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

节点 n20

实现 sign-agreement 融合（稠密 Δμ 基 + 网络符号门控，冲突基因按 γ 收缩）；X3 A 半未超 node 14，按 PLAN step8 出厂纯 node-14 稠密 Δμ 路径（NET_BETA=0）。

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 e0ebf6d69b: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 dc1acabb7b。改动的文件：solution/METHOD.md +83 −79、solution/run.py +51 −7

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex f17a98b..c53a929 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,85 +1,89 @@-# 在父节点基础上，把两阶段型内位移的方向来源从原始伪批量 Δμ 换成 CollecTRI TF→靶基因网络传播方向（typed，β_net=4，EB 收缩+稀疏门控；仅双输入视图激活，单输入与父逐字节一致）--Parent (node 14, 56.39): composition_trend weighted sampling + per-type re-anchoring α=3.0 +-two-stage typed pseudobulk-Δμ direction shift (β=1). PLAN target: lift de_recovery (X3 de_score-−0.026, below floor) by replacing the noisy raw Δμ with a CollecTRI TF→target network-propagated-direction. Family: composition_program; mechanism = network-constrained gene-level direction shift.--## Method change vs parent--Active only when the view has ≥2 input stages (uses the last two, sorted by time). The shift-direction source changes from raw per-type pseudobulk Δμ to:--1. **Network load** (`load_tf_network`): `<view>/prior/tf_regulons/*.tsv.gz` (sorted glob, relative-   layout only — view-agnostic). CollecTRI mouse: 37,497 signed TF→target edges; keeps-   `sign_known==1`, `mor∈{+1,−1}`, both genes in the panel. On X3: 357 TFs kept (≥10 panel targets,-   TF detected in ≥10 cells in both stages), 25,013 edges, 5,208 regulated genes. Missing file or-   <10 kept TFs → automatic fallback to the parent typed-Δμ path (PLAN off-control guarantee).-2. **TF activity change**: Δact[type, tf] = mean of the per-type pseudobulk delta (low-detection-   genes fall back to the global delta, as in the parent) over the tf's panel targets. Aggregating-   ~70 targets/TF is what stabilizes the estimate at X3's ~65 cells/type — the point of the PLAN.-3. **Propagation**: dir[type, g] = Σ_{tf→g} sign(tf→g)·Δact[type, tf]; genes without regulators-   get dir=0 (network sparsity is the gate's skeleton).-4. **EB shrink + rescale + sparse gate**: w_g = n_reg(g)/(n_reg(g)+K_net), K_net=5; dir rescaled so-   median|dir| over regulated genes equals the parent's median|Δμ_shrunk| (β_net comparable to the-   parent's β units); gate |dir| > ZMIN_net·median(|dir| nonzero), ZMIN_net=1.5.-5. **Apply**: x[i,g] += clip(β_net·dir_g, ±MAX_SHIFT=1.0) at existing nonzeros only, clip ≥0-   (sparsity/covariation preserved, same writer as the parent). No RNG in the block; deterministic.--Env knobs: VEC_NET_BETA (0 = off control) / VEC_NET_K / VEC_NET_ZMIN / VEC_NET_MODE / VEC_NET_STACK.--## Mechanism-off control (verified)--- VEC_NET_BETA=0 on X3: X output **byte-identical** to parent node-14 code run (h5 data/indices/-  indptr compared, twice, incl. after the final restructure).-- Single-input view (proxy, E8.5 only): whole block inactive; stderr shows n_input_stages=1 and no-  gene_shift; output path identical to parent's → proxy10 score unchanged (58.80 parent B-half).-  vec-check ok on both views; X3 final run byte-identical to the scored β=4 prediction.--## Screen (vec-score A-half, X3, seed 0; parent node14 = 54.88, node10 = 54.51)--| variant | X3 board | de_score | de_dir | mmd_u | variogram |+实现 sign-agreement 融合（稠密 Δμ 基 + 网络符号门控，冲突基因按 γ 收缩）；X3 A 半未超 node 14，按 PLAN step8 出厂纯 node-14 稠密 Δμ 路径（NET_BETA=0）。++Family: composition_program. Op: improve on node 17 (parent 56.12 = X3 54.78 / proxy10 58.80).+Node 14 (grandparent, best B-half) = 56.39 (X3 55.19 / proxy10 58.80).