总览 · ← 返回运行 20261003-004922-search-t2-embryo-interp-g24

节点 n12 在终选来历上

解耦表达与坐标来源：坐标走父8的 frac=0.5 均匀分层抽取，表达池走组成插值（ADJ=0），型内用对齐帧最近邻配对供体。

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 58f4b8deac: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 819069b545。改动的文件：solution/METHOD.md +57 −49、solution/run.py +159 −29

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex ff2eb07..8f7aac6 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,60 +1,68 @@-mix 基座上实现 T2EI-01 按型组成线性插值：目标比例 p_T=(1−t)p_A+t·p_B 分配型配额，型内两阶段按 t·adj 分配（提交 adj=0，共享型细胞全取早期），去掉父节点的 τ 相似度偏置；表达与坐标不修改。+解耦表达与坐标来源：坐标走父8的 frac=0.5 均匀分层抽取，表达池走组成插值（ADJ=0），型内用对齐帧最近邻配对供体。  ## 方法 -基座 = 父节点 2/8（seed mix）：括号 E6.75+E8.0 → E7.25（t=0.4），procrustes3d 对齐，-scale_damp=0.5，log 线性 RMS，n=5000（manifest min/max 内）。+基座 = mix（procrustes3d，scale_damp=0.5，n=5000）。机制（family T2EI-01，解耦）： -**PLAN 机制（family T2EI-01，组成插值，实现完整）**：替换 `mix_indices`——-1. 按两阶段型比例线性插值得目标组成 p_T=(1−t)·p_A[T]+t·p_B[T]（单阶段型自然退化为-   (1−t)p_A 或 t·p_B）；-2. 总配额 n 按 p_T 最大余数法分配 c_T，超出可用数时封顶并把余额重分给有余量的型；-3. 型内两阶段分配 n_B,T=round(c_T·t·ADJ)（单阶段型全取该阶段，不足从另一阶段补）；-4. 型内均匀无放回抽样；坐标与表达取被抽细胞自身值。-τ 偏置已按 PLAN 删除（无偏路径）。常数 `COMPOS_ADJ=0.0`（环境变量 `VEC_COMPOS_ADJ`-仅本地测试）；`VEC_COMPOS_INTERP=0` 走关闭对照。+1. **坐标路**（决定输出几何与每型细胞数）：父8 路径，tau=0、T_DRAW_FRAC=0.5，+   n_b=round(t·0.5·n)，两阶段各自按自身型谱均匀分层抽样。+2. **表达路**（决定每个输出细胞的表达值）：父10 的 `_compos_mix`，ADJ=0——+   目标组成 p_T=(1−t)p_A+t·p_B 最大余数配额，共享型全取早期细胞、B-only 型全取+   后期细胞，形成每型的表达供体池。+3. **型内配对**：坐标路配额为准；池不足时从另一阶段同型未用细胞补齐（后期优先），+   多余池细胞丢弃。配对方式 `VEC_PAIR`：+   - `order`（PLAN 原文，按顺序赋值）：**已否决**，见下。+   - `nn`（提交默认）：每个坐标细胞在对齐+缩放帧（align_pair→scale_to_rms，与+     interpolate 内部完全相同的确定性变换）中取同型池内最近邻供体的表达。 -**单阶段退路**：目标无括号时 = seed（最后观测阶段分层抽样），未改动。+单阶段退路（目标无括号）：不变（最后观测阶段分层抽样）。 -## 机制生效证据（proxy, seed 0）+## 与 PLAN 的偏离（如实报告） -- 实际型组成 vs 目标 p_T 的 Pearson = 0.99999（>0.99 达标）。-- vs 父节点 8（frac=0.5 统一抽取）：后期(B)细胞 1000→1049（adj=0）；其中共享型 B 细胞-  519→86，B-only 15 型 481→963（恢复完整插值丰度，如 Forebrain 344→…按 c_T 配额），-  A-only 7 型 1139→854（回到插值丰度）。不同型的 delta 方向不同（非全局重加权）。-- 四组分随 ADJ（A 半，见下表）：shape_scale 随 adj 单调升（66.7→77.3），cell_state-  单调降（50.4→33.0），expression_change/local_spatial 在 adj≤0.2 时优于父。+PLAN 的按顺序配对已实现并查分：ec 62.19、cs 56.06 符合预期，但 **ls 44.84（−13.4）、+ss 61.46（−7.1）**，总分 56.14——坐标几何指标（d2_shape/occupancy/scale_log_ratio）+与关闭对照逐位相同，说明崩掉的是表达-空间局部耦合（neighborhood_mmd 0.055→0.094、+variogram 偏离目标）。触发 PLAN 风险条款 1（ls 降幅>1 即检查），故改为型内最近邻+配对（同一机制家族：仍是"表达走组成插值、坐标走 frac=0.5"的解耦，只改配对规则），+ls/ss 完全恢复且保留 cs 增益。 -## 机制关闭对照（mechanism_off_control）+## 机制生效证据（proxy，seed 0，A 半） -`VEC_COMPOS_INTERP=0`（统一 frac=0.5、τ=0）输出与父节点 8 以-`VEC_BIAS_TAU=0 VEC_T_DRAW_FRAC=0.5` 运行**逐位相同**（X、coords、genes 全等已验证）；-该对照 A 半 59.88 {cs 52.56, ec 60.17, ls 58.27, ss 68.53}（父节点 8 已测）。-机制开启（提交配置 adj=0）A 半 59.40：**未达 PLAN 预期**（ss 未回升至 71+，-反而 −1.8；cs −2.2）。如实报告：组成插值在本代理上没有净胜父节点，差距在噪声（~2 分）内。+- 表达路：共享型 B 供体 86/池、B-only 型 963（ADJ=0 生效）；坐标路 B 细胞 1000+  （frac=0.5·t·n），两路 B 占比 0.15 vs 0.20，来源确实不同。+- nn 配对后坐标细胞与表达供体重合率 0.46（<0.5，解耦实际发生）；expr_from_b=741。