[janhelcl]

Summary

• add a second, frozen Jev 1.13 context experiment
• preserve all 77 canonical GFS values at 09:00, 12:00 and 15:00
• add the production site name and every production takeoff with its stored wind direction
• snapshot production metadata from public.sites and public.spots
• fingerprint the 248-site / 535-takeoff snapshot into the resumable cache contract
• retain the same pre-2023 priors, 11 Noul targets, held-out evaluator and model version
• document the completed benchmark and close the Jev 1.13 experiment line

Result

Same 82,584-row xc-temporal-2024-jan-nov-v1 benchmark:

Model	BCE ↓	Brier ↓	ROC-AUC ↑	Mono violation ↓
Conventional MLP (seed 42)	0.15686	0.04762	0.94206	0.00138
Jev semantic context	0.31116	0.08475	0.84939	0.17268
Jev raw weather + production sites	0.30780	0.08240	0.82614	0.110

The raw context modestly improved BCE, Brier and monotonicity over semantic context, but reduced ROC-AUC and remained decisively behind the conventional MLP. Approximate API cost: $32.

Decision: reject this variant, close the Jev 1.13 experiment line, and return to the planned shared per-pressure-level encoder with ordered aggregation.

Experimental boundary

This was one pre-specified context ablation. Because the first 2024 Jev result was already inspected, the repeated 2024 comparison is exploratory and is not fresh promotion evidence.

Validation

• five targeted Jev tests passed locally
• original semantic context and resumability remain covered
• raw context test asserts all 77 features at all three forecast times
• production snapshot validates 248 sites, 535 takeoffs and complete wind-direction metadata
• repository CI passed on the implementation commit