np.setdiff1d(sel, leg).size) / k)+                                    nsz_fr.append(float((ns_g <= 0.0).mean()))+                                    coh_info["nbhdcoh_n_genes_coh"] += 1+                                else:+                                    sel = nz[np.argsort(u[nz], kind="stable")[:k]]+                                    if NBHDCOH:+                                        coh_info["nbhdcoh_n_genes_leg"] += 1+                                exc = float(p_cur[int(g)] - p_t[int(g)])+                                cand.extend((int(c), int(g), exc) for c in sel)+                    groups_e.append({"side": side, "m": m, "xo": xo, "off_g": off_g,+                                     "Xc": Xc, "p_t": p_t, "n_c": n_c,+                                     "cand": cand, "gap_b_side": gap_b_side})+            if move_fr:+                coh_info["nbhdcoh_sel_move_frac"] = float(np.mean(move_fr))+            if nsz_fr:+                coh_info["nbhdcoh_ns_zero_frac"] = float(np.mean(nsz_fr))+            # Phase 2: apply the spatial gate (or the cap fallback) and write.+            cap_k = None+            if NBHDGATE and NBHDGATE_MODE == "cap" and groups_e:+                counts_all = []+                for grp in groups_e:+                    cnt = {}+                    for (c, g, exc) in grp["cand"]:+                        cnt[c] = cnt.get(c, 0) + 1+                    counts_all.extend(cnt.values())+                n_cells_tot = int(sum(grp["n_c"] for grp in groups_e))+                pos = np.asarray(counts_all, dtype=np.float64)+                if pos.size:+                    ng["nbhdgate_pos_q"] = [float(np.quantile(pos, q)) for q in+                                            (0.1, 0.25, 0.5, 0.75, 0.9, 0.99)]+                    ng["nbhdgate_pos_max"] = float(pos.max())+                    ng["nbhdgate_pos_frac"] = float(pos.size / max(1, n_cells_tot))+                # median over POSITIVE per-cell counts (over all cells it is 0+                # because ~68% of one-side cells receive no removal at all).+                cap_k = int(round(NBHDGATE_CAP * float(np.median(pos)))) if pos.size else 0+                ng["nbhdgate_cap_median"] = cap_k+            corr_xs, corr_ys = [], []+            for grp in groups_e:+                Xc, cand, m, xo = grp["Xc"], grp["cand"], grp["m"], grp["xo"]+                n_c = grp["n_c"]+                Xn = Xc.copy()+                n_orig = len(cand)+                ng["nbhdgate_n_off_orig"] += n_orig+                ng["nbhdgate_cells_total"] += n_c+                n_off_t = 0+                gated_on = (NBHDGATE and cand+                            and (NBHDGATE_MODE == "cap" or NBHDGATE_THETA > 0.0))+                if gated_on:+                    by_cell = {}+                    for (c, g, exc) in cand:+                        by_cell.setdefault(c, []).append((exc, g))+                    keep = set()+                    n_gated_cells = 0+                    if NBHDGATE_MODE == "cap":+                        for c, lst in by_cell.items():+                            lst.sort(key=lambda x: (-x[0], x[1]))+                            k_i = min(len(lst), cap_k) if cap_k is not None else len(lst)+                            if k_i < len(lst):+                                n_gated_cells += 1+                            keep.update((c, g) for (_e, g) in lst[:k_i])                     else:-                        ext_info["detr_ext_n_types_b"] += 1-                        ext_info["detr_ext_cells_b"] += n_c+                        gidx = np.where(m)[0] + grp["off_g"]+                        rr_g = r_all[gidx]+                        th = NBHDGATE_THETA+                        for c, lst in by_cell.items():+                            lst.sort(key=lambda x: (-x[0], x[1]))+                            r_i = float(rr_g[c])+                            if r_i < th:+                                k_i = int(round(r_i / th * len(lst)))+                                n_gated_cells += 1+                            else:+                                k_i = len(lst)+                            keep.update((c, g) for (_e, g) in lst[:k_i])+                        if len(by_cell) >= 3:+                            corr_xs.extend(float(rr_g[c]) for c in by_cell)+                            corr_ys.extend(float(len(l)) for l in by_cell.values())+                    ng["nbhdgate_cells_gated"] += n_gated_cells+                    for (c, g, exc) in cand:+                        if (c, g) in keep:+                            Xn[c, g] = 0.0+                            n_off_t += 1+                else:+                    for (c, g, exc) in cand:+                        Xn[c, g] = 0.0+                        n_off_t += 1+                ng["nbhdgate_n_off_kept"] += n_off_t+                if n_off_t:+                    xo[m] = Xn+                p_post = (Xn > 0).mean(axis=0)+                gaps_b_e.append(grp["gap_b_side"])+                gaps_a_e.append(float(np.abs(p_post[act_e] - grp["p_t"][act_e]).mean()))+                ext_info["detr_ext_n_off"] += n_off_t+                if grp["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 len(corr_xs) >= 3 and np.std(corr_xs) > 0 and np.std(corr_ys) > 0:+                ng["nbhdgate_removal_r_corr"] = float(np.corrcoef(corr_xs, corr_ys)[0, 1])             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())+        ext_info.update(ng)+        if NBHDCOH:+            ext_info.update(coh_info)      # AMPSHRINK (this node 39): shrink-only geometric interpolation of nnz     # amplitude means, applied AFTER DETR (order matters: DETR removed entries@@ -1851,15 +2101,13 @@ 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 (DETR_EXT coverage extension; AMPSHRINK stays "-                             "off), reproducing parent node 37 bit-for-bit")+                             "node's mechanisms (NBHDCOH neighborhood-coherent OFF selection; "+                             "NBHDGATE spatial gating), reproducing parent node 39 bit-for-bit")     args = parser.parse_args()-    global RESID_GAMMA, SPATRESID_THETA, DETR_FLIP_ON, AMPS_ENABLE, DETR_EXT+    global RESID_GAMMA, SPATRESID_THETA, DETR_FLIP_ON, AMPS_ENABLE, DETR_EXT, NBHDGATE, NBHDCOH     if args.ablate:-        RESID_GAMMA = 0.0-        SPATRESID_THETA = 0.0-        AMPS_ENABLE = False-        DETR_EXT = False+        NBHDGATE = False+        NBHDCOH = False     proj_asym = None     eigproj_k = 0 if args.ablate else None @@ -1937,6 +2185,17 @@ def main() -> None:                                            "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",+                                           "nbhdgate_enable", "nbhdgate_mode", "nbhdgate_theta",+                                           "nbhdgate_k", "nbhdgate_cap_frac", "nbhdgate_r_mean",+                                           "nbhdgate_r_p10", "nbhdgate_r_p50", "nbhdgate_r_p90",+                                           "nbhdgate_frac_r_gt_half", "nbhdgate_coord_dups",+                                           "nbhdgate_cap_median", "nbhdgate_pos_q", "nbhdgate_pos_max",+                                           "nbhdgate_pos_frac", "nbhdgate_cells_total",+                                           "nbhdgate_cells_gated", "nbhdgate_n_off_orig",+                                           "nbhdgate_n_off_kept", "nbhdgate_removal_r_corr",+                                           "nbhdcoh_enable", "nbhdcoh_p", "nbhdcoh_n_genes_coh",+                                           "nbhdcoh_n_genes_leg", "nbhdcoh_ns_zero_frac",+                                           "nbhdcoh_sel_move_frac",                                            "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",

