Affine fixes NAVWORLD scoring instability with parallel judges

NAVWORLD scoring exhibited extreme variance—identical outputs received different scores (30% to 70% variance) due to single LLM judge bias. The team deployed a fix invoking all judges in parallel and computing median scores per dimension across 3 LLMs (or 2 when unavailable), eliminating single-judge bias. Terminal module completed shadow-run testing and fixes; scoring integration expected Tuesday or Wednesday.

• •LLM judge inconsistency caused >40% score swings on same task
• •New parallel judge median approach ensures reproducible results
• •Terminal module ready for scoring rollout after bug fixes