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

节点 n35 终选程序?按该运行锁定的规则最终选出的程序；可能是候选节点，也可能由护栏回退到基线。在终选来历上

NO_CHANGE: 相关特征子空间投影位移（k=25–200 五档全部噪声内，retention≥0.94 结构性空转）；备选 RESID 残余双峰坍缩、迭代残差核扩散、逐类型 asym、within 核亦全部证否

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 5c4da635ff: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 e5d7615f0a。改动的文件：solution/METHOD.md +49 −46、solution/run.py +267 −13

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 6a15712..096be85 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,61 +1,64 @@-不对称相关扩散：a 侧位移用全量 Δ'=renorm(Δ+η·C·Δ)，b 侧衰减到 η(1−asym)=0.15（asym=0.85），保护近目标细胞的表达-位置配对；--ablate 逐位还原父节点 31。+NO_CHANGE: 相关特征子空间投影位移（k=25–200 五档全部噪声内，retention≥0.94 结构性空转）；备选 RESID 残余双峰坍缩、迭代残差核扩散、逐类型 asym、within 核亦全部证否 -## 方法（family T2EI-01，PLAN 字面实现）+## 结论 -父节点 31 的 CORRPROJ 对 a/b 两侧施加相同的方向校正 Δ'=renorm(Δ+η·C·Δ|_{|Δ|≥0.15})（C=括号两阶段全细胞基因-基因 Pearson 相关，η=1）。本节点引入 asym∈[0,1]（`T2_PROJ_ASYM`，提交默认 0.85）：+提交的 `run.py` 默认输出与父节点 34 **逐位一致**（X/data、indices、indptr、obsm、seed 的 sha256 全部相同；`--ablate mechanism` 亦逐位一致，已验证）。PLAN 指定的机制（EIGPROJ）已实现并按 PLAN 的判据证否；按任务书要求又实现了 4 个针对同一弱项（cell_state / variogram）的备选机制，共 17 个配置、18 次查分，无一超过父节点基线 +0.3（A 半，seed 0），故按 NO_CHANGE 提交，所有机制代码保留在 run.py 中、默认关闭（env 开关见下），供后续节点复用。 -- a 侧（较早括号阶段）细胞位移用全量扩散 Δ'_a（η_a=1）；-- b 侧（较晚阶段、已接近目标）用衰减扩散 Δ'_b（η_b=η·(1−asym)=0.15）；-- 两侧各自的 Δ' 再经父节点的软权重 w_g=clip(|Δ|/0.25,0,1)、λ=6 逐细胞投影加权、α=5 位移、β=0.2 NN 配对收缩，管线其余部分逐位不变；-- asym=0 时两侧共用同一 Δ'（同一对象），输出与父节点 31 逐位一致（本地 digest 验证）；`--ablate mechanism` 即 asym=0。-- `T2_PROJ_ASYM_SIDE=a` 为反向对照（衰减 a 侧、b 侧全量）。+## 基线（proxy A 半，seed 0，与父节点 34 逐位一致的输出） -坐标、细胞抽取、各向异性整形（aniso γ=0.5）均未动。单输入阶段（无括号对）退路不变：分层 copy，asym 不参与。无新增生物知识条目：相关核完全由视图内两个输入阶段现场计算。+board **62.36**；raw：de_score 0.2857 / de_direction 0.3852 / mmd_u 0.01161 / variogram 0.01342 / nbhd 0.05066 / d2_shape 0.0047 / occupancy_dice 0.8066 / scale_log_ratio 0.0162。（父节点 docstring 中 asym=0.85 的网格值也是 62.36，交叉印证默认输出 = 父节点。） -## 机制生效证据（PLAN mechanism_evidence 四项）+## 1. PLAN 机制 EIGPROJ：实现与证否（5 次查分 + 1 次基线） -1. **neighborhood_mmd raw**（A 半 proxy，seed 0；asym=0 基线 0.05158）：0.25→0.05142，0.5→0.05116，0.75→0.05070，**0.85→0.05066**，0.9→0.05083，1.0→0.05540。最优改善 0.00092 ≥ PLAN 有效阈值 0.0005。seed 1：0.05201→0.05074；seed 2：0.04988→0.04890，三种子方向一致。-2. **mmd_u raw**：0.01222（asym=0）→ 0.01161（asym=0.85），父节点 31 的 mmd_u 微损被逆转（A 半口径）。-3. **Δ'_a 与 Δ'_b 的方向差**：cos(Δ,Δ'_a) 均值 0.8616 vs cos(Δ,Δ'_b) 均值 0.9321，差 0.070 > 0.02 要求；作用于全部 10 个共有类型的 b 侧 2000 个细胞的位移方向。-4. **ablation**：`--ablate mechanism` 输出与父节点 31 的 seed-0 预测逐位一致（.X 与 spatial_3D 字节级 digest 相同）。+实现（按 PLAN）：对括号两阶段全细胞的基因-基因 Pearson 相关矩阵 C（498×498，`np.linalg.eigh`，按 |λ| 取前 k 个特征向量 V_k），在不对称相关扩散之后、软权重之前，对每个共有类型的 Δ' 做 Δ''=V_k(V_kᵀΔ')，再经原有 PROJ_NORM 重归一到 ||Δ||。两侧（a 全量、b 衰减 η(1−asym)=0.15）都投影。`T2_EIGPROJ_K`，k=0 或 ≥498 跳过。 -## 查分记录（共 13 次，余 2；A 半，board 为八项 points 之和）+机制生效证据（PLAN mechanism_evidence 逐条核对）：+- 投影保留率 ||Δ''||/||Δ'||：k=25/50/100/150/200 → **0.959/0.963/0.969/0.974/0.980**，全部高于 PLAN 的证据带 0.6–0.95 → 判据本身已宣告投影近似空转；+- 方向改变角：16.5°/15.6°/14.4°/13.0°/11.4°（在 5–30° 带内但偏小端）；+- variogram raw：0.01337/0.01337/0.01332/0.01333/0.01339，对比基线 0.01342，改善 ≤0.0001，**远低于 PLAN 要求的 ≥0.0005**（目标 ≤0.0125 完全未达）；+- de_direction raw：0.3803–0.3818 vs 基线 0.3852，恶化 0.003–0.005，超过 PLAN 的 <0.001 容忍线。 -| # | 配置 | seed | board | nbhd raw | mmd_u raw | variogram raw | de_dir raw | de_score raw |-|---|---|---|---|---|---|---|---|---|-| 1-2 | asym=0（=父节点，基线校准；重复 1 次浪费额度） | 0 | 62.06 | 0.05158 | 0.01222 | 0.013474 | 0.3813 | 0.2857 |-| 3 | asym=0.25 | 0 | 62.11 | 0.05142 | 0.01215 | 0.013445 | 0.3814 | 0.2857 |-| 4 | asym=0.5 | 0 | 62.18 | 0.05116 | 0.01203 | 0.013403 | 0.3817 | 0.2857 |-| 5 | asym=0.75 | 0 | 62.31 | 0.05070 | 0.01177 | 0.013377 | 0.3816 | 0.2857 |-| 6 | asym=1.0 | 0 | 61.20 | 0.05540 | 0.01362 | 0.014745 | 0.3812 | 0.3214 |-| 7 | asym=0.85 | 0 | 62.36 | 0.05066 | 0.01161 | 0.013420 | — | — |-| 8 | asym=0.9 | 0 | 62.34 | 0.05083 | 0.01158 | 0.013560 | — | — |-| 9 | asym=0.85, side=a（反向） | 0 | 61.53 | 0.05477 | 0.01216 | 0.013539 | — | — |-| 10 | asym=0.85 + CLIPAWARE_TH=0.05 | 0 | 55.80 | 0.07325 | 0.01981 | 0.029699 | 0.3548 | 0.2143 |-| 11 | asym=0（基线） | 1 | 61.56 | 0.05201 | 0.01173 | 0.013803 | — | — |-| 12 | asym=0.85 | 1 | 61.89 | 0.05074 | 0.01111 | 0.013817 | — | — |-| 13 | η=1.25/asym=0.88 | 0 | 62.38 | 0.05058 | 0.01162 | 0.013424 | — | — |-| 14 | η=1.5/asym=0.9 | 0 | 62.38 | 0.05053 | 0.01165 | 0.013424 | — | — |-| 15 | η=2.0/asym=0.925 | 0 | 62.39 | 0.05047 | 0.01166 | 0.013424 | — | — |-| 16 | η=2.0/asym=0.925 | 1 | 61.92 | 0.05052 | 0.01115 | — | — | — |-| 17 | asym=0.85 | 2 | 62.46 | 0.04890 | 0.01097 | — | — | — |-| 18 | asym=0（基线） | 2 | 62.16 | 0.04988 | 0.01151 | — | — | — |+| k | board | variogram raw | mmd_u raw | nbhd raw |+|---|---:|---:|---:|---:|+| 0（=父节点） | 62.36 | 0.01342 | 0.01161 | 0.05066 |+| 25 | 62.24 | 0.01337 | 0.01180 | 0.05118 |+| 50 | 62.24 | 0.01337 | 0.01170 | 0.05137 |+| 100 | 62.37 | 0.01332 | 0.01149 | 0.05091 |+| 150 | **62.43** | 0.01333 | 0.01139 | 0.05053 |+| 200 | 62.41 | 0.01339 | 0.01144 | 0.05049 | -（编号 1-2 为同一预测的两次查分；实际消耗 18/20 额度。）