++## What the PLAN asked, what I built++PLAN mechanism = **sign-agreement gated fusion**: keep node 14's dense per-type pseudobulk Δμ as the+shift *base* (preserves cell-state coverage for mmd_u), and use the CollecTRI TF→target network+direction (exactly as node 17: typed, EB shrink K_net=5, rescaled to parent units, sparse gate+ZMIN_net=1.5) only as a **binary sign filter**:++- gene entry with network dir==0 → Δμ unchanged;+- sign(dir) == sign(Δμ) → Δμ unchanged (network confirms);+- sign(dir) != sign(Δμ) → Δμ ← Δμ·(1−γ) (shrink toward 0; γ = suppression strength).++Then apply with node 14's writer `apply_gene_shift_typed` (nonzero entries only, β=SHIFT_BETA=1.0,+ZMIN=1.5, MAX_SHIFT=1.0, clip ≥0). Implemented as `fusion_gate()` + a new `VEC_NET_MODE=fusion`+branch. Env knobs: `VEC_NET_MODE` (type|global|fusion), `VEC_FUSION_GAMMA`, `VEC_NET_BETA`.+No RNG in the block; deterministic. Active only when n_input_stages ≥ 2.++## Mechanism-off control (verified byte-identical)++- `VEC_NET_MODE=fusion VEC_FUSION_GAMMA=0` on X3: fused = np.where(disagree, Δμ·1.0, Δμ) = Δμ exactly+  → output h5 `X/data`, `indices`, `indptr` **byte-identical** to `VEC_NET_BETA=0` (node 14 path). Compared, equal.+- **Shipped default** (`NET_BETA=0`, no env) on X3: byte-identical to node 14 (verified). On proxy+  (single input, n_input_stages=1): whole two-stage block inactive (0 gene_shift lines), byte-identical+  to node 10/14/17 proxy → proxy10 unchanged. Both views pass `vec-check` (status ok).++## Mechanism evidence (X3, seed 0, A-half)++Gate membership is γ-independent (γ only scales amplitude): 387,420 total (type,gene) entries,+26,030 network-covered (post ZMIN_net gate), **9,728 sign-disagree = 37.4% of covered** → the gate+*bites* (well above the >90%-unchanged failure signal in PLAN risk #1). Mean |Δμ| of suppressed+entries = **0.0153**, small vs overall mean |shift| = 0.043 → the gate preferentially removes+*low-amplitude* genes, which is exactly why de_score (a rank/sign metric) barely moves.++| γ | X3 board | de_score | de_direction | mmd_u | variogram | |---|---|---|---|---|---|-| net typed β=1 | 54.68 | −0.014 | 0.084 | 0.0272 | 0.00120 |-| net typed β=2 | 54.94 | +0.014 | 0.083 | 0.0271 | 0.00120 |-| net typed β=3 | 55.20 | +0.043 | 0.081 | 0.0271 | 0.00119 |-| **net typed β=4 (ships)** | **55.33** | **+0.057** | 0.079 | 0.0270 | 0.00119 |-| net typed β=5 | 55.07 | +0.029 | 0.078 | 0.0270 | 0.00119 |-| net typed β=6 | 54.95 | +0.014 | 0.077 | 0.0270 | 0.00120 |-| net typed β=8 | 54.78 | 0.000 | 0.075 | 0.0270 | 0.00120 |-| net β=4, ZMIN=2.0 | 55.33 | +0.057 | 0.080 | 0.0270 | 0.00119 |-| stack (parent Δμ β=1 then net β=4) | 55.29 | 0.000 | 0.072 | 0.0257 | 0.00120 |-| (parent node14, typed Δμ β=1) | (54.88) | (−0.026) | (0.092) | (0.0251) | (0.00121) |--- **de_score flips positive for the first time in this tree** (+0.057 at β=4, skill >0.5, above-  floor) and is non-monotone in β with a clear peak at 4 — consistent with "right direction,-  wrong amplitude" (PLAN risk #3): over-amplification (β≥6) degrades it again.-- **Stacking with the raw typed Δμ kills the DE signal** (de_score back to 0) even though it keeps-  the mmd_u gain (0.0257): the noisy Δμ shift re-orders genes and drowns the network direction.-  Ships unstacked — a clean test of the PLAN mechanism, at the cost of node-14's mmd_u edge-  (0.0270 vs 0.0251, ~+0.6 pts given back; net de_recovery/direction gain outweighs it: +0.45 A-half).-- de_direction slightly lower than the parent (0.079 vs 0.092) — the network direction is sparser-  (5,208 regulated genes vs ~28k for Δμ), so unregulated genes contribute nothing to the rank-  correlation; the gated-gene ordering is what improves de_score.-- Mechanism evidence per PLAN: 357 TFs × 10 types, 26,030 (type,gene) entries pass the gate and are-  shifted (192,899 nonzero entries touched, mean |shift| 0.135 at β=4); per-type direction differs-  from the type-mean by 0.0022 on average (typed mode is genuinely per-type, not a global vector);-  gene set shifted = regulated genes above the gate, a sparse subset structurally distinct from the-  parent's dense Δμ set (Jaccard ≪1 by construction: only 5,208/28,000 genes are even eligible).