+- 补齐：fill_late=325、fill_early=286（坐标配额>池时从另一阶段同型补）。+- 四组分（A 半）： -## 查分记录（A 半，proxy，seed 0，共 4/20 次）--| 配置 | board | cell_state | expr_change | local_spatial | shape_scale |+| 配置 | 榜分 | ec | cs | ss | ls | |---|---:|---:|---:|---:|---:|-| 父 8 对照 τ=0,frac0.5 | 59.88 | 52.56 | 60.17 | 58.27 | 68.53 |-| compos adj=1.0 | 55.63 | 32.95 | 60.79 | 51.50 | 77.28 |-| compos adj=0.5 | 57.72 | 42.26 | 59.84 | 57.54 | 71.25 |-| compos adj=0.2 | 59.19 | 48.02 | 61.46 | 59.47 | 67.81 |-| **compos adj=0.0（提交）** | **59.40** | 50.36 | 61.19 | 59.33 | 66.72 |--关键发现（供后续节点）：① PLAN 假设"按型组成插值回升 shape_scale"被证伪——ss 由共享型-的 B 细胞量驱动（adj 单调），全局组成换成插值丰度反而略降 ss；② cell_state 主要由-总 B 细胞量与 A-only 过采样驱动，父节点 frac=0.5 的近统一抽取已接近 cs/ss 权衡前沿；-③ 组成插值稳定带来 ec(+1.0~1.4) 与 ls(+1.1~1.2)，但 cs 损失更大。--## 验证过 / 没验证--- 验证：关闭对照逐位等于父 8 对照；seed 0 重跑逐位确定；adj=0 输出与查分文件逐位一致；-  n=5000、X 有限、vec-check ok；运行 ~2s。-- 没验证：B 半分；adj∈(0,0.2) 细扫（前沿单调，预期不超父）；组成阻尼 d（t_comp=t·d）-  与型内 split 解耦的第二自由度（时间不足未实现）；真实括号（t=1/3、共有型 11）上-  adj=0 使共享型几乎全取 E7.25 之前的早期细胞，方向与代理一致但未直接检验。-- 生物学知识来源：无外部先验，仅用视图内数据（型计数、时间差）现场计算。-- 视图无关：不读 board/mode/路径/绝对时间；t、括号、n 均由 manifest 现场算出；-  adj 为常数。+| 关闭对照（=父8 tau0 frac0.5） | 59.88 | 60.17 | 52.56 | 68.53 | 58.27 |+| 解耦+顺序配对（PLAN 原文） | 56.14 | 62.19 | 56.06 | 61.46 | 44.84 |+| **解耦+nn 配对（提交）** | **60.88** | 60.83 | 55.87 | 68.53 | 58.30 |+| 解耦+nn，ADJ=0.3 | 60.61 | 60.01 | 54.86 | 68.53 | 59.02 |+| 参考：父10（ coupled ADJ=0） | 60.29 | 62.09 | 51.65 | 67.14 | 60.27 |++机制开 vs 关：+1.00（cs +3.31、ec +0.66、ss/ls 持平），主要来自 cell_state——+解耦让表达组成回到插值丰度（B-only 型足量、共享型年轻化）而几何不受损。+ADJ 扫描（0 / 0.3）确认 0 最优，未扫 0.6（趋势单调向下）。++## 机制关闭对照++`VEC_DECOUPLE=0`：表达与坐标同走父8 均匀抽取（tau=0、frac=0.5），代码路径与父10 的+`VEC_COMPOS_INTERP=0` 分支逐行相同（父10 已验证与父8 `VEC_BIAS_TAU=0+VEC_T_DRAW_FRAC=0.5` 逐位一致）；A 半 59.88 vs 开启 60.88。+`VEC_DECOUPLE=0 VEC_COMPOS_INTERP=1` 复现父10 行为（仅诊断用）。++## 已验证 / 未验证++- 已验证：proxy 视图 seed 0/1 跑通、vec-check 通过、同 seed 逐位确定（X 与坐标+  bit-identical）；无绝对时间/路径依赖（仅用 bracket 的 t、类型标签与表达），时间平移+  不变；nn 配对用的对齐帧在 run 内重算，与 interpolate 内部变换一致。+- 未验证：真实 final 视图（括号类型数更多、n 更大）；B 半分数；nn 配对在共有类型少+  的括号上的稳健性（此时代码自动退回顺序补齐路径，不会崩）。+- 生物学知识来源：无新增外部知识；仅复用父节点的谱系类型标签（视图内提供）。++## 参数++`PARAMS={"align":"procrustes3d","scale_damp":0.5}`；`DECOUPLE=True`、`COMPOS_ADJ=0.0`、+`T_DRAW_FRAC=0.5`、`VEC_PAIR=nn`（环境变量仅本地测试用，默认值即提交配置）。diff --git a/solution/run.py b/solution/run.pyindex 3f037a7..ccac6b4 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,28 +1,42 @@ #!/usr/bin/env python3-"""mix (T2 interpolation) + per-type compositional interpolation (T2EI-01).+"""mix (T2 interpolation) + decoupled expression / coordinate drawing (T2EI-01).  