2.5 KiB
H1 comparison design v0.1 addendum (Claude, 2026-11-13)
Amends v0 (Drive 1b1175XqWBbSW3_tXSTdjQfzfS9n467Wj) after Frank's review. Still a design draft; nothing implemented.
A. Partition matching must be verified, not assumed (replaces "match by construction")
Changing edge weights in the same balanced partitioner does not guarantee matched shape or connectivity. Balanced partitioners bound size imbalance but may return disconnected parts, and the edge weights can shift the aspect distribution.
Acceptance per part and per arm (G1/G2/G3):
- Same group count K.
- Every group connected on the true tet adjacency (BFS check).
- Volume-weighted group sizes within the preset imbalance tolerance.
Acceptance over the train/validation pool (distributions, G2 vs G1 and G2 vs G3):
- volume-weighted group size;
- group aspect (eigenvalue ratios of the volume-weighted coordinate covariance);
- boundary-surface fraction per group;
- alignment score with type boundaries (expected to differ; this is the manipulated variable).
Tolerances and test statistic are preset before partitions are generated.
Repairs (splitting disconnected components, rebalancing) are logged, and repair rates are reported per arm, because unequal repair rates are themselves a confound.
A part failing acceptance after repair is excluded from ALL arms, and the exclusion is reported.
B. Whole-part arm definition
- G0 = the same aggregation layer with the trivial partition K = 1. The "group mean" is the volume-weighted whole-part mean broadcast to all nodes. Parameters are identical to G1-G3.
- G0 still has the 64-slice learned global attention. It is the strong unpartitioned reference, not an unstructured model.
- Removing the aggregation layer is a different architecture (different parameter count) and is not an H1 arm.
- Note: G0's K = 1 mean overlaps the global slices, so G0 vs G1 tests "partition pooling vs none" on top of global slices. The type question is G2 vs G1 and G2 vs G3.
C. Uncertainty
- A bootstrap over test source groups captures test-sample uncertainty only, conditional on one training run. It does not estimate training-seed variance.
- Single-seed results are reported as "conditional on this training run" and stay exploratory.
- Any claim needs ≥3 seeds per arm. The paired difference is computed per seed on the same test parts, and the CI uses a two-level resampling: seeds (outer) and test source groups (inner). With few seeds, also report the seed-to-seed spread explicitly.
- The 10 % margin remains a candidate until fixed before the first claim-bearing run.