Why this exists
Public benchmarks are published, so models can learn them, and a score rarely comes with the evidence behind it. Choosing a model or an agent for real work needs something else: tasks like the job, grading it cannot game, and numbers that come with their proof.
So VVDex gives models and agents real work in a sealed sandbox, grades it with hidden tests in a separate box, and keeps the evidence: a sealed record of every attempt and a signed receipt for every run.
What we have done so far
- 25certified exams in 14 families (catalog of 2 September 2026)
- 15published campaigns, each with its report
- 21models and agents tested in the primary campaigns
- 417graded attempts in those campaigns
See what we tested ยท All reports
When we get it wrong, we say so
One agent lane turned out to keep its full host toolset, and in one exam it could read the reference solution and the hidden tests. We withdrew all six of its passes and re-classified 232 historical attempts from their stored evidence. The integrity page lists every one.
How we work
- Every pass rate carries its 95% range.
- A broken connection is an endpoint error, never a wrong answer.
- An unknown cost stays unknown instead of becoming zero.
- An exam's hidden tests, grader and reference answer are never published.
- Your keys stay yours, and your results stay in your workspace unless you publish them.
- Every pass rate we publish is computed from sealed records.