SUITE v3.2
Offline evaluation
- Safety recall 99.2% ≥99% · citation precision 97.4% ≥95%
- Unsupported advice 0.2% ≤0.5% · escalation accuracy 98.8% ≥98%
Capability 5.6 · Screen 51
Evaluation & experiments
Offline eval runs, A/B and shadow lanes, MRT variant comparison, reproducible outputs export. 5.6.
State: ab detail.This is a dedicated, reviewable product state at its own route.
EXPERIMENT LANES
Online evidence
- A/B: A 602 · B 602
- Shadow delta 1.8% · MRT decision record linked
- Export reproducible bundle