Base pipeline is the seed ``mix`` (parent node 2): bracket the target with the nearest inputs, align frames (procrustes3d), rescale both clouds to exp(log r_a + SCALE_DAMP*t*dlog r), draw n cells stratified by type. -Mechanism (family T2EI-01, compositional interpolation): instead of drawing a-uniform fraction t from the later stage with each stage's own type profile,-build the target type composition p_T = (1-t)*p_A[T] + t*p_B[T], allocate the-output quota per type (largest remainder), then split each type's quota between-stages as n_B,T = round(c_T * t * ADJ) (all from the stage that has the type-when the other lacks it; deficits are filled from the other stage). Within a-type, cells are drawn uniformly without replacement. The tau similarity bias-of parent node 8 is removed (tau = 0 path). Expression values and coordinates-are never modified — only *which* cells are drawn changes.--ADJ controls how many of a shared type's cells come from the later stage:-ADJ=1 reproduces (in expectation) the plain mix stage totals with the exact-interpolated per-type composition; ADJ<1 keeps shared types "younger" while-late-only types still get their full interpolated abundance.--Mechanism-off control: VEC_COMPOS_INTERP=0 routes to parent node 8's uniform-stratified draw with T_DRAW_FRAC=0.5 and tau=0 (bit-identical to parent 8 run-with VEC_BIAS_TAU=0 VEC_T_DRAW_FRAC=0.5).+Mechanism (family T2EI-01, decoupled sources): coordinates and expression come+from *different* draws, paired within each cell type.++- Coordinate path (unchanged from parent node 8 with tau=0): draw n cells with+  a uniform later-stage fraction T_DRAW_FRAC (n_b = round(t*frac*n)), each+  stage stratified by its own type profile. This fixes the output geometry and+  the per-type cell counts of the prediction.+- Expression path (parent node 10's compositional interpolation, ADJ=0): build+  the interpolated per-type composition p_T = (1-t)*p_A[T] + t*p_B[T], allocate+  quotas (largest remainder, capped and redistributed), split each type's quota+  between stages as n_B,T = round(c_T * t * ADJ) (ADJ=0: shared types take+  early-stage cells only; late-only types take their full interpolated+  abundance from the late stage). This yields per-type *pools* of expression+  donor cells.+- Pairing: for each type, the expression pool cells are assigned in order to+  that type's coordinate cells (coordinate-path quota wins). If the pool is+  smaller than the quota, the deficit is filled with unused cells of the same+  type (late stage first, then early stage); surplus pool cells are dropped.+  Expression values are copied verbatim from the donor cells; coordinates are+  copied verbatim (after the shared alignment/rescaling) from the coordinate+  cells.++Rationale: parent node 10 showed ADJ=0 compositional interpolation reliably+gains expression_change (+2.04) but, because it also changed which cells+provide coordinates, lost cell_state (-1.51) and shape_scale (-1.85). Here the+validated expression gain is captured while the geometry stays on parent 8's+frontier draw.++Mechanism-off control: VEC_DECOUPLE=0 routes both expression and coordinates+through the same parent-8 uniform draw (bit-identical to parent node 8 run+with VEC_BIAS_TAU=0 VEC_T_DRAW_FRAC=0.5). VEC_DECOUPLE=0 VEC_COMPOS_INTERP=1+reproduces parent node 10 (composition draw for both).  