总览 · ← 返回运行 20261003-093415-search-t1-r2-D-s0

节点 n9

composition_trend + A_FATE=0；两阶段组成 log-ratio 机制已实现但在 X3 上为负、默认关闭

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 787f9f1f22: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 911056ceda。改动的文件：solution/METHOD.md +28 −57、solution/run.py +46 −6

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 3bc5139..8fc82da 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,72 +1,43 @@-# composition_trend — type-level composition reweighting + within-type maturity selection+# composition_trend + A_FATE=0；两阶段组成 log-ratio 机制已实现但在 X3 上为负、默认关闭 -**Provenance.** Derived from agent-produced node 33 of run `20261002-034201-search-t1-abc-r1-A-era`-(programs.git `refs/nodes/33`, commit `10400ee`; the program behind the official T1:val 50.1, submissions.tsv-2026-10-02). The mechanism and every constant are the node's own defaults; nothing was re-tuned. What was removed,-for the seed contract only (2026-10-03):+在 node 6（54.01）基础上做两件事：(1) 把型内拟时序权重 A_FATE 从 0.6 改为 0（正式验证了上一节点未提交的声称：X3 48.92 → 52.17，+3.25）；(2) 实现 PLAN 指定的 composition_program 家族机制——由两个输入阶段的类型比例 log-ratio（经验贝叶斯收缩 + 截断）调制类型采样权重（A_COMP），并在两把尺子上量化：该机制在 X3 上双向为负（A_COMP=+0.5: 50.44；+0.25: 51.77；−0.25: 51.66；−0.5: 51.54；0: 52.17），故提交配置默认 A_COMP=0，代码保留、可用 VEC_A_COMP 打开。proxy10 是单输入视图，A_COMP 在其上恒等于 0，输出与 A_FATE=0 的父代码逐字节一致（md5 验证）。 -- the external-maturity axis `A_EXT` (read manifest `source == "external"`; a view-identity read; it never fired on-  final / proxy / patched X3);-- the multi-stage pooling `POOL`, the composition-trajectory axis `A_TR`, the kNN expression smoothing `A_S`, the-  apoptosis / cell-level proliferation terms (all 0 in the node's submitted defaults) and the CellRank fate mode;-- the `inputs_by_time(include_external=False)` base selection: the base is simply the latest input stage by time;-- the `VEC_*` environment overrides (constants are fixed in the code).+## 提交配置（出厂默认，无环境变量） -On every current neutral view (final, proxy, X1, X3–X6) the seed and node 33 select the same cells; see the admission-record for the digest check. The mechanism-split study (`notes/reports/dev/2026-10-02_mech_split_50.md`, candidate-`r1A:CUS` = `r1A:CU`) is the analysis of exactly this procedure.+A_TP=−0.55, A_CM=1.2, **A_FATE=0**, **A_COMP=0**（机制实现但关闭）, K_SHRINK=30, GAMMA=2, K_STRAT=5。+其余方法与 node 3/6 相同：取最新输入阶段的真实细胞，类型层 exp(A_TP·z(型均增殖)·收缩) 重加权 + 细胞层 exp(A_CM·z_met) 选择，分层加权无放回抽样，表达值不改。 -**Naming.** "trend" is descriptive, not a two-stage model: the type weights come from a single snapshot (types whose-cells are, on average, less proliferative get a larger share). On the official final view this moves the predicted-pseudobulk along the previous step's direction (cosine with E9.5 − E8.5: 0.33 for the composition part, 0.34 for the-whole procedure; mech_split §4), and in the E8.5 → E9.5 rehearsal the composition part carries most of the gain-(+10.0 of +11.4 points; mech_split §2.1). Nothing in the code reads a second stage.+## 方法：两阶段组成 log-ratio（A_COMP，本节点新增代码） -Contract: `python run.py --data <view> --out <pred.h5ad> --seed <int>`; CPU only (`EXECUTION.json {"gpu": false}`).+视图有 ≥2 个输入阶段时（按 time 排序取最后两个，只用时间差/顺序，视图无关）：对每个在两阶段都出现的类型 c，p1、p2 为加伪计数（1/类型数）的比例，lr_c = log(p2_c/p1_c)，收缩 lr·m/(m+K_COMP)（m = min(两阶段该型细胞数)，K_COMP=30），截断 |lr|≤1，类型权重乘 exp(A_COMP·lr_c)。