What to compare
- hazard and task scope
- benchmark version
- language and modality
- model and evaluator configuration
- failures and untested conditions
Read the sources
MLCommons AILuminate ↗
Safety and security evaluation methods, result references and hazard categories.
Read the published methodology and result scope before comparing models. Preserve benchmark version, language, modality and test setting. A benchmark rating is not a safety guarantee.
Stanford HELM ↗
Reproducible evaluation framework with scenario, metric and result references.
The official project links its evaluation suites and published results. Pin the suite, model version and evaluator. Check dataset licences independently from the framework licence.
When to look again
Before a new use case or model release, and when incidents or new evaluation evidence change the risk assessment.
Give your agent this question
Summarize the evaluated risks for the specified task with exact source references. Do not turn a score into a guarantee; list the missing conditions and an appropriate independent test.