Single-stage fallback (target not bracketed): unchanged from the seed. """@@ -39,12 +53,13 @@ import numpy as np import src.task2_spatial.methods as methods_mod from src.task2_spatial.methods import interpolate from src.task2_spatial.sample import take+from src.task2_spatial.transport import as_dense from src.task2_spatial.view_io import board_params, interp_bracket, load_manifest, panel_genes, read_stage, write_t2  PARAMS = {"align": "procrustes3d", "scale_damp": 0.5}-COMPOS_INTERP = True-COMPOS_ADJ = 0.0  # later-stage share of a shared type's quota: t * COMPOS_ADJ-T_DRAW_FRAC = 0.5  # only used by the mechanism-off control path+DECOUPLE = True+COMPOS_ADJ = 0.0  # later-stage share of a shared type's expression quota: t * COMPOS_ADJ+T_DRAW_FRAC = 0.5  # coordinate-path later-stage fraction   def _compos_mix(labels_a, labels_b, tt: float, n: int, rng: np.random.Generator, adj: float, diag: dict | None = None):@@ -151,6 +166,107 @@ def _weighted_stratified(labels, n: int, rng: np.random.Generator) -> np.ndarray     return np.concatenate(picks)  +def _uniform_mix(labels_a, labels_b, tt, n: int, frac: float, rng: np.random.Generator):+    """Parent-8 coordinate path: uniform stratified draw with later fraction t*frac."""+    n_b = int(np.clip(int(round(float(tt) * frac * n)), 0, n))+    n_a = int(n) - n_b+    ia = _weighted_stratified(labels_a, n_a, rng) if n_a > 0 else np.array([], dtype=int)+    ib = _weighted_stratified(labels_b, n_b, rng) if n_b > 0 else np.array([], dtype=int)+    return ia, ib+++def _decoupled_expr(stage_a, stage_b, ia, ib, tt: float, adj: float, rng: np.random.Generator, diag: dict,+                    ca=None, cb=None, pair: str = "order"):+    """Rebuild the expression matrix: per-type pools from _compos_mix(adj), paired+    with the coordinate-path cells of the same type.++    pair="order": pool cells assigned in order (PLAN's literal mechanism).+    pair="nn": each coordinate cell takes the spatially nearest same-type pool+    donor in the aligned frame (ca/cb), preserving local expression-spatial+    structure while keeping the compositional pool per type.+    """+    la = np.asarray(stage_a.labels).astype(str)+    lb = np.asarray(stage_b.labels).astype(str)+    n_out = int(len(ia)) + int(len(ib))+    coord_types = np.concatenate([la[ia], lb[ib]]) if n_out else np.array([], dtype="<U64")+    ia_e, ib_e = _compos_mix(la, lb, tt, n_out, rng, adj, diag)+    la_e = la[ia_e] if len(ia_e) else np.array([], dtype=la.dtype)+    lb_e = lb[ib_e] if len(ib_e) else np.array([], dtype=lb.dtype)++    n_genes = stage_a.X.shape[1]+    side = np.zeros(n_out, dtype=np.int8)+    idx = np.zeros(n_out, dtype=np.int64)+    fill_late = fill_early = overlap_num = overlap_den = 0+    for T in np.unique(coord_types):+        pos = np.flatnonzero(coord_types == T)+        poolA = ia_e[la_e == T]+        poolB = ib_e[lb_e == T]+        seq_a = list(poolA.astype(np.int64))+        seq_b = list(poolB.astype(np.int64))+        need = int(len(pos))+        