+**证否的结构性原因（供后续节点参考，勿再沿此方向）**：(I+ηC) 与 V_kV_kᵀ 可交换（同为 C 的特征基上的对角算子），扩散 Δ+ηCΔ 已把 Δ 集中到 C 的主特征方向（谱权 (1+ηλᵢ)），事后再投影到同一矩阵的前 k 特征空间只能移除 ≤6% 的能量（k=25 时也只移除 4%）——投影与扩散不是独立约束，PLAN 的前提（"加性校正保留了 Δ 中与 C 正交的分量"）在数学上不成立。k<25 的档位按 PLAN 风险条款跳过（k=50 时 de_dir=0.3816<0.39）。 -## 结论与选择+## 2. 备选机制（针对同一弱项 cell_state / variogram，均已实现、均证否） -- asym 在 0→0.85 单调改善、0.85 处峰、1.0 处崩（b 侧完全无扩散时邻域配对反而变差：两侧位移方向差过大，交界处表达不连续，PLAN 风险 1 在 asym=1 兑现）。三种子配对差 +0.30/+0.34/+0.29 board，收益通道全部是 neighborhood_mmd（+0.05~0.11 points）与 mmd_u（+0.04~0.06），与 PLAN 预期一致。-- 提高 η 同时保持 η_b=0.15（η=1.25/1.5/2.0）只在噪声内微增（+0.02~0.03 board，nbhd raw 差 0.0002 < 0.0005 显著阈值），按 PLAN 保持 η=1 单参数改动。-- **提交配置**：asym=0.85、side=b、η=1、support=0.15（代码默认值，无需环境变量）。-- **clip-aware attenuation（PLAN 辅助修正）被证否**：TH=0.05 时 board 55.80（−6.3），mmd_u 0.0198、variogram 0.0297 均崩——衰减位移使本应被 clip≥0 整流的近零对被保留在原值，破坏了位移管线赖以工作的稀疏整流结构。未再试更低阈值（PLAN 风险注明 <0.02 预期无效，且 0.05 已显示方向性错误而非力度问题）。-- 反向（衰减 a 侧）61.53 劣于基线，交叉验证了机制的方向特异性：需要保护的是 b 侧。+诊断动机：全管线跑完后，输出内每个共有类型中 a 来源与 b 来源细胞的残余质心分离仍达 within-type std 的 **0.55–0.76 倍**（run.py stderr 的 `resid_sep_rel_before`=0.629），即混合双峰是 variogram/mmd 低于地板的直接嫌疑。 -## 验证过 / 未验证+### 2a. RESID 残余双峰坍缩（`T2_RESID_GAMMA/TAU/SIDE/LAMBDA`，8 次查分）+测出输出内的残余分离 s=μ_out(b)−μ_out(a)，用与主管线相同的解码（软权重 w=clip(|s|/0.15,0,1)、nnz-only、clip≥0）坍缩 γ 比例；both-sides 时系数 c_a=γ·n_b/(n_a+n_b)、c_b=γ·n_a/(n_a+n_b) 保持合并均值不动（DE 指标结构性不受扰）。+- both γ=0.5 → **62.27**；γ=1.0 → 62.14；γ=0.5+λ=6 → 62.26；γ=1.0+λ=6 → 61.79；a-only γ=0.5 → 60.90；a-only γ=1.0 → 56.58（dp 被移动、de_score 0.2857→0.1786，证实"保均值"设计必要）。+- 一致规律：**mmd_u 显著改善**（0.01161→0.01081，+0.22 分，全树从未有机制做到）但 **neighborhood_mmd 一致恶化**（0.05066→0.05296…0.05655，−0.27…−0.67 分），净负。结论：残余双峰并非 nbhd/variogram 能吃的免费午餐——把 a/b 两群细胞的表达拉近会破坏表达-位置配对（两群细胞坐标不同）。零率不变（0.9584→0.9586），排除稀疏结构解释。 -- 已验证：seed 0/1/2 三种子 A 半配对比较；ablate 与父节点逐位一致；vec-check 通过；运行 ~9s、内存 <1GB（限额 30min/28GB 内）；对 seed 确定（default_rng）。-- 未验证：B 半正式分（A 半 +0.3 小于 T2 噪声 ~1，但三种子方向一致且 nbhd/mmd_u raw 改善均超阈值，机制真实性由 ablation 与方向特异性支撑）；真实括号（E6.75→E8.0 之外的 t、共有类型数）下 asym=0.85 的最优性——t 变大时 b 侧离目标更远，最优 asym 可能偏小，但 0.5–0.9 区间在 A 半上平坦（62.18–62.36），预期不敏感；未做伪装视图重跑（改动不含任何视图路径/绝对时间依赖，asym 只依赖括号两侧的数据顺序）。+### 2b. ITERDIFF 迭代残差核扩散（父节点建议 #1，`T2_ITERDIFF_ENABLE/ETA2`，2 次查分）+两遍管线：第一遍输出后按型内残差重算相关核 C2（与 C 的余弦仅 0.48，确是不同的核），第二遍用 C2 做方向校正。η2=1.0 → **61.90**（variogram 0.01378 反而变差）；η2=2.0 → 62.30（de_score 0.2857→0.3214 是唯一亮点，但 nbhd 0.05295、variogram 0.01353 均恶化）。证否。 -## 知识来源+### 2c. ASYMPT 逐类型不对称（父节点建议 #2，`T2_ASYMPT_ZETA/CAP`，2 次查分）+asym_type = clip(0.85 + ζ·(||Δ_type||/median −1), 0, 0.95)。ζ=+0.3 → **62.35**（asym 实际范围 0.74–0.91）；ζ=−0.3 → 62.29。均噪声内。证否。 -无新增外部生物知识。相关核 C、逐类型 Δ、细胞抽取全部由视图 manifest 指定的两个输入阶段现场计算；未使用任何保留阶段/保留基因型的测量值。父节点沿用的通路先验接口（prior/ go_bp、reactome）在本节点默认路径（PROJ_MODE=corr）中不读取。+### 2d. within 型内残差核（`T2_PROJ_CORR_KERNEL=within`，1 次查分）+扩散核换成两阶段型内残差相关（去掉组成结构）→ **61.80**，variogram 0.01381 恶化。全细胞核（含组成结构）才是对的方向。证否。++## 验证过什么 / 没验证什么++- 验证：默认输出与 `--ablate` 输出逐位一致且与父节点行为一致（哈希对比）；`vec-check` 通过；seed 0/1/2 均能跑通、确定；全程 1.6 s / <1 GB。所有统计量（相关核、残余分离、类型范数）均从视图输入现场计算，无硬编码、无绝对时间依赖（只用 t 与时间差），满足视图无关。+- 未验证：所有配置只在 proxy A 半 seed 0 上筛过（个别双配置）；没有任何配置达到"双 seed 确认"门槛（≥+0.3），故未花 seed 1/2 查分；B 半行为未测。+- 生物学知识来源：本节点全部机制均为输入数据的现场统计（基因-基因相关、类型均值差、特征分解），未使用任何外部生物学知识条目或保留阶段信息。++## 查分账目（18/20）++基线×2（其中 1 次为重复，浪费）、EIGPROJ k 网格×5、RESID×6、ITERDIFF×2、ASYMPT×2、within×1。++## 给后续节点的建议++1. variogram（skill≈0.39，低于地板）对**类型级 Δ 方向**的一切修正都不敏感（CORRPROJ 是唯一动过它的机制，η≈1 已饱和）；对**细胞级分布**的修正（RESID）会先伤 nbhd。它可能与 nbhd 存在结构冲突：任何把两群细胞表达拉近的操作都同时降 mmd_u 和升 nbhd。若再攻 cell_state，应找只降 mmd_u 不动表达-位置配对的路径（例如按空间邻域条件化地坍缩：只坍缩空间上已混居的 a/b 细胞对——RESID 代码已就位，加一个邻域条件即可试）。+2. de_score 在 ITERDIFF η2=2.0 下 0.2857→0.3214（+0.18 分）是本次唯一超噪声的单项收益，但被 nbhd/variogram 恶化淹没；若单独把"残差核扩散"只用于 DE 通道（如叠加 TSHIFT/PBPROJ 式的伪批量校正而不动细胞级位移），可能变现。+3. 本父节点（34）在 A 半 62.3 附近是一个尖锐局部最优：17 个新配置全部 ≤ +0.07。继续在位移方向/幅度上找增益的期望值很低。