+| 0 (=node14 off-ctrl) | 54.877 | −0.0429 | 0.0748 | 0.02589 | 0.001206 |+| 0.3 | 54.852 | −0.0429 | 0.0758 | 0.02597 | 0.001206 |+| 0.5 | 54.831 | −0.0429 | 0.0759 | 0.02602 | 0.001206 |+| 0.7 | 54.819 | −0.0429 | 0.0768 | 0.02607 | 0.001206 |+| 1.0 | 54.898 | −0.0286 | 0.0770 | 0.02615 | 0.001206 | -## Knowledge sources+Four-group decomposition (γ=0 == node14, and shipped config): de_recovery 48.62, direction 53.06,+cell_state 60.01, covariation 57.25 (skill: de_score 0.486, de_dir 0.531, mmd_u 0.600, vario 0.573).++**Fusion fails the PLAN ship criterion** (board ≥55.5 AND mmd_u ≤0.0255 AND de_score ≥−0.010):+1. de_score is *scale-invariant* (scoring brief §1), so γ=0.3/0.5/0.7 leave it at −0.0429 — shrinking+   disagree genes without changing their rank order does nothing. Only γ=1.0 (zeroing them) reorders,+   moving de_score to −0.0286 = **+1 quantization step (1/70≈0.0143)** — exactly the single-seed+   quantization luck the parent ANALYSIS warns must not ship.+2. mmd_u *degrades monotonically* with γ (0.02589→0.02615): removing the disagree mass erodes the+   dense cell-state coverage that gave node 14 its mmd_u edge (PLAN risk #2 realized).+3. Best board (γ=1.0, 54.898) beats node 14 (54.877) by **+0.02**, far below the ~2-pt T1 noise.++Conclusion: the sign gate cannot simultaneously hold mmd_u and lift de_score, because the genes it+suppresses are low-amplitude (no DE-rank effect) yet carry cell-state mass (mmd_u loss). This matches+the parent lesson that the two direction signals are neither stackable nor swappable — and now, per+PLAN, not sign-fusable either. -- CollecTRI signed TF→target network: provided in `prior/tf_regulons/` of each view (general-  regulatory knowledge, not derived from any held-out stage/genotype; allowed per task rules).-- No held-out stage (E10.5/E12.5, 9.5<E≤13.5) or held-out genotype information used; the program-  reads only the view's inputs, `genes.txt`, and `prior/`. No absolute-time, path, filename, or-  manifest-identity branching (only input count and time *differences*).+## Shipped configuration & expected score++Per PLAN step 8 (all γ fail the criterion), ship the **pure node-14 dense typed-Δμ path**+(`NET_BETA=0.0` default → typed-dmu fallback; `VEC_NET_MODE=fusion`/`VEC_NET_BETA=4` kept for+reproducibility only). This reverts node 17's sparse-net swap and recovers node 14's behavior, which+had the best **B-half** X3 in the tree (55.19 vs node 17's 54.78). Expected node score ≈ node 14's+56.39 (proxy10 unchanged 58.80 single-input; X3 = node 14 path), vs parent 56.12 → +0.27.+Decision rests on B-half superiority (change table) + fusion's A-half failure, NOT on the A-half+net-β=4 gain (node 17 A-half 55.33 did not transfer to its B-half 54.78 — the A/B trap).  ## Verified / not verified -- Verified: off control byte-identity; single-input inactivity; both views run + vec-check; β/ZMIN-  screen above (A-half only, seed 0; quantized de_score steps of 1/70 mean ±1 step is luck).-- Not verified: B-half reproduction (+0.45 A-half is below the ~2-pt noise); whether β=4 transfers-  to final (E8.5→E9.5 official, larger cells/type — the rescale-to-Δμ-median makes β units-  comparable, but the optimum could shift); no multi-seed run within this node's budget.-- proxy10 unchanged by construction (single input); not re-scored (quota preserved).+- Verified: γ=0 and shipped default byte-identical to node 14 on X3; single-input block inactive on+  proxy; both views vec-check ok; γ sweep {0,0.3,0.5,0.7,1.0} on X3 A-half (5 vec-score queries); gate+  membership fraction (37.4%) and suppressed |Δμ| (0.0153). CPU-only, deterministic, ~130 s / 6.9 GB.