have = len(seq_a) + len(seq_b)+        if have < need:+            usedA = set(seq_a)+            usedB = set(seq_b)+            remB = np.array([j for j in np.flatnonzero(lb == T) if j not in usedB], dtype=np.int64)+            remA = np.array([j for j in np.flatnonzero(la == T) if j not in usedA], dtype=np.int64)+            short = need - have+            if remB.size:+                k = int(min(short, remB.size))+                pick = rng.choice(remB, k, replace=False) if k < remB.size else remB+                seq_b.extend(pick.astype(np.int64))+                short -= k+                fill_late += k+            if short > 0 and remA.size:+                k = int(min(short, remA.size))+                pick = rng.choice(remA, k, replace=False) if k < remA.size else remA+                seq_a.extend(pick.astype(np.int64))+                fill_early += k+        seq_a = seq_a[:need]+        room = need - len(seq_a)+        seq_b = seq_b[:room]+        don_side = np.array([0] * len(seq_a) + [1] * len(seq_b), dtype=np.int8)+        don_idx = np.array(list(seq_a) + list(seq_b), dtype=np.int64)+        coordA = ia[la[ia] == T]+        coordB = ib[lb[ib] == T]+        if pair == "nn" and ca is not None and cb is not None and don_idx.size and don_idx.size != need:+            sel = np.arange(min(need, don_idx.size))+        elif pair == "nn" and ca is not None and cb is not None and don_idx.size:+            from scipy.spatial import cKDTree++            pts_coord = np.vstack([ca[coordA], cb[coordB]]) if (len(coordA) + len(coordB)) else np.zeros((0, ca.shape[1]))+            pts_don = np.vstack(+                [p[k] for p, k in ((ca, don_idx[don_side == 0]), (cb, don_idx[don_side == 1])) if k.size]+            )+            tree = cKDTree(pts_don)+            _, sel = tree.query(pts_coord, k=1)+        else:+            sel = np.arange(min(need, don_idx.size))+        side[pos[: len(sel)]] = don_side[sel]+        idx[pos[: len(sel)]] = don_idx[sel]+        # diagnostic: overlap between coord donors and expr donors for this type+        inter = len(set(coordA.tolist()) & set(don_idx[don_side[sel] == 0].tolist())) + len(+            set(coordB.tolist()) & set(don_idx[don_side[sel] == 1].tolist())+        )+        overlap_num += inter+        overlap_den += need++    expr = np.zeros((n_out, n_genes), dtype=np.float32)+    mA = side == 0+    mB = side == 1+    if mA.any():+        expr[mA] = as_dense(stage_a.X, idx[mA]).astype(np.float32)+    if mB.any():+        expr[mB] = as_dense(stage_b.X, idx[mB]).astype(np.float32)+    np.clip(expr, 0.0, None, out=expr)+    diag["expr_from_b"] = int(mB.sum())+    diag["coord_from_b"] = int(len(ib))+    diag["fill_late"] = int(fill_late)+    diag["fill_early"] = int(fill_early)+    diag["donor_overlap_frac"] = float(overlap_num / max(overlap_den, 1))+    return expr++ def main() -> None:     parser = argparse.ArgumentParser()     parser.add_argument("--data", required=True)@@ -158,7 +274,8 @@ def main() -> None:     parser.add_argument("--seed", type=int, default=0)     args = parser.parse_args() -    compos = os.environ.get("VEC_COMPOS_INTERP", "1" if COMPOS_INTERP else "0") not in ("0", "", "false")+    decouple = os.environ.get("VEC_DECOUPLE", "1" if DECOUPLE else "0") not in ("0", "", "false")+    compos_only = (not decouple) and os.environ.get("VEC_COMPOS_INTERP", "0") not in ("0", "", "false")     adj = float(os.environ.get("VEC_COMPOS_ADJ", COMPOS_ADJ))     t_draw_frac = float(os.environ.get("VEC_T_DRAW_FRAC", T_DRAW_FRAC)) @@ -175,21 +292,22 @@ def main() -> None:     stage_b = read_stage(args.data, b, genes)     params = board_params(manifest, "mix", PARAMS, args.seed) -    diag: dict = {"compos_interp": bool(compos), "adj": adj, "t_draw_frac": t_draw_frac}+    diag: dict = {"decouple": bool(decouple), "compos_only": bool(compos_only), "adj": adj, "t_draw_frac": t_draw_frac}+    holder: dict = {}     orig_mix = methods_mod.mix_indices-    if compos:++    if compos_only:          def patched(labels_a, labels_b, tt, n, rng):             ia, ib = _compos_mix(labels_a, labels_b, tt, n, rng, adj, diag)+            holder["ia"], holder["ib"] = ia, ib             return ia, ib      else:          def patched(labels_a, labels_b, tt, n, rng):-            n_b = int(np.clip(int(round(float(tt) * t_draw_frac * n)), 0, n))-            n_a = int(n) - n_b-            ia = _weighted_stratified(labels_a, n_a, rng) if n_a > 0 else np.array([], dtype=int)-            ib = _weighted_stratified(labels_b, n_b, rng) if n_b > 0 else np.array([], dtype=int)+            ia, ib = _uniform_mix(labels_a, labels_b, tt, n, t_draw_frac, rng)+            holder["ia"], holder["ib"] = ia, ib             return ia, ib      methods_mod.mix_indices = patched@@ -198,6 +316,18 @@ def main() -> None:     finally:         methods_mod.mix_indices = orig_mix +    if decouple and "ia" in holder:+        from src.task2_spatial.frame import align_pair, log_interp, rms_radius, scale_to_rms++        pair = str(os.environ.get("VEC_PAIR", "nn"))+        aligned_a, aligned_b, _ = align_pair(stage_a.coords, stage_b.coords, stage_a.labels, stage_b.labels, str(params.get("align", "procrustes")))+        target_rms = log_interp(rms_radius(stage_a.coords), rms_radius(stage_b.coords), t, float(params.get("scale_damp", 1.0)))+        ca = scale_to_rms(aligned_a, target_rms)+        cb = scale_to_rms(aligned_b, target_rms)+        rng_e = np.random.default_rng([int(args.seed), 20261003])+        expr = _decoupled_expr(stage_a, stage_b, holder["ia"], holder["ib"], t, adj, rng_e, diag, ca=ca, cb=cb, pair=pair)+        diag["pair"] = pair+     keep = {k: info.get(k) for k in ("t", "n", "rms_a", "rms_b", "out_rms", "n_shared_types", "z_dot", "z_flipped", "align")}     print(json.dumps({"bracket": [a["stage"], b["stage"]], **keep, **diag}, default=float), file=sys.stderr)     write_t2(args.out, expr, coords, genes, seed=args.seed)