diff --git a/solution/run.py b/solution/run.pyindex 4d9377e..4dce15b 100644--- a/solution/run.py+++ b/solution/run.py@@ -8,7 +8,23 @@ 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 (34, PROJ_ASYM, family T2EI-01): the corr-diffusion direction+This node (35, NO_CHANGE, family T2EI-01): the PLAN mechanism (EIGPROJ:+project each type's diffused Δ' onto the top-k eigenvectors of the corr+kernel C, then renormalize) was implemented and falsified: post-diffusion+retention is 0.94-0.98 at every k in 25..200 because (I+ηC) and V_kV_k^T+commute — the diffusion already concentrates Δ into C's leading eigenspace,+so the projection is structurally a near no-op. Board deltas over the parent+(A half, seed 0): k=25/50/100/150/200 → −0.12/−0.12/+0.01/+0.07/+0.05,+variogram raw moves ≤0.0001 (target ≥0.0005). Four alternative mechanisms+targeting the same weakness (cell_state) were also falsified: RESID residual+sub-population collapse (mmd_u 0.01161→0.01081 but nbhd 0.05066→0.053+,+net −0.1..−5.8 over 6 configs), ITERDIFF residual-kernel second-pass+diffusion (61.90/62.30), per-type ASYMPT (62.35/62.29), within-type corr+kernel (61.80). All mechanisms stay in the code behind default-off env+flags; defaults reproduce parent node 34 bit-for-bit, and --ablate disables+them (identical output). See METHOD.md.++Parent node (34, PROJ_ASYM, family T2EI-01): the corr-diffusion direction correction is applied ASYMMETRICALLY by source side: the a-side (earlier bracket stage) displacement uses the full diffusion Δ'_a = renorm(Δ + η·C·Δ), the b-side (later stage, already close to the target) uses the attenuated@@ -162,6 +178,16 @@ PROJ_CORR_KERNEL = os.environ.get("T2_PROJ_CORR_KERNEL", "all") # side is attenuated ("b" default per PLAN, "a" for the reverse cross-check). PROJ_ASYM = float(os.environ.get("T2_PROJ_ASYM", "0.85")) PROJ_ASYM_SIDE = os.environ.get("T2_PROJ_ASYM_SIDE", "b")+# Eigenspace projection of the diffused displacement (EIGPROJ, family T2EI-01,+# this node 35): after the asymmetric corr diffusion (and its support+# restriction), each type's Δ' is projected onto the span of the top-k+# eigenvectors of C (by |eigenvalue|), Δ'' = V_k (V_k^T Δ'), then the existing+# norm-preservation renormalizes to ||Δ||. This removes the component of the+# displacement orthogonal to the tissue's dominant co-variation programs, which+# is where the displacement pipeline injects spurious gene-gene variation+# (variogram skill 0.407 < floor at the parent). k = 0 (or k >= n_genes, or+# --ablate) skips the projection and reproduces parent node 34 bit-for-bit.+EIGPROJ_K = int(os.environ.get("T2_EIGPROJ_K", "0")) # Clip-aware attenuation (optional auxiliary, PLAN): for gene-cell pairs whose # displaced value would fall below CLIPAWARE_TH (i.e. gets rectified/distorted # by the clip≥0), scale that pair's correction by clip(displaced/TH, 0, 1) so@@ -173,6 +199,43 @@ CLIPAWARE_TH = float(os.environ.get("T2_CLIPAWARE_TH", "0")) # to 5× the biological range, which is what collapses mmd_u) while keeping the # correlation structure the diffusion imprints. 0 = off (parent behavior). BRACKET_CLIP = float(os.environ.get("T2_BRACKET_CLIP", "0"))+# Residual sub-population collapse (RESID, family T2EI-01, this node 35's+# submitted mechanism after the PLAN eigenspace projection was falsified):+# after the full expression pipeline (displacement + pair shrinkage), the+# a-derived and b-derived output cells of each shared type still sit apart —+# measured residual type-mean separation is 0.55–0.76 × within-type std on the+# proxy, i.e. the pipeline leaves a strong two-subpopulation bimodality that a+# real intermediate stage does not have. RESID measures that residual+# separation s = μ_out(b-derived) − μ_out(a-derived) per type and collapses a+# fraction γ of it with the SAME decode the main displacement uses: per-cell+# soft gene weights w_g = clip(|s_g|/τ_r, 0, 1)^p (noise genes barely move),+# shift applied only to entries already expressed in the cell, clipped at 0.+# Side coefficients c_a = γ·n_b/(n_a+n_b), c_b = γ·n_a/(n_a+n_b) keep the+# pooled output mean fixed (dp / DE metrics structurally blind, coordinates+# untouched). γ = 0 (or --ablate) reproduces parent node 34 bit-for-bit.