File size: 3,224 Bytes
c5054f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
{
  "protocol_version": "P1D-v3.0-focused-ap1-retrospective",
  "frozen_at": "2026-10-30T09:04:00.409696+00:00",
  "seed": 20261031,
  "candidate_sha256": "6dc7405e9240d0435d2d275f1c0d5a94acebda518ff06a7451620aab8babe24d",
  "candidate_scope": "original18, no reselection; historical CM030 residual identity retained",
  "geometry_population": "complete left frozen fine KCa'b'-ap1 parent148; focal pilot, not whole-KC generalization",
  "pair_split": "reuse P1-v2 unordered pair hash seed20261030; all148*147 directions, all nonedges, no case-control subsampling",
  "features": "contact-free exact skeleton point-to-segment proximity FULL/CA/PED/ML/VL at Gaussian scales1,2,5um; cable midpoint quadrature1um with0.5um check",
  "evaluation": "retrospective not blind; train-only graph features; subtract direct target edge from degree predictors",
  "capacity": "all M0-M4 RFF128 + logistic L2; same C=[.1,1] and gamma=[.5,2] =4 settings, validation selection; equal nominal129 coefficients not guaranteed equal effective capacity",
  "M0": "known fine/coarse type and side (constant within this parent)",
  "M1": "soma, distance, pair-excluded training degree and contact-free tree descriptors",
  "M2": "M1 plus real region arbor opportunities",
  "M3": "M2 plus frozen18 pre/post/both indicators",
  "M4": "M2 plus training graph continuous embedding; not continuous geometry",
  "cell_block": "root-ID hash disjoint60/20/20 cell sets; evaluate only within-set pairs; geometry-only M1/M2 excludes graph degrees/embeddings/labels to avoid frozen candidate selection leakage",
  "weak_edge_conditions": [
    "count",
    "binary",
    "count_ge2",
    "count_ge3",
    "count_ge5",
    "training_contact_cleft_score_lower_quartile_removed",
    "training_contact_connection_score_lower_quartile_removed"
  ],
  "rosette": "unknown unless source annotation found; weak-edge is NOT rosette label",
  "candidate_stability": "20 fresh pair-drop20 seeds2026111000..19; K2/3/4 inside fullparent using recomputed globaltraining embeddings + same frozen soma/tree fusion; best-overlap Jaccard, retention, purity; no stable promotion",
  "matched_null": "1000 groups18 without replacement from other130 in same fine type/side; covariates soma XYZ, logtrain degree and tree PCA8; propensity nearest-neighbor pools20 per target, random ordering and unused-control sampling; report SMD and balance failure; no route-based selection",
  "routing": "ALPN input and MBON output partner types; equalcell Hellinger profiles; matched null contrast of group-vs-complement mean profiles; two endpoints Holm correction; not biologically independent and prior project route exposure remains retrospective",
  "geometry_null": "target geometric dispersion vs same fixed matched sets; no posthoc rematching",
  "crossbrain": "audit original MaleCNS release hashes and endpoint coverage before external mechanism claims; BANC specimen independence literature primary source must be traced",
  "stop_rules": "source graph/count conflict or split leakage => block geometry model claims; numerical quadrature check difference>.02 => refine explicitly; unmatched covariates => no matched-null significance claim",
  "no_new_functional_units": true
}