[janhelcl]

Summary

• generalize glideator-ml confirm-seeds xc to accept multiple controls while training the candidate only once per seed
• preserve the existing single-control command/report shape
• add tests for multi-control pairing, candidate reuse and legacy behavior
• record the paired 42–46 AGL/mask result
• add ADR 0015 and close the current vertical Conv1D profile family

Confirmation result

The AGL/mask candidate does not clear the frozen conventional MLP promotion bar:

Metric	MLP mean	AGL/mask mean	Delta	AGL wins
BCE ↓	0.156677	0.156987	+0.000310	2/5
Brier ↓	0.047601	0.047635	+0.000034	3/5
ROC-AUC ↑	0.941865	0.941867	+0.000002	3/5
Mono ↓	0.002076	0.005656	+0.003580	2/5

Against the plain CNN, AGL/mask is also slightly worse on BCE/Brier; only ROC-AUC shows a small signal (+0.000191, 4/5 wins).

Decision

Reject AGL/mask as a promotion candidate and stop the current Conv1D family. Do not tune convolution width/depth/kernel size or add another coordinate channel to this architecture. The frozen conventional MLP remains the promotion reference.

Keep both CNN configs as reproducible negative experiments. Preserve AGL/masking as possible inputs for a different encoder family.

The next vertical-profile hypothesis should use a shared per-pressure-level encoder with simple ordered aggregation; level attention comes later as a separate experiment.

Harness improvement

A two-control five-seed confirmation now takes 15 trainings rather than 20: 5 candidate + 5 per control. The same candidate run for each seed is paired with both controls, preserving the statistical comparison while avoiding duplicate training.