+RESID_GAMMA = float(os.environ.get("T2_RESID_GAMMA", "0"))+RESID_TAU = float(os.environ.get("T2_RESID_TAU", "0.15"))+RESID_P = float(os.environ.get("T2_RESID_P", "1.0"))+RESID_MIN_CELLS = int(os.environ.get("T2_RESID_MIN_CELLS", "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:+# protect the b-side cells that are already close to the target)+RESID_SIDE = os.environ.get("T2_RESID_SIDE", "both")+# Iterative (residual-kernel) correlation diffusion (ITERDIFF, family T2EI-01,+# node-34 next_suggestion #1): run the full pipeline once, recompute the gene-+# gene correlation kernel C2 from the WITHIN-TYPE RESIDUALS of the pass-1+# output cloud (what actually remains unexplained after displacement), then+# redo the whole pipeline with C2 in place of C for the direction correction+# (Δ'' = renorm(Δ + η2·C2·Δ), same support/asym/soft-weight decode). η2 = 0 or+# ITERDIFF_ENABLE=0 (or --ablate) keeps the single-step parent behavior.+ITERDIFF_ENABLE = os.environ.get("T2_ITERDIFF_ENABLE", "0") == "1"+ITERDIFF_ETA2 = float(os.environ.get("T2_ITERDIFF_ETA2", "1.0"))+# Per-type asymmetry (ASYMPT, node-34 next_suggestion #2): attenuated-side+# diffusion strength per type, asym_type = clip(asym + ζ·(||Δ_type||/median − 1),+# 0, cap). ζ = 0 (or --ablate) reproduces parent node 34 bit-for-bit.+ASYMPT_ZETA = float(os.environ.get("T2_ASYMPT_ZETA", "0"))+ASYMPT_CAP = float(os.environ.get("T2_ASYMPT_CAP", "0.95")) # Pseudobulk-level prior projection (PBPROJ, this node's submitted mechanism). # Motivation: the stratified (1−t, t) draw weights make the per-cell # displacement cancel EXACTLY in the output pseudobulk, so de_score /@@ -488,7 +551,15 @@ def pathway_boost_graded(abs_d: np.ndarray, M, k: int):     return boost, order[:k]  -def project_delta(d: np.ndarray, M, names, k: int, eta: float, corr: np.ndarray | None):+def eig_basis(corr: np.ndarray, k: int):+    """Top-k eigenvectors of the symmetric corr kernel, by |eigenvalue|."""+    evals, evecs = np.linalg.eigh(corr.astype(np.float64))+    order = np.argsort(-np.abs(evals), kind="stable")[:k]+    return np.ascontiguousarray(evecs[:, order])+++def project_delta(d: np.ndarray, M, names, k: int, eta: float, corr: np.ndarray | None,+                  eigV: np.ndarray | None = None):     """Direction-correct one type's Δ with the prior. Returns (Δ', info)."""     ad_ = np.abs(d).astype(np.float64)     if M is not None:@@ -506,11 +577,23 @@ def project_delta(d: np.ndarray, M, names, k: int, eta: float, corr: np.ndarray             # |Δ| < support keep their measured Δ (no recruited noise genes,             # no support/sign changes outside the DE set)             dp = np.where(np.abs(d64) >= PROJ_SUPPORT, dp, d64)+        eig_ret, eig_cos = None, None+        if eigV is not None:+            pre = dp+            dp = eigV @ (eigV.T @ pre)+            n_pre = np.linalg.norm(pre)+            n_post = np.linalg.norm(dp)+            eig_ret = float(n_post / max(n_pre, 1e-12))+            eig_cos = float(pre @ dp / max(n_pre * n_post, 1e-12))         if PROJ_NORM:             dp = dp * (np.linalg.norm(d64) / max(np.linalg.norm(dp), 1e-12))         dp = dp.astype(np.float32)         cos = float(d @ dp / max(np.linalg.norm(d.astype(np.float64)) * np.linalg.norm(dp.astype(np.float64)), 1e-12))-        return dp, {"cos": cos, "mode": "corr"}+        info = {"cos": cos, "mode": "corr"}+        if eig_ret is not None:+            info["eig_ret"] = eig_ret+            info["eig_cos"] = eig_cos+        return dp, info     return d, {"mode": "none"}  @@ -601,8 +684,29 @@ def type_aniso_reshape(coords, labs_out, ca, cb, la_all, lb_all, t: float):     return c, info  +def corr_from_residuals(X: np.ndarray, labels: np.ndarray) -> np.ndarray | None:+    """Gene-gene Pearson corr of within-type residuals of an arbitrary cloud."""+    X = np.asarray(X, dtype=np.float64)+    labels = np.asarray(labels).astype(str)+    parts = []+    for lab in np.unique(labels):+        m = labels == lab+        if m.sum() >= 10:+            parts.append(X[m] - X[m].mean(axis=0, keepdims=True))+    if not parts:+        return None+    R = np.vstack(parts)+    sd = R.std(axis=0)+    sd[sd < 1e-12] = np.inf+    C = (R.T @ R) / R.shape[0]+    C = C / np.outer(sd, sd)+    np.fill_diagonal(C, 0.0)+    return C.astype(np.float32)++ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: str | None = None,-                 proj_eta: float | None = None, proj_asym: float | None = None):+                 proj_eta: float | None = None, proj_asym: float | None = None,+                 eigproj_k: int | None = None, corr_override: np.ndarray | None = None):     t = float(t)     damp = float(params.get("scale_damp", 1.0))     align = str(params.get("align", "procrustes"))@@ -645,6 +749,8 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s         proj_cos, proj_boost = [], []         if eta != 