只在一个阶段出现的类型 lr=0。单输入视图（proxy10）该项恒为 0。 -## Method+## 机制证据（PLAN mechanism_evidence） -Base = the latest input stage of the view (final: E9.5; proxy: E8.5; rulers: their last input), all cells, its own-`celltype` labels (whatever vocabulary the view provides; `Unknown` is an ordinary type).+X3 视图（E8.75→E9.0，A_COMP=0.5, K_COMP=30）实测每型收缩 log-ratio（run 日志打印）：+IFT-CM +0.69（w×1.41）、AVC-CM +0.56（×1.33）、aSHF +0.21（×1.11）预测扩张；SV-CM −0.65（×0.72）、Unknown −0.28（×0.87）、OFT/RV-CM −0.25（×0.88）预测收缩；无极端值。它确实改变了类型配额（X3 输出 500–652 细胞中每型份额按上表移动），但四项指标全部变差：mmd_u 0.0305→0.0341，de_score −0.043→−0.071，de_direction 0.056→0.034，variogram 0.00134→0.00137。**方向反了也变差**（A_COMP=−0.25/−0.5 亦低于 A_COMP=0），说明 E8.75→E9.0 这 0.25 天的组成趋势不是 E9.0→E9.5 变化的可靠预测（Qiu 心脏数据两阶段样本小、E9.0 仅 2174 细胞），任何偏离输入比例的调制都损失 mmd。 -1. **Cell scores** (generic gene-set means, z-scored over the stage, clipped at ±3): proliferation (34 cell-cycle-   genes), metabolic maturity `z_met` = OXPHOS (13 genes) − glycolysis (10 genes).-2. **Within-type commitment** `z_fate`: HVG 2000 → PCA 30 → kNN 30 → diffusion map 15 → diffusion pseudotime rooted-   at the most progenitor-like cell (argmax 2·z_prolif − z_met, first index on ties); z-scored, the type mean-   removed, z-scored again (only the order inside a type matters).-3. **Weights.** Type layer `w_t = exp(−0.55 · z(type-mean z_prolif))` (types with < 5 cells: 1). Cell layer-   `w_i = exp(1.2 · z_met_i + 0.6 · z_fate_i)`. `w = clip(w_t · w_i, 1e-6, 1e6)`.-4. **Sampling.** n = the stage's cell count clipped to [min_cells, max_cells] (final: 5,118 of 17,057). Per-type-   quota ∝ the type's weight sum (largest remainder, capped by type size, overflow re-apportioned); inside a type 5-   equal-frequency `z_met` strata, stratum quotas ∝ stratum weight sums, Efraimidis–Spirakis weighted sampling-   without replacement inside a stratum. Expression values are copied unchanged.+## 对照（PLAN mechanism_off_control） -Mechanism-off controls (for analysis; not code switches): `A_TP = 0` leaves the input composition (mech_split "U");-`A_CM = A_FATE = 0` with the same per-type quotas is uniform within type (mech_split "C").+- 新代码 VEC_A_COMP=0 与加代码前的 A_FATE=0 运行在 X3 上输出 **md5 逐字节一致**（d36ef183…），确认 A_COMP=0 时无副作用。+- A_FATE=0 对照 A_FATE=0.6（父配置）：X3 52.17 vs 48.92（de_score skill 0.452→0.486，de_direction 0.509→0.523，mmd_u 0.476→0.538，variogram 0.531→0.541，四项全升）；proxy10 60.18 vs 64.19（四项全降：z_fate 选择只对官方单阶段视图有利）。加权节点分 (2·X3+proxy10)/3：54.84 vs 54.01。 -## Data / knowledge used+## 查分记录（A 半，seed 0） -Only the latest input stage of the view. Generic knowledge: textbook cell-cycle, OXPHOS and glycolysis gene sets-(pathway membership, not stage measurements). No held-out stage, no information from (E9.5, E13.5], no external-dataset, no `uns.celltype_palette`, no frozen probe, no pre-trained weights.