Behavior
Controls produce continuous, visible change.
Interactive agent evaluation
Can AI build simulations that teach?
121 simulations test behavior, scientific fidelity, and classroom usefulness.
Live benchmark surface
Energy conservation in motionOpen lab ->
121Interactive simulations
242Benchmark variants
4Benchmark models
0Verified outputs
Generation coverage
Live results from published, browser-verified candidates.
| Model | Verified | Coverage | Average check score | Median time | Status |
|---|
Coverage is measured against 242 tasks per model. Only candidates passing all automated browser checks are published.
4
Signals separate working science from a convincing animation.
Science
Relationships remain physically defensible.
Teaching
Evidence supports a classroom inquiry.
Robustness
Repeated use and reset states hold up.
Benchmark task catalog
Search every simulation, inspect its tasks, and run generated model outputs.