Visual Artifact Bench · v2
Does the generated interface actually work?
VAB asks coding agents to build interfaces, data visualizations, simulations, games and graphics from a product brief, then operates the artifacts like a user. Executable checks, reference solutions and seeded faulty variants make the judgement inspectable.
Task roster
77
One task roster
Results & generation progress
Loading current data…
Every model configuration below uses the same 77 cases. An artifact can be available before its run completes or its checks are judged. Full-roster scores stay pending until that evidence is available.
Loading model configurations…
Reading the coverage
- Generated
- An artifact was found for the case.
- Qualified
- The artifact meets generation validity requirements.
- Native completed
- The agent ended normally. Time and step limits stay separate.
- Judged / audited
- Checks have evidence / failed checks have an artifact review.
Generation coverage describes execution progress. A per-case check result and a full-roster score require their own judgement evidence. Model and harness configurations are kept distinct.