Automate discovery.

Recursive scientific improvement

Recursive self-improvement.

Research OS turns agents into self-improving scientists: each experiment tests a theory, captures what matters, and makes the next discovery smarter.

Mechanism-first Failure-aware Promotion-gated
Research State
01 Hypothesis
02 Intervention
03 Evidence
04 Next State

Stop chasing scores. Build understanding that compounds.

Each action is chosen for useful gain and information about why the system behaves the way it does.

Theory

Turn ideas into testable claims.

Design

Specify the mechanism and falsifier.

Sandbox

Move fast without mutating truth.

Trace

Capture what actually happened.

Promote

Commit only what survives evidence.

Exploration stays wide. Truth stays protected.

Agents can create, test, and learn aggressively. Promotion is a separate act with stronger evidence and clean provenance.