/ the lab · delegation experiment registry
Experiments enter the Lab when they test a delegation boundary.
The Lab publishes only artifacts that can be shared: public repos, synthetic or redacted corpora, approved traces, frozen protocols, and reproducible results. Selected Free Experiments may enter this registry when they help answer whether an agent should proceed, ask, escalate, or stop. Non-shareable Zartis, client, or project work stays out.
Describe a task for an AI agent. I'll generate a behavioral contract clause for it — the specification that helps decide when delegated work may proceed, ask, escalate, or stop before output evaluation misses the failure. Use public, synthetic, or generalized examples only; do not paste confidential tasks, traces, customer data, credentials, or client material.
Schema: experiment-card-v0.md. Registry data: experiments.json. Nulls and overturned results are not hidden.
Have a hypothesis?
Submit a free experiment idea.
Start with the agent interview. It helps turn a real agent behavior problem into a falsifiable, public-safe experiment. Selected ideas may become free Lab experiments or Behavior Watch case files.
Talk to the lab →Featured · interactive demo
OPBR canonical · refund commit
The touchable face of EXP-001 and EXP-004. Same correct refund, two different paths. Click through the trace pair, detector verdicts, and evidence cards live on the page. Output evaluation sees two successful runs; the behavior contract sees one regression.
9 detectors · McNemar paired tests · trace diff · open the demo →