Current V1
Current V1: clean-v1 reruns and updated evaluation sets, currently tested on fewer frontier models. The overall view allows at most two missing capability benchmarks and weights observed scores by configured evaluation counts. BioSecBench-Refusal never contributes to Capabilities; it is reported separately under Safeguards.
42 model-and-harness rows · data updated 2026-11-05
Historical V0
Legacy V0 with broader historical model coverage. Available benchmark scores are equally weighted; missing scores are excluded rather than treated as zero.
45 model-and-harness rows · data updated 2026-10-09
Sources and benchmark pages
- Benchmark directory — canonical benchmark pages, Markdown results and paper links.
- V1 methodology and provenance — scoring rules and complete source/import notes.
- Version manifest — structured dates, sources and download URLs.
URL migration
Paths now match displayed versions: legacy V0 moved from /results/v1/ to /results/v0/, and current V1 moved from /results/v2/ to /results/v1/. Old versioned URLs are not maintained or redirected; update saved links and integrations. See the explicit version labels above.