+| 配置 | X3 | proxy10 | 加权 |+|---|---|---|---|+| 父 node 6（A_FATE=0.6, 无 comp） | 48.92 | 64.19 | 54.01 |+| A_FATE=0, A_COMP=0（**提交**） | **52.17** | 60.18 | **54.84** |+| A_FATE=0, A_COMP=0.5 | 50.44 | （同 60.18） | 53.69 |+| A_FATE=0, A_COMP=0.25 / −0.25 / −0.5 | 51.77 / 51.66 / 51.54 | 同上 | ≤54.57 |+| A_FATE=0.3, A_COMP=0 | 50.31 | 未测 | — |+| A_FATE=0, A_CM=1.6 | 未测 | 57.22 | — | -## Hyper-parameters+## 验证过 / 没验证 -| Name | Value | Where it came from |-|---|---|---|-| `A_TP` | −0.55 | node 33's default (tuned by the agent on the old proxy / proxy2) |-| `A_CM` | 1.2 | node 33's default (same) |-| `A_FATE` | 0.6 | node 33's default (agent scan +0.2 … +1.0 on proxy seed 0, peak at 0.6) |-| `K_STRAT`, `MIN_TYPE_CELLS` | 5, 5 | node 33's defaults |-| HVG / PCA / kNN / diffmap | 2000 / 30 / 30 / 15 | node 33's defaults |+- 验证过：A_FATE=0 在两把尺子上的完整分数；A_COMP 四个幅度（含负向）在 X3 上；A_CM=1.6 在 proxy10 上更差；A_COMP=0 时新代码与父代码路径逐字节一致；输出通过 vec-check。+- 没验证：A_COMP 在双官方输入视图（final/proxy2 形态）上的效果——本次没有这种尺子，X3 的负面结果不一定迁移到官方 E8.5+E9.5→E10.5；K_COMP=120 的完整跑分（时间不足）；A_FATE=0 与 A_CM 的交互（A_CM=0.8/1.6 只在 proxy10 上试过 1.6）。+- 风险：提交配置相对父节点的 +0.83 加权分在 T1 约 2 分的噪声内；A 半/B 半差异也可能吞掉它。选择依据是 X3（权重 2）上四项指标全部同向改善，方向一致性比总分更可信。 -All were chosen on the old single-input proxy (E8.5 → E9.5), which is also where they look best; on the two-input-rulers the mechanism is near neutral (admission record).+## 数据 / 知识来源 -## Known failure modes--- The type layer is a single-snapshot heuristic: it bets that low-proliferation types expand next. Where the next-  step's composition change does not go that way, every ranked metric gets worse (mech_split zero-development control).-- Within-type selection pushes the sample towards OXPHOS-high / late-pseudotime cells; on patched X3 this part was-  −2.4 points (MMD −1.8) on top of the composition (mech_split §3).-- Real cells only: no new expression states, no new cell types.+与 node 3/6 相同：只用视图输入阶段的表达与标签；细胞周期 / OXPHOS / 糖酵解基因集为教科书通路成员知识（非任何阶段的测量）。两阶段组成机制只用视图内两个输入阶段的标签计数。无保留阶段信息、无外部数据集、无 pre-trained 权重。diff --git a/solution/run.py b/solution/run.pyindex 2d139e2..5c0ba0c 100644--- a/solution/run.py+++ b/solution/run.py@@ -35,7 +35,10 @@ from src.task1_temporal.view_io import (  A_TP = float(os.environ.get("VEC_A_TP", "-0.55"))   # type layer: lower mean proliferation -> larger share A_CM = float(os.environ.get("VEC_A_CM", "1.2"))     # cell layer: metabolic maturity (OXPHOS - glycolysis)-A_FATE = float(os.environ.get("VEC_A_FATE", "0.6")) # cell layer: within-type diffusion pseudotime+A_FATE = float(os.environ.get("VEC_A_FATE", "0"))   # cell layer: within-type diffusion pseudotime (0 = off)+A_COMP = float(os.environ.get("VEC_A_COMP", "0"))   # type layer: two-stage composition log-ratio trend (measured harmful on X3; default off)+K_COMP = float(os.environ.get("VEC_K_COMP", "30"))  # empirical-Bayes shrink of the composition log-ratio+LR_CAP = float(os.environ.get("VEC_LR_CAP", "1.0")) # cap on |shrunk log-ratio| K_STRAT = 5       # z_met strata per type MIN_TYPE_CELLS = 5 @@ -206,6 +209,32 @@ def direction_cosine(X, idx, e):     return float(dp @ ee / (na * ne))  +def composition_logratio(prev_adata, uniq, inv, counts):+    """Two-stage composition trend: shrunk log-ratio of per-type proportions between the+    previous and the latest input stage. Types present in only one stage get 0."""