节点 n6

GENESPACE+α=0.5 → 49.06, beats zero control (48.32). Tuning α and κ.

lowMatcher); integrate from E9.5 with damped Euler steps to predict E10.5.",
  "vs_constant_shift": "Velocity is a function of cell state x (and time t): two cells of the same type at different PCA positions receive different displacement vectors. Node 4 applies one EB-shrunk mean-difference vector per cell type to all cells of that type. Here, within-type velocity dispersion should be substantial and correlated with cell position, not zero.",
  "mechanism_evidence": "Engineer should report: (a) within-type standard deviation of the predicted displacement vectors (norm of v(x,1)*alpha) averaged over types; compare to between-type mean displacement. If within-type std < 10% of between-type, the field collapsed to per-type constants. (b) Pairwise cosine similarity of displacements within a type (should be < 0.9 if truly state-dependent). (c) Four-group score decomposition vs copy_last and vs node 4: specifically whether covariation improves relative to node 4 while cell_state remains competitive. (d) PCA variance ratio output/input per dimension to detect universal diffusion.",
  "mechanism_off_control": "Run the same script with --velocity_mode zero: skip training and integration, output = E9.5 cells decoded back (i.e., copy_last equivalent). Expected: score ≈ node 1 (47-49). Additionally --velocity_mode random: replace learned v with Gaussian noise matched to the trained field's output variance; expected: worse than learned field on direction/DE if the field learned meaningful structure, similar on covariation. If learned ≈ random, mechanism did not engage.",
  "sources": ["arXiv:2302.00482 — Tong et al., OT-CFM: ExactOptimalTransportConditionalFlowMatcher API, minibatch-OT pairing, conditional flow regression (used for torchcfm API and training recipe)", "arXiv:2210.02747 — Lipman et al., Flow Matching: straight-line conditional paths x_t=(1-t)x0+t*x1, target u=x1-x0, time-conditioned vector field (used for fallback manual implementation and conceptual framing)"]
}