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 verified. Final state of `solution/`:

**做了什么**：实现 PLAN 的 NBHDGATE（15-NN 空间组成门控 DETR_EXT 置零预算，θ=0.15/0.25/0.35）及 cap 备选（中位数退化，改正计数中位数 K=14）——全部证否：门控确实单调改善 nbhd（0.05232→0.05196），但每削减置零量，variogram 损失（~0.125 分/1e-4）约为 nbhd 收益（~0.047 分/1e-4）的 3 倍，6 配置净分全 ≤ 父。据此改交备选机制 **NBHDCOH**：保持每基因移除数不变，只把选细胞从随机键改为「15-NN 邻域内该基因平均表达（排除自身）最低者优先」（cKDTree，确定性 lexsort），并把强度 0.35→0.45（无 coh 时 s=0.45 净负 64.33，coh 保护下 +0.71，归因干净）。

**查分结果**（A 半同会话，20/20 用尽；本会话锚点与 node 39 记录不同，父同文件重测 64.34）：提交配置 seed 0/1/2 = **65.04/64.64/64.68** vs 父 64.34/63.89/63.97，配对差 **+0.71/+0.75/+0.71（均值 +0.72）**；nbhd raw 0.0493–0.0500（父 0.0521–0.0535）且 variogram 0.0070–0.0072（父 0.0075–0.0076）首次同向改善。`--ablate` 逐位还原同强度无 coh 管线（64.33），`T2_NBHDCOH_P=1.0` 逐位还原父管线；seed 0 双跑逐位一致；vec-check 通过；~2s / <1GB。

**下一步最值得试**：① 在 coh 保护下继续扫 s=0.5–0.55 与 noself/self 键（A 半 65.02–65.05 平顶，B 半可能分岔）；② 把 NBHDCOH 同样应用于共有类型的 DETR 置零（目前只覆盖单侧类型，共有类型仍是随机键）；③ occupancy_dice（skill 0.42，最弱单项）仍无人撼动，但历史所有坐标整形均劣化它，需谨慎。

