OUR METHODOLOGY
Numbers with context.
A benchmark result is a piece of evidence. Understanding its source and settings is part of reading the score.
What the leaderboard ranks
The main table ranks published observations from one selected benchmark and source cohort. Its score column is the original reported metric. It is not an overall model rating.
Version and configuration
Benchmark versions, dates, model identifiers and source revisions are retained. The initial OpenEvals feed groups source-reported results; complete prompting, reasoning and sampling settings are not available. These rankings show reported results and do not establish a controlled comparison.
Missing evidence
A dash means no published observation is available. It is never treated as zero. Coverage reflects available evidence, not model quality.
Models and agent systems
Agent benchmarks depend on the model, scaffold and tools. Results lacking a recoverable system identity are not attached to a base model. Their benchmark definitions remain in the explorer while source integration is pending.
Overall index
We do not currently publish a composite Singularity index. A future index would need an explicit benchmark basket, normalization method, minimum coverage, and versioned weights. Preference ratings, cost and latency will stay distinct from accuracy.
History and corrections
Imported observations are retained with checksums and source snapshots. Only new evidence creates a historical observation. Rank movement is not invented from one snapshot. Withdrawing a publication retains the underlying audit trail.