+    prev_labels = labels_of(prev_adata) if "celltype" in prev_adata.obs.columns else np.full(prev_adata.n_obs, "all")+    puniq, pinv = np.unique(prev_labels, return_inverse=True)+    pcounts = np.bincount(pinv, minlength=len(puniq))+    ppos = {t: i for i, t in enumerate(puniq)}+    n_prev = max(int(pcounts.sum()), 1)+    n_last = max(int(counts.sum()), 1)+    k = len(uniq)+    pc_last = 1.0 / max(k, 1)+    pc_prev = 1.0 / max(len(puniq), 1)+    lr = np.zeros(k, dtype=np.float64)+    for t, name in enumerate(uniq):+        j = ppos.get(name)+        if j is None:+            continue+        p2 = (counts[t] + pc_last) / (n_last + pc_last * k)+        p1 = (pcounts[j] + pc_prev) / (n_prev + pc_prev * len(puniq))+        raw = float(np.log(p2 / p1))+        m = min(int(counts[t]), int(pcounts[j]))+        shrunk = raw * m / (m + K_COMP)+        lr[t] = float(np.clip(shrunk, -LR_CAP, LR_CAP))+    return lr++ def main() -> None:     ap = argparse.ArgumentParser()     ap.add_argument("--data", required=True)@@ -215,7 +244,8 @@ def main() -> None:      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)-    last = sorted(manifest["inputs"], key=lambda e: float(e["time"]))[-1]   # every input stage treated alike+    inputs = sorted(manifest["inputs"], key=lambda e: float(e["time"]))+    last = inputs[-1]   # every input stage treated alike     adata = read_stage(args.data, last, genes, missing="fill")     X = adata.X     labels = labels_of(adata) if "celltype" in adata.obs.columns else np.full(adata.n_obs, "all")@@ -237,7 +267,17 @@ def main() -> None:     shrink = counts.astype(np.float64) / (counts.astype(np.float64) + K_SHRINK)     z_tp = z_prolif_t * shrink -    w_type = np.exp(A_TP * z_tp)+    # mechanism 3 (composition_program core): two-stage composition log-ratio trend on the type layer+    lr_comp = np.zeros(len(uniq), dtype=np.float64)+    if A_COMP != 0.0 and len(inputs) >= 2:+        prev_adata = read_stage(args.data, inputs[-2], genes, missing="fill")+        lr_comp = composition_logratio(prev_adata, uniq, inv, counts)+        print(f"[comp] A_COMP={A_COMP} K_COMP={K_COMP} per-type shrunk log-ratio:", flush=True)+        for t in range(len(uniq)):+            if lr_comp[t] != 0.0:+                print(f"  {uniq[t]}: n_last={int(counts[t])} lr={lr_comp[t]:+.4f} w_mult={float(np.exp(A_COMP * lr_comp[t])):.4f}", flush=True)++    w_type = np.exp(A_TP * z_tp + A_COMP * lr_comp)     w_type[counts < MIN_TYPE_CELLS] = 1.0     w_cell = np.exp(A_CM * z_met + A_FATE * z_fate)     w = np.clip(w_type[inv] * w_cell, 1e-6, 1e6)@@ -253,15 +293,15 @@ def main() -> None:         print(f"[guard] cosine(dp, trend) = {c:.4f}, genes_used = {int((e != 0).sum())}", flush=True)         if c < 0.0:             damp = float(np.exp(c * GAMMA))     # in (0, 1): shrinks w_type toward 1-            w_type = np.exp(A_TP * z_tp * damp)+            w_type = np.exp(damp * (A_TP * z_tp + A_COMP * lr_comp))             w_type[counts < MIN_TYPE_CELLS] = 1.0             w = np.clip(w_type[inv] * w_cell, 1e-6, 1e6)             idx = stratified_sample(w, z_met, inv, len(uniq), n_out, K_STRAT, rng)             print(f"[guard] fired: damping = {damp:.4f}", flush=True) -    print(f"[shrink] k = {K_SHRINK}; type weights (raw -> shrunk):", flush=True)+    print(f"[shrink] k = {K_SHRINK}; type weights (raw -> final):", flush=True)     for t in range(len(uniq)):-        wr = float(np.exp(A_TP * z_prolif_t[t]))+        wr = float(np.exp(A_TP * z_prolif_t[t] + A_COMP * lr_comp[t]))         ws = float(w_type[t])         if abs(wr - ws) > 1e-6:             print(f"  {uniq[t]}: n={int(counts[t])} w_raw={wr:.3f} w_final={ws:.3f}", flush=True)