0.0 and shared and view is not None:             def _kernel():+                if corr_override is not None:+                    return corr_override                 if PROJ_CORR_KERNEL == "within":                     k = corr_kernel_within(stage_a, stage_b)                     if k is not None:@@ -656,12 +762,24 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s                     proj_corr = _kernel()             elif PROJ_MODE == "corr":                 proj_corr = _kernel()+        # Eigenspace projection basis (EIGPROJ, this node): top-k eigenvectors+        # of the same corr kernel used by the diffusion. k >= n_genes spans the+        # full space (V V^T = I) and is treated as off, exactly like k = 0.+        eig_k = EIGPROJ_K if eigproj_k is None else int(eigproj_k)+        proj_eigV = None+        if eig_k > 0 and proj_corr is not None and eig_k < proj_corr.shape[0]:+            proj_eigV = eig_basis(proj_corr, eig_k)         proj_cos_att = []+        eig_rets, eig_coss = [], []          def _prep(d0: np.ndarray, eta_side: float, record: bool) -> np.ndarray:             d = d0             if eta_side != 0.0 and (proj_M is not None or proj_corr is not None):-                d, pinfo = project_delta(d, proj_M, proj_names, PROJ_K, eta_side, proj_corr)+                d, pinfo = project_delta(d, proj_M, proj_names, PROJ_K, eta_side, proj_corr,+                                         eigV=proj_eigV)+                if "eig_ret" in pinfo:+                    eig_rets.append(pinfo["eig_ret"])+                    eig_coss.append(pinfo["eig_cos"])                 if record:                     if "cos" in pinfo:                         proj_cos.append(pinfo["cos"])@@ -679,18 +797,40 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s                 d = np.where(np.abs(d) >= CONVERGE_MIN_DELTA, d, np.float32(0.0)).astype(np.float32)             return d +        # Per-type asymmetry (ASYMPT, node-34 next_suggestion #2): scale the+        # attenuated side's diffusion strength per type by how much that type+        # moves over the bracket: r = ||Δ_type|| / median(||Δ||),+        # asym_type = clip(asym + ζ·(r − 1), 0, cap). ζ = 0 (or --ablate)+        # reproduces the global-asym parent bit-for-bit.+        dnorms = {lab: float(np.linalg.norm((means_b[lab] - means_a[lab]).astype(np.float64)))+                  for lab in shared}+        dmed = float(np.median(list(dnorms.values()))) if dnorms else 0.0+        asym_pts = []         for lab in shared:             d0 = (means_b[lab] - means_a[lab]).astype(np.float32)-            if asym == 0.0 or eta_att == eta:+            if ASYMPT_ZETA != 0.0 and dmed > 0:+                r = dnorms[lab] / dmed+                a_t = float(np.clip(asym + ASYMPT_ZETA * (r - 1.0), 0.0, ASYMPT_CAP))+                eta_att_t = eta * (1.0 - a_t)+                asym_pts.append(a_t)+            else:+                eta_att_t = eta_att+            if asym == 0.0 or eta_att_t == eta:                 d_full = _prep(d0, eta, True)                 delta_a[lab] = d_full                 delta_b[lab] = d_full             elif asym_side == "b":                 delta_a[lab] = _prep(d0, eta, True)-                delta_b[lab] = _prep(d0, eta_att, False)+                delta_b[lab] = _prep(d0, eta_att_t, False)             else:-                delta_a[lab] = _prep(d0, eta_att, False)+                delta_a[lab] = _prep(d0, eta_att_t, False)                 delta_b[lab] = _prep(d0, eta, True)+        if asym_pts:+            conv["asympt_zeta"] = ASYMPT_ZETA+            conv["asympt_cap"] = ASYMPT_CAP+            conv["asympt_asym_mean"] = float(np.mean(asym_pts))+            conv["asympt_asym_min"] = float(np.min(asym_pts))+            conv["asympt_asym_max"] = float(np.max(asym_pts))         if eta != 0.0 and (proj_M is not None or proj_corr is not None):             conv["proj_eta"] = eta             conv["proj_mode"] = "pathway" if proj_M is not None else "corr"@@ -705,6 +845,12 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s             if proj_cos and proj_cos_att:                 conv["proj_cos_full_minus_att"] = float(np.mean(proj_cos) - np.mean(proj_cos_att))             conv["proj_boost_mean"] = float(np.mean(proj_boost)) if proj_boost else None+        if proj_eigV is not None and eig_rets:+            conv["eigproj_k"] = int(eig_k)+            conv["eigproj_retention_mean"] = float(np.mean(eig_rets))+            