instrument index / research

Worlds that only make sense when they move.

Five real-time systems you can perturb, measure, and run controlled experiments against. No account and no fixed demo.

live model organisms

Open a world

Each simulation runs in the browser and exposes the same sandbox, challenge, and hypothesis-lab structure.

petri-bench

Can your model do science?

A contamination-resistant benchmark for hypothesis isolation, controlled experimentation, inferential statistics, and calibrated claims. The ground truth is generated and verified at task creation, then hidden from the agent.

inspect the benchmark
objective no model judge in the headline score auditable every number traces to an episode artifact procedural method rigor is scored alongside correctness

the hypothesis lab

Measured, not anecdotal.

State a hypothesis, define a control and treatment, run seeded replicates, and inspect effect sizes and corrected significance tests. The lab turns each playground into an experimental instrument rather than a slider toy.

for agents

The same engines over MCP.

Model agents can describe a world and run deterministic experiments without opening a browser.

npx -y petri-labs-mcp open the npm package