Lumina Reasoning Index

Each benchmark contributes exactly one third. Converted source scores retain full precision until the final combined score is rounded. Matches require the same model release and explicit setting.

57 exact 3/3 configurations; 509 estimates in total: 12 supported and 497 lower-confidence provisional scores. All evidence tiers are enabled by default and ranked together by full-precision score. Display can exclude any tier; unknown settings remain unknown.

Validated same-release estimates are retained. New general three-source predictors did not beat the previous approach in provider-held-out testing and are not adopted. Weak proxies remain provisional and explicitly labelled. Hover information separates publisher standard errors from empirical prediction error ranges; neither establishes decisive order for small score gaps. Missing resource measurements are never estimated. Calibration: reasoning-estimation-v5.

Reasoning contributes to Expanded Capabilities. Method: reasoning-index-v3.