+- Not verified: β interaction sweep at γ=0.5 (skipped — de_score is provably scale-invariant so β only+  rescales amplitude, cannot move the rank-based DE metrics; would not meet the criterion). No second+  seed run (shipped config is byte-identical to node 14, whose B-half is already known). proxy2 (2-input+  official E8.5 + Qiu E9.0) not separately scored here, but the block is view-agnostic and would activate+  identically to node 14's typed-Δμ path.++## Knowledge sources++CollecTRI signed TF→target regulatory network from `<view>/prior/tf_regulons/*.tsv.gz` (general+lineage/TF-regulation prior, not window-specific), read via relative layout only (view-agnostic). No+retained-stage / retained-genotype measurements used. Program reads no manifest identity field, no+absolute stage time (only time ordering / differences), no external dataset — same guarantees as parent.diff --git a/solution/run.py b/solution/run.pyindex 7c22922..5a68104 100644--- a/solution/run.py+++ b/solution/run.py@@ -75,9 +75,16 @@ SHIFT_K = 30.0         # EB shrinkage constant (detection-count based) SHIFT_ZMIN = 1.5       # sparse gate: multiple of median |dmu_shrunk| SHIFT_MAX = 1.0        # per-gene shift cap (log space) -NET_BETA = 4.0         # CollecTRI network-propagation direction shift (replaces raw pseudobulk-                       # dmu as the direction source when >=2 input stages and the network loads).-                       # Per TF t: d_act_t = mean of the (per-type or global) pseudobulk delta over+NET_BETA = 0.0         # CollecTRI network-propagation direction shift. SHIPS OFF (0.0): node 20+                        # screening showed the sign-agreement FUSION of net-direction with dense dmu+                        # (VEC_NET_MODE=fusion) does NOT beat node 14 on X3 A-half (best gamma=1.0+                        # -> 54.898 vs node14 54.877, +0.02 = noise; de_score is scale-invariant so+                        # gamma<1 leaves it unchanged, mmd_u degrades monotonically). Per PLAN step 8,+                        # ship the pure node-14 dense-typed-dmu path (NET_BETA=0 -> typed dmu fallback).+                        # Net-direction (node 17) and fusion remain available via env for reproducibility:+                        #   VEC_NET_BETA=4 VEC_NET_MODE=type    -> node 17 behavior+                        #   VEC_NET_MODE=fusion VEC_FUSION_GAMMA=g -> sign-agreement fusion (g=0 == node14)+                        # Per TF t: d_act_t = mean of the (per-type or global) pseudobulk delta over                        # t's panel targets; per gene g: dir_g = sum_{t->g} sign(t->g)*d_act_t over                        # kept regulators; EB shrink w_g = n_reg(g)/(n_reg(g)+NET_K); rescaled to the                        # parent's |dmu_shrunk| median so NET_BETA is comparable to SHIFT_BETA;@@ -98,6 +105,14 @@ NET_ZMIN = 1.5         # sparse gate: multiple of median |dir| over regulated ge NET_MIN_TF_TARGETS = 10  # TF needs >= this many panel targets to be kept NET_MIN_TF_CELLS = 10    # TF needs detection in >= this many cells in BOTH stages to be kept +NET_MODE = "type"      # type (node 17 net direction) | global | fusion (sign-agreement gate, below)+FUSION_GAMMA = 0.5     # fusion suppression strength: entries where the network direction sign+                       # disagrees with the dense per-type pseudobulk dmu get dmu *= (1-gamma).+                       # gamma=0 -> byte-identical to node 14 (dense dmu unmodified; off control).+                       # Network dir computed exactly as node 17 (typed, K_net, ZMIN_net gate);+                       # only its SIGN is used, amplitude comes from dense dmu with SHIFT_BETA.+                       # Env: VEC_NET_MODE / VEC_FUSION_GAMMA.