[janhelcl]

Summary

• add the site-relative AGL + above-ground mask representation on top of the frozen vertical CNN experiment
• keep the CNN, raw bypass, shared [64,32] MLP, fusion tower and optimizer policy unchanged
• derive z_AGL = geopotential_height - site_altitude from raw values
• fit AGL scaling only on valid above-terrain training/noon levels
• mask pressure surfaces below GFS terrain and append an explicit validity channel

Result

Seed 42 looked clearly positive against both the plain CNN and frozen conventional MLP, but paired seeds 42–46 did not confirm the improvement.

Against the conventional MLP, AGL/mask averaged:

• BCE +0.000310 worse, wins 2/5
• Brier +0.000034 worse, wins 3/5
• ROC-AUC +0.000002, effectively tied, wins 3/5
• monotonic violation +0.003580 worse, wins 2/5

Against the plain CNN, AGL/mask averaged:

• BCE +0.000245 worse, wins 2/5
• Brier +0.000024 worse, wins 2/5
• ROC-AUC +0.000191 better, wins 4/5
• monotonic violation -0.001227, wins 2/5

Decision

Do not promote AGL/mask. The small ROC-AUC signal versus the plain CNN does not improve the primary losses and does not clear the conventional promotion bar.

Stop the current Conv1D profile family rather than tuning width, depth, kernel size or adding another coordinate channel. Keep the implementation/config as a reproducible negative experiment; the next weather-specific architecture should change the inductive bias to a shared per-pressure-level encoder.

The multi-control confirmation workflow and durable decision are recorded in stacked PR #124 / ADR 0015.