Python · open source
Architecture prototype of a continually learning autonomous research system.
How could an autonomous research system keep learning without hallucinating results or acting unsafely? CONTINUUM sketches the architecture and runs it against a simulated objective.
View on GitHubRead the READMEWhat it does
Four-layer memory
SQLite backend with two-buffer consolidation.
Bayesian world model
Gaussian process with uncertainty estimates and next-experiment proposals.
Verification layer
Lets no unsupported claim through.
Governance gate
Hazardous-material screening plus an audit log.
Multi-agent pipeline
Hypothesis generation, runnable with a mock LLM — no API key, GPU or hardware.
Honest limits
Read this first. Phase 0: every experiment runs against a simulated objective function. Continual weight learning and real lab hardware are later, documented-interface phases. Not a validated scientific result.