Benchmark dossiers / independent analysis

The benchmark behind the number.

Inspect what each benchmark asks, how its score is produced, and which conclusions its design can support. These are original analyses by Arcophos of the benchmark authors’ work. Any reproduced measurements are labeled with their published source and version.

Dossier01

Original Nature study, corrected version of record

MultiMedQA ↗

Seven question sources do not produce one clinical score.

UnitOne question; human evaluation uses a selected 140-question subset.MeasureTask-specific accuracy and human ratings

What we contribute. Source reconciliation, task and metric interpretation, and tools that make assumptions inspectable. We do not claim to have created these benchmarks or run the reported models.