Benchmark Radar, an Evidence Trail for AI Benchmarks

Field notes

Benchmark Radar counts model-card adoption without allowing a long appendix to outweigh a different vendor: each card contributes once per benchmark, while organization counts break ties. Its score history uses values readable in cited documents and dates them to publication, leaving visible gaps when no newer figure can be verified. The site explicitly warns that its paths are reported-record histories, not like-for-like scientific trends. A future comparison of score, cost, and speed is described but not yet represented in the current data.