conv["eigproj_retention_min"] = float(np.min(eig_rets))+            conv["eigproj_cos_mean"] = float(np.mean(eig_coss))+            conv["eigproj_angle_deg_mean"] = float(np.degrees(np.arccos(np.clip(np.mean(eig_coss), -1.0, 1.0))))         if n_part:             conv["soft_enable"] = True             conv["soft_tau"] = SOFT_TAU@@ -938,6 +1084,74 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s                 variso_pb_maxshift=float(np.max(pbshifts)),             ) +    resid_info = {"resid_enable": bool(RESID_GAMMA != 0.0), "resid_gamma": RESID_GAMMA,+                  "resid_tau": RESID_TAU, "resid_p": RESID_P, "resid_n_types": 0,+                  "resid_sep_rel_before": None, "resid_sep_rel_after": None,+                  "resid_shift_norm_mean": None, "resid_zero_rate_before": None,+                  "resid_zero_rate_after": None}+    if RESID_GAMMA != 0.0 and ia.size and ib.size:+        labs_a_r = np.asarray(stage_a.labels).astype(str)[ia]+        labs_b_r = np.asarray(stage_b.labels).astype(str)[ib]+        shared_r = sorted(set(labs_a_r.tolist()) & set(labs_b_r.tolist()))+        rb, ra_, snorms, zr0, zr1 = [], [], [], [], []+        for lab in shared_r:+            ma = labs_a_r == lab+            mb = labs_b_r == lab+            na, nb = int(ma.sum()), int(mb.sum())+            if na < RESID_MIN_CELLS or nb < RESID_MIN_CELLS:+                continue+            A = xa[ma].astype(np.float64)+            B = xb[mb].astype(np.float64)+            s = B.mean(axis=0) - A.mean(axis=0)+            w = np.clip(np.abs(s) / max(RESID_TAU, 1e-12), 0.0, 1.0) ** RESID_P+            if RESID_SIDE == "a":+                ca_, cb_ = RESID_GAMMA, 0.0+            elif RESID_SIDE == "b":+                ca_, cb_ = 0.0, RESID_GAMMA+            else:+                ca_ = RESID_GAMMA * nb / (na + nb)+                cb_ = RESID_GAMMA * na / (na + nb)+            comb0 = np.vstack([A, B])+            sep0 = np.linalg.norm(s)+            spread = max(np.linalg.norm(comb0.std(axis=0)), 1e-12)+            ds = w * s+            if RESID_LAMBDA != 0.0:+                m_ = ds != 0+                if m_.any():+                    dm = ds[m_].astype(np.float64)+                    denom = float(dm @ dm) + CONVERGE_EPS+                    dev_a = A[:, m_].astype(np.float64) - A.mean(axis=0)[m_][None, :]+                    p_a = dev_a @ dm / denom+                    wa = np.clip(1.0 - RESID_LAMBDA * p_a, 0.2, 2.5)+                    dev_b = B[:, m_].astype(np.float64) - B.mean(axis=0)[m_][None, :]+                    p_b = dev_b @ dm / denom+                    wb = np.clip(1.0 + RESID_LAMBDA * p_b, 0.2, 2.5)+                else:+                    wa = np.ones(A.shape[0]); wb = np.ones(B.shape[0])+                Anew = np.clip(A + ca_ * wa[:, None] * ds[None, :], 0.0, None)+                Bnew = np.clip(B - cb_ * wb[:, None] * ds[None, :], 0.0, None)+            else:+                Anew = np.clip(A + ca_ * ds[None, :], 0.0, None)+                Bnew = np.clip(B - cb_ * ds[None, :], 0.0, None)+            Anew = np.where(A > 0, Anew, A)+            Bnew = np.where(B > 0, Bnew, B)+            xa[ma] = Anew.astype(np.float32)+            xb[mb] = Bnew.astype(np.float32)+            comb1 = np.vstack([Anew, Bnew])+            s1 = comb1[na:].mean(axis=0) - comb1[:na].mean(axis=0)+            rb.append(sep0 / spread)+            ra_.append(np.linalg.norm(s1) / spread)+            snorms.append(float(np.linalg.norm(w * s)))+            zr0.append(float((comb0 == 0).mean()))+            zr1.append(float((comb1 == 0).mean()))+            resid_info["resid_n_types"] += 1+        if rb:+            resid_info["resid_sep_rel_before"] = float(np.mean(rb))+            resid_info["resid_sep_rel_after"] = float(np.mean(ra_))+            resid_info["resid_shift_norm_mean"] = float(np.mean(snorms))+            resid_info["resid_zero_rate_before"] = float(np.mean(zr0))+            resid_info["resid_zero_rate_after"] = float(np.mean(zr1))+     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]:@@ -1049,6 +1263,7 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s     info.update(pbproj_info)     info.update(tshift_info)     info.update(clamp_info)+    info.update(resid_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,@@ -1062,10 +1277,15 @@ 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: 'mechanism' (or any name) sets PROJ_ASYM=0 "-                             "(both sides full diffusion), reproducing parent node 31 