Before comparing AI models by their scores, it helps to establish which model each score describes and when that model became available. A mismatched name or release date can make results from different periods look directly comparable.
EvalEval’s evaluation cards collect model information and reported evaluation results. I found 9 July 2024 in the release-date field of its Grok 4 card and compared it with xAI’s official announcement and developer release notes. The official date was 9 July 2025: the same day and month, one year later.
Leaving a route someone else could check
On 18 September 2026, I filed public issue #76 with the card URL, the field to inspect, the observed value, and the official sources. The request is conditional on what the field means: if “Released” denotes the model’s public release, the year should be 2025. Specifying the field and the two values gives someone reviewing the issue a concrete check to repeat.
This is a metadata problem—an issue with the information describing what a result refers to. It is closer to checking the title and date attached to a table than recalculating the numbers inside it. The report does not establish that Grok 4’s benchmark scores are wrong or that its evaluations are invalid.
At the 28 September check, the issue remained open with no response. A correction request is the confirmed outcome so far; I have not confirmed that the card has been corrected. This small date check is a record of following a model description back to its source before relying on the associated evaluation results.