[janhelcl]

Summary

• add a shared per-pressure-level MLP encoder for the canonical 5 × 13 XC weather profile
• apply the same [16, 8] level encoder at every pressure surface and forecast time
• preserve 1000 → 500 hPa order through flattening and a 32d profile projection
• add an isolated seed-42 config against the frozen conventional MLP
• record the completed result in ADR 0017 and close the experiment

Controlled experiment boundary

The candidate added only the profile branch:

standardized [u, v, T, RH, z] × 13 levels
  -> shared MLP 5 -> 16 -> 8 per level
  -> ordered flatten (13 × 8)
  -> Linear(104 -> 32)

The raw bypass, shared per-time MLP [64, 32], fusion [64, 32], 32d site embedding, independent sigmoid head, dropout 0.10, optimizer policy, and xc-temporal-2024-jan-nov-v1 benchmark remained unchanged.

Result

Metric	Conventional MLP	Shared level MLP	Delta
Macro BCE ↓	0.15686	0.15794	+0.00108
Macro Brier ↓	0.04762	0.04785	+0.00023
Macro ROC-AUC ↑	0.94206	0.94233	+0.00027
Mono violation ↓	0.00140	0.00209	+0.00069
Best epoch	86	86	—
Parameters	48,523	58,259	+9,736

Decision

Reject at seed 42. The candidate regresses both primary calibration losses, worsens monotonicity, and adds about 20% more parameters. The isolated +0.00027 ROC-AUC gain does not earn paired-seed confirmation.

Do not run seeds 43–46, tune the level MLP, add AGL/masking, or proceed to level attention on this encoder. Preserve it as a reproducible negative experiment; the frozen conventional MLP remains the promotion baseline.

Validation

• focused weather-profile tests: 10 passed
• full ML suite: 103 passed
• git diff --check: clean