bit-for-bit")+                        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")     args = parser.parse_args()-    proj_asym = 0.0 if args.ablate else None+    global RESID_GAMMA+    if args.ablate:+        RESID_GAMMA = 0.0+    proj_asym = None+    eigproj_k = 0 if args.ablate else None      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)@@ -1079,8 +1299,33 @@ def main() -> None:     stage_a = read_stage(args.data, a, genes)     stage_b = read_stage(args.data, b, genes)     params = board_params(manifest, "mix", PARAMS, args.seed)-    expr, coords, info = mix_converge(stage_a, stage_b, t, params, CONVERGE_ALPHA,-                                      view=args.data, proj_asym=proj_asym)+    iterdiff_info = {"iterdiff_enable": bool(ITERDIFF_ENABLE and not args.ablate),+                     "iterdiff_eta2": ITERDIFF_ETA2, "iterdiff_kernel_cos_vs_C": None}+    if ITERDIFF_ENABLE and not args.ablate:+        expr1, coords1, info1 = mix_converge(stage_a, stage_b, t, params, CONVERGE_ALPHA,+                                             view=args.data, proj_asym=proj_asym, eigproj_k=eigproj_k)+        rng_mix = np.random.default_rng(args.seed)+        ia_r, ib_r = mix_indices(stage_a.labels, stage_b.labels, t, int(info1["n"]), rng_mix)+        labs_out = np.concatenate([np.asarray(stage_a.labels).astype(str)[ia_r],+                                   np.asarray(stage_b.labels).astype(str)[ib_r]])+        C2 = corr_from_residuals(expr1, labs_out)+        if C2 is not None:+            C1 = corr_kernel(stage_a, stage_b, len(stage_a.genes)).astype(np.float64)+            n1 = np.linalg.norm(C1); n2 = np.linalg.norm(C2.astype(np.float64))+            iterdiff_info["iterdiff_kernel_cos_vs_C"] = float(+                (C1 * C2.astype(np.float64)).sum() / max(n1 * n2, 1e-12))+            expr, coords, info = mix_converge(stage_a, stage_b, t, params, CONVERGE_ALPHA,+                                              view=args.data, proj_asym=proj_asym,+                                              eigproj_k=eigproj_k, corr_override=C2,+                                              proj_eta=ITERDIFF_ETA2)+            info.update(iterdiff_info)+        else:+            expr, coords, info = expr1, coords1, info1+            info.update(iterdiff_info)+    else:+        expr, coords, info = mix_converge(stage_a, stage_b, t, params, CONVERGE_ALPHA,+                                          view=args.data, proj_asym=proj_asym, eigproj_k=eigproj_k)+        info.update(iterdiff_info)     keep = {k: info.get(k) for k in ("t", "n", "rms_a", "rms_b", "out_rms", "n_shared_types", "z_dot", "z_flipped",                                      "align", "alpha", "n_shared_types_converged", "within_type_std_rel",                                      "between_type_mean_dist_rel", "n_from_a", "n_from_b",@@ -1091,6 +1336,15 @@ def main() -> None:                                          "proj_cos_mean", "proj_cos_min", "proj_boost_mean",                                          "proj_asym", "proj_asym_side", "proj_eta_att",                                          "proj_cos_att_mean", "proj_cos_full_minus_att",+                                         "eigproj_k", "eigproj_retention_mean", "eigproj_retention_min",+                                         "eigproj_cos_mean", "eigproj_angle_deg_mean",+                                         "resid_enable", "resid_gamma", "resid_tau", "resid_p",+                                         "resid_n_types", "resid_sep_rel_before", "resid_sep_rel_after",+                                         "resid_shift_norm_mean", "resid_zero_rate_before",+                                         "resid_zero_rate_after",+                                         "iterdiff_enable", "iterdiff_eta2", "iterdiff_kernel_cos_vs_C",+                                         "asympt_zeta", "asympt_cap", "asympt_asym_mean",+                                         "asympt_asym_min", "asympt_asym_max",                                         "pbproj_enable", "pbproj_eta", "pbproj_k", "pbproj_src",                                         "pbproj_n_pathways", "pbproj_n_boosted", "pbproj_boost_max",                                         "pbproj_s_min", "pbproj_s_max", "pbproj_s_mean",

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 done. Final verification passed: default output bit-for-bit identical to parent 34 and to `--ablate`, vec-check ok on seeds 0/1/2, deterministic, 1.6 s runtime, CPU-only.

