Benchmark Radar catalogs 12,916 AI scores

Benchmark Radar’s September 10 paper catalogs 1,283 benchmark source records and 12,916 numeric score observations across 790 records, alongside source links, score histories, and a searchable dashboard, according to the arXiv report. Its authors warn that only 82 scored records meet their known percentage-scale rule, so the database is an audit trail—not a universal leaderboard. Our BenchMIRT analysis makes the same buyer point from question-level evidence: evaluation teams should pin model version, task, scale, and date before comparing a number.