调研来源?调研员查到并用到的知识条目和文献检索结果（只列标题和编号）。

用到的知识库条目

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

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

只给统计和最后回答的摘录；完整对话请到原始记录位置里列出的文件看。

分析员

ANALYSIS.json written

工程师

**Done.** Node 12 submitted: decoupled expression/coordinate sources on the mix base.

- **What**: Coordinates from parent-8's uniform frac=0.5 stratified draw (unchanged geometry); expression donors from parent-10's compositional interpolation (ADJ=0), paired within type by nearest neighbor in the aligned frame. PLAN's literal in-order pairing collapsed local_spatial (−13) and shape_scale (−7), so I switched to nn pairing per PLAN's risk clause — documented as a deviation.
- **Scores (A half, 5 queries)**: off-control 59.88; order-pairing 56.14 (rejected); **nn-pairing 60.88** (ec 60.83, cs 55.87, ss 68.53, ls 58.30) vs parent 60.29; adj=0.3 60.61 (kept adj=0). Determinism verified bit-identical across reruns; vec-check passes on seeds 0/1.
- **Next best directions**: sweep T_DRAW_FRAC (coord path) now that expression is decoupled — cs/ss frontier may shift; or blend nn-paired compos pool with uniform donors per type to trade ec↔cs.

调研员

ion": "节点10证明组成插值ADJ=0稳定带来ec+2.04（62.09 vs 60.05），但同时改变了坐标来源导致cs−1.51、ss−1.85。ANALYSIS建议①：表达矩阵走组成插值（捕获已验证的ec增益），坐标仍走父8的frac=0.5抽取（保持cs 53.16/ss 68.99），检验ec增益能否在cs不受损时转化为净榜分。预期净增+1.5~2.5（ec+2，cs/ss不损）。",
  "approach": "步骤：\n1. 复制父10的run.py为基座，保留_compos_mix函数和坐标对齐流程。\n2. 新增解耦逻辑（环境变量VEC_DECOUPLE，默认1）：\n   a. 坐标抽取（coord_indices）：使用父8的均匀分层抽样（frac=0.5，tau=0），即对每个型按两阶段细胞数比例统一抽50%后期。这决定输出的坐标和型标签。\n   b. 表达抽取（expr_indices）：使用_compos_mix（ADJ=0），按插值组成p_T分配型配额，共享型全取早期、B-only型全取后期。这决定输出的表达矩阵。\n   c. 型内配对：对每个型，将expr_indices中该型的细胞按顺序赋给coord_indices中该型的细胞（型内细胞数相同时一一对应；若不等，多出的坐标细胞复用该型最后一个表达细胞的表达，多出的表达细胞丢弃）。实际实现中，两路抽取都按相同的目标总数n=5000、相同的型配额（来自frac=0.5的型比例）抽取，确保每型细胞数一致。\n   注意：坐标路的型配额由frac=0.5的两阶段型比例决定（父8逻辑）；表达路的型配额由插值组成p_T决定。两者型配额可能不同——此时以坐标路的型配额为准（因为输出细胞数由坐标路决定），表达路在每个型内按坐标路配额数抽取表达细胞（从_compos_mix给出的该型细胞池中无放回抽）。若该型表达池不足，从另一阶段同型细胞补齐。\n3. 单阶段退路：目标无括号时，两路均退化为种子分层抽样，行为不变。\n4. vec-score快速筛选：先跑VEC_DECOUPLE=1（机制开）和VEC_DECOUPLE=0（关闭对照），比较四组分。若ec提升≥1.5且cs不降，提交；否则回退。\n5. 关键参数：ADJ=0.0（固定，已由父10验证为ec最优）；frac=0.5（坐标路，已由节点5/8验证）。无需扫描。\n6. 时间预算：实现~10分钟，测试~5分钟（两次vec-score），总计<30分钟。",
  "expected_groups": ["expression_change", "cell_state"],
  "risks": "1. 表达与坐标来源不一致可能损害local_spatial（局部表达-空间相关性被打破）——但父10中ls随组成插值+0.74，且型内配对保持型内表达一致性，风险可控。Engineer应在第一次查分时检查ls是否下降>1。2. 型内配对逻辑实现错误（型细胞数不匹配）导致表达错位——Engineer应验证每型坐标细胞数==表达细胞数。3. ec增益在B半不重现（A/B半差异）——差距<1分时多查一次确认。",
  "family_id": "T2EI-01",
  "mechanism": "将组成插值（ADJ=0，共享型表达全取早期、B-only型全取后期）仅作用于表达矩阵的来源选择，坐标仍由frac=0.5均匀抽取决定，实现表达组成与几何组成的解耦。",
  "vs_constant_shift": "常数位移对所有细胞施加相同向量；本机制按型差异化处理——共享型表达取早期细胞（表达更\"年轻\"），B-only型表达取后期细胞（表达更\"成熟\"），且只改表达来源不改坐标，不同型的改变方向和幅度不同。",
  "mechanism_evidence": "Engineer应检查：1. 表达路中共享型的B细胞占比≈0（ADJ=0生效），B-only型占比=100%；2. 坐标路中B细胞占比≈frac=0.5（与父8一致）；3. 四组分对比：ec应≥61.5（+1.5以上），cs应≥52.5（不降），ss应≥68（不降）；4. 表达路与坐标路的细胞索引集合不同（共享型中overlap<50%），证明解耦实际发生。",
  "mechanism_off_control": "VEC_DECOUPLE=0时，表达和坐标使用同一个frac=0.5抽取结果（即父8的tau=0, frac=0.5路径），输出应与父节点8以VEC_BIAS_TAU=0 VEC_T_DRAW_FRAC=0.5运行逐位相同。预期差别：关闭时ec≈60.05，开启时ec≈62；cs/ss不变。",
  "sources": []
}

审查员

REVIEW.json written