+ CYCLE = [     "Mki67", "Top2a", "Pcna", "Ccna2", "Ccnb1", "Ccnb2", "Ccnd1", "Ccneg", "Ccne1",     "Cdk1", "Cdk2", "Cdk4", "Cdk6", "Mcm2", "Mcm3", "Mcm4", "Mcm5", "Mcm6", "Mcm7",@@ -568,6 +583,24 @@ def net_direction(view, entries, genes, labels, k_net, zmin_net, typed=True,     return dir_t, diag  +def fusion_gate(dmus_t, dirs_t, gamma):+    """Sign-agreement gate: shrink dense per-type dmu toward 0 where the (post-gate, nonzero)+    network direction sign disagrees. Entries without network info, or with dmu==0, untouched.+    gamma=0 returns dmus_t numerically unchanged (off control -> node 14 path)."""+    has_net = dirs_t != 0+    disagree = has_net & (dmus_t != 0) & (np.sign(dirs_t) != np.sign(dmus_t))+    fused = np.where(disagree, dmus_t * (1.0 - float(gamma)), dmus_t)+    diag = {+        "entries_total": int(dmus_t.size),+        "entries_with_net": int(has_net.sum()),+        "entries_disagree": int(disagree.sum()),+        "frac_disagree_of_net": float(disagree.sum() / max(int(has_net.sum()), 1)),+        "mean_abs_dmu_suppressed": float(np.abs(dmus_t[disagree]).mean()) if disagree.any() else 0.0,+        "gamma": float(gamma),+    }+    return fused, diag++ def apply_gene_shift_typed(Xsel, dmus_t, inv_out, beta, zmin, max_shift):     from scipy import sparse as sp @@ -684,17 +717,28 @@ def main() -> None:     net_beta = float(os.environ.get("VEC_NET_BETA", NET_BETA))   # 0 -> off control (parent dmu path)     net_k = float(os.environ.get("VEC_NET_K", NET_K))     net_zmin = float(os.environ.get("VEC_NET_ZMIN", NET_ZMIN))-    net_typed = os.environ.get("VEC_NET_MODE", "type") != "global"+    net_mode = os.environ.get("VEC_NET_MODE", NET_MODE)          # type | global | fusion+    fusion_gamma = float(os.environ.get("VEC_FUSION_GAMMA", FUSION_GAMMA))+    net_typed = net_mode != "global"     net_stack = os.environ.get("VEC_NET_STACK", "0") == "1"   # 1: parent typed dmu shift THEN net shift     if shift_beta != 0.0 and n_inputs >= 2:         entries = sorted(manifest["inputs"], key=lambda e: float(e["time"]))         dirs_t, ndiag = None, {}-        if net_beta != 0.0:+        if net_beta != 0.0 or net_mode == "fusion":             dirs_t, ndiag = net_direction(args.data, entries, genes, labels, net_k, net_zmin,                                           typed=net_typed)             if dirs_t is None:                 print(f"[gene_shift] net fallback to parent dmu: {ndiag}", file=sys.stderr)-        if net_beta == 0.0 or dirs_t is None or net_stack:+        use_fusion = net_mode == "fusion" and dirs_t is not None+        if use_fusion:+            dmus_t = gene_delta_by_type(args.data, entries, genes, labels, inv, shift_k)+            fused, fdiag = fusion_gate(dmus_t, dirs_t, fusion_gamma)+            Xsel, sdiag = apply_gene_shift_typed(Xsel, fused, inv[idx], shift_beta, shift_zmin, shift_max)+            sdiag.update(fdiag)+            sdiag.update(ndiag)+            sdiag["mode"] = "fusion"+            print(f"[gene_shift] {sdiag}", file=sys.stderr)+        if (not use_fusion) and (net_beta == 0.0 or dirs_t is None or net_stack):             if shift_mode == "type":                 dmus_t = gene_delta_by_type(args.data, entries, genes, labels, inv, shift_k)                 Xsel, sdiag = apply_gene_shift_typed(Xsel, dmus_t, inv[idx], shift_beta, shift_zmin, shift_max)@@ -709,7 +753,7 @@ def main() -> None:                 sdiag["n_dmus_nonzero"] = int((np.abs(dmu_raw) > 1e-9).sum())                 sdiag["mode"] = "global"             print(f"[gene_shift] {sdiag}", file=sys.stderr)-        if dirs_t is not None:+        if dirs_t is not None and not use_fusion:             if not net_typed and dirs_t.shape[0] == 1:                 Xsel, sdiag = apply_gene_shift(Xsel, dirs_t[0], net_beta, 0.0, shift_max)             else:

