[janhelcl]

Summary

• add a resumable hosted Jev 1.13 benchmark for the canonical XC contract
• use a fixed semantic weather adapter and fit-only historical priors
• evaluate all 82,584 rows in xc-temporal-2024-jan-nov-v1
• document the completed result and reject Jev as the main XC model in ADR 0016

Result

Model	BCE ↓	Brier ↓	ROC-AUC ↑	Mono rate ↓
Conventional MLP	0.15686	0.04762	0.94206	0.00138
Jev 1.13	0.31116	0.08475	0.84939	0.17268

Jev shows useful zero-shot semantic signal, especially at higher thresholds, but the predictive and monotonicity gaps are decisive. The experiment line is closed without tuning against the 2024 holdout.

Validation

• full ML test suite: 98 passed
• Jev unit coverage includes non-leaky state/questions and resumable partial-to-full execution
• implementation tree previously verified against the locally tested commit