**做了什么**：实现 PLAN 的 EIGPROJ（扩散后 Δ' 投影到相关核前 k 特征子空间，k=25–200 五档）并按 PLAN 判据证否——retention 0.94–0.98（远超证据带上限 0.95）、variogram raw 仅动 ≤0.0001（要求 ≥0.0005）、board 最好 +0.07（噪声内）；根因是 (I+ηC) 与 V_kV_kᵀ 可交换，扩散已把 Δ 集中到同一特征空间，投影结构性空转。随后实现 4 个针对同一弱项（cell_state）的备选机制并各证否：RESID 残余双峰坍缩（6 配置：mmd_u 0.01161→0.01081 全树最佳单项收益，但 nbhd 0.0507→0.053–0.057 一致恶化，净负；a-only 变体崩到 56.6）、ITERDIFF 残差核二遍扩散（61.90/62.30）、逐类型 ASYMPT（62.35/62.29）、within 核（61.80）。

**查分结果**：18/20 次，基线 62.36（A 半，seed 0），17 个新配置全部 ≤62.43，无一达到 +0.3 确认门槛。按 NO_CHANGE 提交，机制代码保留（默认关闭），METHOD.md 首行为 `NO_CHANGE:` 并附完整证否账目。

**下一步最值得试**：(1) ITERDIFF η2=2 下 de_score 0.2857→0.3214 是本次唯一超噪声单项收益——把残差核扩散只接到伪批量/DE 通道（不动细胞级位移）可能变现；(2) RESID 的 mmd_u 收益真实但被 nbhd 抵消——加空间邻域条件（只坍缩空间上已混居的 a/b 细胞对）或可解耦两者；(3) variogram 对类型级 Δ 方向修正完全不敏感，别再在位移方向上做文章。

调研员

{
  "name": "相关特征子空间投影位移：去除Δ中与组织共变结构正交的分量以修复variogram",
  "motivation": "父节点34的variogram skill仅0.407（低于地板0.5，得分5.09/12.5），是全树最弱单项指标，说明位移管线主动破坏了基因-基因共变结构。cell_state组51.51是四组最低。节点31的相关扩散（Δ+ηCΔ）将variogram raw从0.01382降至0.01300（skill 0.391→0.407），方向正确但力度不足：加性校正保留了Δ中与C正交的分量，该分量在位移后产生虚假基因间变异。所有直接操控方差的尝试（VARISO节点17/19、共变重着色节点21）均失败，因为它们改变表达幅度而非位移方向。本方案从位移方向入手，用投影而非加性校正，是未试过的结构改动。",
  "approach": "在父节点34的相关扩散步骤之后、软权重之前，新增特征子空间投影：(1) 对括号两阶段全细胞计算基因-基因Pearson相关矩阵C（~498×498，已有）；(2) 对C做特征分解，取前k个特征向量组成V_k（498×k）；(3) 对每个共有类型的扩散后位移Δ'执行投影 Δ''=V_k(V_k^T·Δ')，再重归一化到||Δ'||；(4) 投影后的Δ''进入后续管线（support=0.15、w_g、λ=6、α=5、β=0.2、asym=0.85）不变。参数：k的初值与搜索范围——k∈{50,100,150,200}（498基因面板），预期最优在100-150（捕获主要共变模式但不过度约束）。vec-score快速筛选：每个k只跑seed 0，看variogram raw（目标：从0.01296降至≤0.0125）和board；若variogram raw改善≥0.0005且de_score/de_direction/neighborhood_mmd raw恶化均<0.001，再跑seed 1/2确认。单输入阶段退路：无括号对时走分层copy，投影不参与（与父节点一致）。实现注意：特征分解用numpy.linalg.eigh（对称矩阵），取绝对值最大的前k个特征值对应的特征向量；投影是矩阵乘法，498维向量×(498×k)矩阵，计算量可忽略。",
  "expected_groups": ["cell_state"],
  "risks": "1) k过小（<50）时Δ被过度约束，位移方向偏离真实时间变化，de_score/de_direction下降——Engineer应在k=50时检查de_dir raw，若<0.39则跳过更小k。2) C的顶部特征向量可能由少数高变基因主导，投影后Δ集中在这些基因上，反而加剧不均衡——检查投影后Δ的基因支持集是否仍覆盖≥30个基因。3) variogram对稀疏结构（93%零值）敏感，投影可能改变哪些基因被clip到0——检查投影后nnz比例变化，若>2%则需加nnz保护。4) 收益可能仍在噪声内（<1分）——用双seed确认，若board差<0.5则停止。",
  "family_id": "T2EI-01",
  "mechanism": "将逐类型位移向量投影到基因-基因相关矩阵的前k个特征向量张成的子空间上，移除与组织共变结构正交的分量，使位移只沿协同变化的基因程序方向推动细胞，从而保护基因-基因共变结构（variogram）。",
  "vs_constant_shift": "常数位移对型内所有细胞施加相同向量；本机制改变的是位移方向本身——通过子空间投影去除与组织共变结构不一致的分量，不同型的Δ方向不同、投影结果不同，且型内逐细胞投影加权（λ=6）进一步调制幅度。这是方向约束，不是幅度平移。",
  "mechanism_evidence": "1) variogram raw变化：父节点0.01296，目标≤0.0125（改善≥0.0005）；2) 投影保留率||Δ''||/||Δ'||：应在0.6-0.95之间，过低说明k太小；3) 逐型Δ方向改变角度：arccos(Δ'·Δ''/(|Δ'||Δ''|))均值应在5°-30°，为0则投影未生效；4) de_score/de_direction raw不应下降>0.01；5) 三种子配对比较确认方向一致。",
  "mechanism_off_control": "k设为基因总数（498）或设置环境变量T2_EIGPROJ_K=0跳过投影：此时V_k·V_k^T=I，Δ''=Δ'，输出与父节点34逐位一致。用--ablate mechanism触发，验证digest相同。",
  "sources": []
}

审查员

REVIEW.json written