Skip to content
D-CSIL

Project File

Autonomous R&D Lab

A governed research loop: objective → plan → agents → experiments → evidence → critique → revision, with claim gates and immutable audit logs.

PrototypeClaims must survive the Skeptic to become results.

Overview

OBSERVED

The lab engine turns human objectives into governed agent loops: nine specialist roles with scoped permissions, deterministic task planning, sandboxed experiments, tool budgets, and evidence-backed memory consolidation. Every claim passes Skeptic, Statistical Evaluator, and Auditor review gates before it counts.

Honest Status

EXPERIMENTAL

The MVP is feature-complete and self-validates with a toy experiment playbook (dcsil demo), with a React mission-control dashboard and CLI. The default provider is still a mock: live-LLM research quality is the open frontier, and container-grade isolation is not done.

The Governed Research Loop

OBJECTIVEPLANAGENTSEXPERIMENTEVIDENCECRITIQUECLAIM GATES:SKEPTIC · STATS · AUDITOR

Every stage writes to an immutable audit log; claims must pass review gates to become results.

Current Limitations

  • −No end-to-end autonomous discovery on a real research question yet.
  • −Mock provider by default; live providers are wired but lightly exercised.

Next Steps

  • +First supervised closed-loop run on a real (small) lab question.
  • +Use Remy as the execution arm for experiment stages.