All benchmarks contribute to this ranking. General columns show the Diverse and Contextual subset already included in the task scores, with no extra weight. This subset does not establish unseen-task generalization.

How a model ranks against other models (0–100; higher is better). 1. Rank models by their adjusted score on each benchmark. 2. Award 100 points to first place and 0 to last, evenly spaced in between. Ties share the average points. 3. Average the points across all benchmarks, with equal weight per benchmark. Tasks with more benchmarks have more weight. General benchmarks count once. Only rank matters, not the size of score gaps. Adding or removing models can change this score. Requires at least two measured models, each with complete results and valid adjusted scores for every benchmark.

Task Avg and task columns are baseline-adjusted scores, not accuracy. 0 means at or below the constant baseline; 100 means the reference ceiling.

Choice: (target mass − baseline) / (1 − baseline). Noul: 2 × balanced accuracy − 1. Score: 1 − MAE / constant baseline MAE.

Clip each benchmark to 0–100, average benchmarks within each task, then average the three tasks equally. Complete measured coverage is required. Baselines are fitted to evaluation targets.