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. Summary:

**做了什么**：在 node 6 基础上 (1) 正式验证并采用 A_FATE=0（去掉型内拟时序权重，上一节点声称但未提交的配置）；(2) 实现 PLAN 的 composition_program 家族机制——两输入阶段类型比例 log-ratio（伪计数 + m/(m+K) 经验贝叶斯收缩 + |lr|≤1 截断）调制类型采样权重，并做了正反四个幅度的量化。

**查分结果**（A 半，seed 0；加权 = (2·X3 + proxy10)/3）：
- 提交配置（A_FATE=0, A_COMP=0）：X3 **52.17**（父 48.92，四项指标全升）、proxy10 **60.18**（父 64.19）→ 加权 **54.84** vs 父 54.01
- A_COMP 机制在 X3 上双向为负（+0.5: 50.44；+0.25: 51.77；−0.25: 51.66；−0.5: 51.54），E8.75→E9.0 的 0.25 天组成趋势不能预测 E9.0→E9.5，故默认关闭、代码保留并留全证据；A_COMP=0 对照与父代码路径 md5 逐字节一致
- A_CM=1.6 在 proxy10 更差（57.22），保留 1.2；两视图出厂默认（无环境变量）跑通并过 vec-check，输出与查分文件逐字节一致

**下一步最值得试**：proxy10 的损失全部来自去掉 z_fate（官方单阶段数据上它有效、外部 X3 上有害）——找一个数据自适应、视图无关的 z_fate 可靠性门控（如 DPT 稳定性/标签粒度指标）可能在两把尺子上同时保住收益；以及在双官方输入形态的视图上复核 A_COMP（X3 的负面结论未必迁移到 final）。

调研员

finishes late. (4) Types with very few cells in one stage produce noisy log-ratios — the shrinkage (K_COMP) handles this, but Engineer should print per-type lr_c_shrunk to verify no extreme values.",
  "family_id": "composition_program",
  "mechanism": "Estimate per-type composition change directly from the two input stages' type proportions (log-ratio), shrink toward zero by type cell count (empirical Bayes), and use the shrunk log-ratio to modulate type sampling weights — replacing the proliferation-only proxy with data-driven composition trend. Additionally remove within-type pseudotime weighting (A_FATE=0) which hurts external-data generalization.",
  "vs_constant_shift": "This changes type sampling proportions based on observed two-stage proportion change (with shrinkage and capping), not a uniform shift of all expression values. The per-type weight change is proportional to each type's own log-ratio of proportions, not a constant. Expression values are never modified. The composition change is separately estimable and separately switchable (A_COMP=0 turns it off) from the within-type selection mechanism.",
  "mechanism_evidence": "Engineer should report: (1) per-type lr_c_shrunk values for X3 and final views (which types are predicted to expand/contract and by how much); (2) the resulting type-quota changes vs parent (which types get more/fewer cells); (3) four-group scores and per-ruler metric breakdown for A_COMP=0 vs A_COMP>0 to confirm the composition term is what moves the score; (4) confirm proxy10 output is byte-identical to parent (single-stage fallback).",
  "mechanism_off_control": "Run with VEC_A_COMP=0 VEC_A_FATE=0: this disables the two-stage composition modulation and the within-type pseudotime weighting. Output should differ from parent (which has A_FATE=0.6) only by the removal of z_fate weighting. Then run VEC_A_COMP=0 VEC_A_FATE=0.6: should be byte-identical to parent node 6 output. Any difference indicates a code bug.",
  "sources": []
}
```