调研来源?调研员查到并用到的知识条目和文献检索结果（只列标题和编号）。

用到的知识库条目

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

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

只给统计和最后回答的摘录；完整对话请到原始记录位置里列出的文件看。

分析员

ANALYSIS.json written

工程师

Done. Final state verified and clean.

## Summary

**What I did:** Implemented the PLAN's sign-agreement fusion in node 17's `run.py` — a new `fusion_gate()` + `VEC_NET_MODE=fusion` branch that keeps node 14's dense per-type pseudobulk Δμ as the shift base and uses the CollecTRI network direction (exactly as node 17) as a binary sign filter: genes where sign(dir)≠sign(Δμ) get Δμ·(1−γ). Then screened γ on X3 A-half.

**Off-control verified:** γ=0 fusion is byte-identical to node 14 (h5 data/indices/indptr equal). Shipped default (NET_BETA=0) is byte-identical to node 14 on X3; proxy single-input leaves the block inactive. Both views pass `vec-check`.

**Screen result (5 vec-score queries, 15 left):** Fusion **fails** the PLAN ship criterion. Gate bites (37.4% of net-covered entries disagree, 9728 entries) but suppressed genes are low-amplitude (mean |Δμ|=0.0153 vs overall 0.043), so de_score — being scale-invariant — is unchanged at γ=0.3/0.5/0.7 and moves only +1 quantization step (noise) at γ=1.0, while mmd_u degrades monotonically (0.02589→0.02615). Best board 54.898 vs node14 54.877 = +0.02, far below ~2pt noise.

**Shipped:** Per PLAN step 8, the pure node-14 dense typed-Δμ path (`NET_BETA=0.0`), reverting node 17's sparse-net swap. This had the best B-half X3 in the tree (55.19 vs node17's 54.78). Expected node score ≈ 56.39 vs parent 56.12 (+0.27), resting on B-half superiority, not the A-half net gain that didn't transfer for node 17.

**Next best direction:** The three-way evidence (stack→0, swap→mmd_u loss, sign-fuse→no DE gain) says Δμ and network directions can't be combined at the amplitude level. To actually lift de_recovery one must change the *rank order* of top-DE genes without eroding dense coverage — e.g. use network direction only to re-rank within the already-gated high-|Δμ| gene set (magnitude-preserving reordering), or attack mmd_u via the composition lever instead. Both are separate methods, not sign-fusion.

调研员

n steps of 1/70≈0.014. A single seed may show ±1 step by chance. Engineer must confirm with ≥2 seeds if de_score change is <2 steps.\n4) 30-min budget: this is a code modification, not new infra. Network loading and Δμ computation already exist. Risk is low but Engineer should test on X3 subset first.",
  "family_id": "composition_program",
  "mechanism": "Sign-agreement gated fusion: dense per-type Δμ provides the shift base (preserving cell-state coverage for mmd_u), while the CollecTRI network direction acts as a binary sign filter that suppresses genes where the two direction estimates disagree (reducing ordering noise for de_score). Suppression strength γ controls how much disagreement genes are shrunk.",
  "vs_constant_shift": "Not a constant per-type vector: the gate is gene-specific and conditioned on two independent direction estimates (pseudobulk Δμ and network propagation). Genes with conflicting signals are selectively suppressed; genes without network info are untouched. A constant shift would move all genes in a type equally regardless of directional evidence.",
  "mechanism_evidence": "Engineer should report: (1) fraction of genes gated (sign disagreement × has network info) at each γ; (2) X3 mmd_u raw vs node 14 baseline (0.02514) — should stay ≤0.0255; (3) X3 de_score raw vs node 14 (−0.026) — should move toward 0 or positive; (4) de_direction raw change; (5) number of genes actually suppressed (dir magnitude reduced) and their mean |Δμ|; (6) vec-score four-group decomposition at γ=0 (must match node 14) and at shipped γ.",
  "mechanism_off_control": "VEC_NET_MODE=typed (or equivalently VEC_FUSION_GAMMA=0): the sign gate is bypassed, dense Δμ is applied unmodified, and output must be byte-identical to node 14's run (compare h5 data arrays). Additionally, single-input views (proxy10) never activate the two-stage block, so proxy10 output is byte-identical to node 10. If γ=0 output differs from node 14, the implementation has a bug.",
  "sources": []
}

审查员

REVIEW.json written