Active
ResearchAgentLab: Control Boundaries for Autonomous Research Agents
Back to Projects
0 / 18
Drifted protocols that passed
18 / 18
Apparatus mutations blocked
15 / 18
Escaping without integrity check
13.0% → 27.1%
Target selection, benign to adversarial
An autonomous research agent, and a control architecture built for it after it silently ran a different experiment from the one that had been approved. Protocol Before Code freezes approved scientific intent in a machine-checkable protocol, binds trusted components by content hash, limits how much the model is allowed to generate, and checks scientific conformance independently of software correctness. Stress-tested with generator-blind holdouts and fault injection.