单 Agent 运行

SELECTION — run 20261003-155754-t3-gata4-komix-b

Board in this run's harness header: T3:gata4 only.

T3:gata4 — chosen submission

final/T3__gata4.h5ad

python scripts/t3_komix.py \
  --wt data/raw/official/t3/WT_E8.75.h5ad \
  --ko data/raw/official/t3/E9.5_mab21l2_ko.h5ad \
  --out final/T3__gata4.h5ad \
  --beta 0.40 --k 0.55 --seed 0 --max-cells 7449 --zero-gene Gata4
python scripts/verify_submission.py --input final/T3__gata4.h5ad --board T3:gata4   # passed

Method: T3 method card v2 main line komix — real cells of the observed training knockout (Mab21l2 KO E9.5, official training data) mixed into the stage-matched WT carrier (WT E8.75), each cell keeping its own expression and its own spatial_3D; no averaging, no shift transfer, no per-gene editing, no densification. Plus the card's optional --zero-gene component for the board's knocked-out gene.

Disclosure required by the method card: the output contains real cells of the training knockout sample (Mab21l2 KO E9.5, official training data); they are 45.0 % of the submitted cells (uns["t3_komix"]: n_ko 3 352 / n 7 449, ko_fraction 0.4500), and their coordinates come from the same file.

Why this variant

1. The card's main line at this run's start point. The harness gave variant b the start β=0.40, k=0.55; the a/b/c variants differ only there, and the comparison across runs is by official quota, so the start point is kept unchanged (see 3: the knobs are second-order).
2. --zero-gene Gata4 is on. It is the only component whose sign is verified by allowed training data: in the official Mab21l2 KO file the targeted gene's transcript is essentially gone (pseudobulk 0.0081 and 0.19 % of cells nonzero, versus 0.976 / 18.7 % in WT E9.5), so a knockout of Gata4 should make Gata4 the board's largest down-response. Setting it to 0 in every output cell only turns nonzeros into zeros (never densifies; density ratio 0.940). In the hypothesis scoring below it is worth +0.2 to +0.6 response-group points and is never negative, under every hypothesis about the unseen truth that I could build from allowed files.
3. β and k are second-order. dp = (1−k)·d_mix + k·ε with std(ε) ≈ 0.023 versus std(d_mix) ≈ 0.427, and the two rank-based response metrics are invariant to the scale of dp: over β∈{0.40,0.55,0.80,1.00} × k∈{0.45,0.55,0.65} the response groups move by ≤1.5 points (all of it through severity_slope) and the 20-point cell-state group by ≤1 point (analysis/t3_grid.py, analysis/t3_cellstate.py, analysis/LOG.md). No local ruler score was used to pick them (method card v2: lgo is documented as rank-inverted against the official board; lgo875's target is komix's own carrier difference, so scoring komix with it is circular — record only, see below).

Honest statement of the residual risk

d_mix = pb(KO E9.5) − pb(WT E8.75) = d_dev + d_ko is dominated by the E8.75→E9.5 progression (||d_dev|| = 9.84 vs ||d_ko|| = 5.64; corr(d_mix,d_dev) = +0.83, corr(d_mix,d_ko) = +0.28, and the two components themselves correlate −0.30). If the Gata4 truth were a pure developmental arrest (dt ≈ −d_dev), komix would land ~7 points below the floor on the response groups; if it is shaped like the observed knockout (dt ≈ d_ko), ~14 points above it (analysis/LOG.md §2). This cannot be settled locally.

Two things decide the bet in favour of keeping the komix line:

• k is not a hedge. Because de_score and de_direction are rank-based and std(ε) is 18× smaller than std(d_mix), the direction of dp — and therefore the whole downside under H_delay — is essentially identical for every k ≤ 0.75 (32.8 / 32.8 / 32.9 response points at k = 0.45 / 0.55 / 0.65). Raising k does not buy protection; it only shrinks severity_slope towards its floor and, in the limit k = 1, becomes a WT copy, whose dp is pure sampling / batch noise (the case the card flags as "照抄 WT 不是 50 分").
• The curators' official-score calibration behind card v2 ranks programs by exactly this carrier direction. scripts/ruler_calibration_t2t3.py (in this workspace) documents three official T3:gata4 points for pre-komix programs — 45.9 / 53.4 / 41.0 — with the most conservative one at 41.0, and states that the lgo875 ruler ("rewards moving from WT E8.75 towards the knockout sample, penalises developmental delay") reproduced their ordering and sign, while lgo was inverted. That is the evidence card v2 rests on; it is three points, it is not independent of the scoring service's view, and I could not reproduce the calibration here (it re-runs programs from other runs' workspaces), so it is cited as corroboration, not proof.

Local ruler (record only, not a selection criterion, not an estimate of the official score)

Pipeline check (task book §5.1): scripts/score_proxy.py --board T3:gata4 --baselines reproduces the local ruler's floor / ceiling — wt_identity 50.00, split_half 100.00 — and the submitted file is reproducible from the command above (re-ran it and compared X, spatial_3D and the labels: identical).

final/proxy/T3__gata4.h5ad is the identical method and parameters, generated with --ko-exclude-half B so that the cells the local ruler uses as truth (the second split_half(seed=0) half of the Mab21l2 KO file) are not in the file — without that flag the recording would be circular. Scored with scripts/score_proxy.py --board T3:gata4 (T3:gata4/proxy_mab21l2, reference t2_heart/E9.5, truth = Mab21l2 KO half B, 25 147 cells):

proxy_score T3:gata4 69.71

Groups: response_genes 62.0, direction 69.2, magnitude 96.1, cell_state 48.8 (modeling/experiments/logs/proxy/T3:gata4__komix_b040_k055_zgata4_exclB.json; ledger rows in notes/changelogs/experiments.tsv).

This is a local proxy ruler on a different gene and a different stage, and it is partly circular even with half B excluded: the knockout cells in the prediction come from half A of the very sample whose half B is the ruler's truth (same embryo, same slide, same batch), which is why magnitude reads 96.1. The method card also documents this ruler family (lgo/lgo10) as rank-inverted against the official board. Per the card it is recorded here for the harness's re-scoring check only; the submission above was chosen on mechanism, not on this number.

External data disclosure (外部数据披露)

无 / none. Only the official training files under data/raw/official/t3/ (WT_E8.75.h5ad, WT_E9.5.h5ad, E9.5_mab21l2_ko.h5ad) and data/reference/panels/ were read. No network access, no external dataset, no pretrained model. No held-out genotype (Gata4 or β-catenin knockout, any allele, any phenotypically similar perturbation) and no forbidden-window data was read or used.

Files