调研员

（raw 0.01016→0.00741，+1.03 分），但 neighborhood_mmd raw 从 0.04775 恶化到 0.04996（skill 0.620→0.609，−0.27 分），是唯一变坏的组（local_spatial −1.09，60.90 vs 父 61.99）。结构原因：DETR_EXT 对 37% 输出细胞（24 个单侧类型）的 OFF 置零逐细胞独立决策，不考虑空间邻域；当单侧类型细胞与共享类型或其他来源细胞混居时，孤立置零在 15-NN 尺度上制造表达-位置不连续，恰好被 neighborhood_mmd（15-NN 平均表达分布的 MMD，25 分+结构门）捕捉。",
  "approach": "步骤：1) 在 DETR_EXT（mode=shared, s=0.35）置零前，对每个单侧类型输出细胞在输出坐标空间建 KD-tree 取 15-NN（与 neighborhood_mmd 同尺度）。2) 统计 15-NN 中同源同侧单侧类型细胞数 h_i，计算占比 r_i = h_i/15。3) 门控：r_i < θ 时，该细胞置零条目数缩减为 k_i = round(r_i/θ × k_orig)，保留超额概率最大的 k_i 条（确定性排序不变）；r_i ≥ θ 时全量置零。4) θ 初值 0.25，搜索 {0.15, 0.25, 0.35}，每个配置跑 seed 0 vec-score。5) 快筛标准：nbhd raw ≤ 0.0489（父 37 与 39 中点）且 variogram raw ≤ 0.0085 才继续；否则试备选——每细胞置零上限 K = median(全部单侧细胞置零数)，超额按超额概率截断。6) 实施前先统计 r_i 分布：若 >80% 细胞 r_i > 0.5（单侧类型已高度聚集），组成门控无区分度，直接跳到备选 K 上限。7) 单输入阶段退路：无括号则无 DETR_EXT，门控自然不激活，输出等同上游。8) 选定配置后跑 seed 0/1/2 确认，均值提升 ≥0.3 才提交。",
  "expected_groups": ["local_spatial"],
  "risks": "1) 门控过度（θ 偏大）导致置零量不足、variogram 回退：Engineer 每个 θ 配置同时监控 variogram raw，>0.009 即放弃该配置。2) 单侧类型细胞在 mix 分层抽样后已按类型聚集，大多数细胞 r_i 本来就高，门控几乎不改变输出（r_i 分布无区分度）：此时应立即转备选 K 上限，不要耗查分次数在无效网格上。3) 净收益可能 <1 分（T2 噪声 ~1 分）：需 3 种子均值确认；若 3 种子均值提升 <0.3 则如实报 NO_CHANGE。4) 15-NN 建在输出坐标上，若坐标有退化（全重合等）会导致门控失效：先检查坐标唯一性。",
  "family_id": "T2EI-01",
  "mechanism": "将 DETR_EXT 的 OFF 置零从逐细胞独立改为空间耦合：在 15-NN 邻域内同源同侧单侧细胞占比低于阈值 θ 时按比例缩减该细胞的置零条目数，使移除决策与 neighborhood_mmd 的度量尺度对齐，防止混合邻域中的孤立置零破坏表达-位置配对。",
  "vs_constant_shift": "不是每型常数位移或全局组成重加权。机制在支持集（零/非零）通道操作，按逐基因超额概率做确定性二值移除，且移除量受细胞空间邻域组成调制——同一类型的两个细胞因邻域不同会获得不同的移除集合，这是任何型级常数位移或全局重加权无法实现的。",
  "mechanism_evidence": "1) 门控前后对比：被缩减的移除应集中在混合邻域（r_i < θ 的细胞），而非均匀分布——输出被门控细胞的占比及其 r_i 分布。2) nbhd raw 从 0.04996 回落（目标 ≤0.0489），variogram raw 保持 ≤0.0085，两项同时满足才说明门控在正确尺度上生效。3) 四组分变化：local_spatial 升、cell_state 持平或微降（≤0.2 分）、expression_change 与 shape_scale 不变（坐标未动）。4) 逐细胞移除数与 15-NN 同源占比的相关系数应为正。",
  "mechanism_off_control": "T2_NBHDGATE=0（或 --ablate mechanism）：跳过 15-NN 计算和门控，所有单侧类型细胞按原 DETR_EXT 逻辑全量置零，输出应与父节点 39 逐位一致（sha256 校验 seed 0/1/2）。"
}
```

审查员

REVIEW.json written