How pkjaslam/cambium_ai works
Cambium is a local-first, human-governed AI research institute: a task is routed to specialist work, paused at human decision gates, independently checked, and carried into reporting and close-out. The graph includes its documented council stages and major capabilities, plus the sampled router, web bridge, and MCP entry point. The README documents broader council and learning behavior than the sampled source verifies, so those links reflect documented workflow rather than inspected implementation details.
Main entry points: cambium_run.py Run orchestrator · server.py MCP server · task_router.py Task router · engine.py Streamed run engine
Components 21
Research workflow
- Run orchestrator
tools/cambium_run.py - Task router
tools/task_router.py - Scouting
tools/cambium_run.py - Lab design
tools/cambium_run.py - Execution
tools/cambium_run.py - Verification board
tools/cambium_run.py - Reporting
tools/cambium_run.py
Governance and records
- Human gates
tools/gate.py - Enforcement checks
tools/enforce.py - Run records
tools/run_state.py - Audit ledger
tools/audit_log.py - Learning delivery
tools/learning_delivery.py
User interfaces
- 3D web interface
web/frontend-r3f/src/App.jsx - Bridge API
web/server/app.py - Streamed run engine
web/server/engine.py - MCP server
mcp_server/cambium_mcp/server.py
Research support
- Proposal support
tools/budget_review.py - AI-use disclosure
tools/ai_disclosure.py - Finding recall
tools/memory_recall.py - Concept graph
tools/concept_graph.py
Other components
- Researcher
Connections 25
- Researcher → Run orchestrator: submits task
- Run orchestrator → Task router: requests plan
- Task router → Run orchestrator: returns phases
- Run orchestrator → Scouting: dispatches work
- Run orchestrator → Lab design: dispatches work
- Run orchestrator → Execution: dispatches work
- Run orchestrator → Verification board: dispatches work
- Run orchestrator → Reporting: dispatches work
- Run orchestrator → Human gates: pauses at gates
- Researcher → Human gates: decides
- Run orchestrator → Run records: records run
- Human gates → Audit ledger: logs decision
- Run orchestrator → Enforcement checks: checks controls
- Run orchestrator → Learning delivery: delivers packet
- Researcher → 3D web interface: uses
- 3D web interface → Bridge API: sends requests
- Bridge API → Streamed run engine: creates run
- Streamed run engine → Task router: routes task
- Streamed run engine → Researcher: streams gate events
- Researcher → MCP server: invokes tools
- MCP server → Task router: plans task
- MCP server → Proposal support: runs review
- MCP server → AI-use disclosure: runs tool
- MCP server → Finding recall: queries findings
- MCP server → Concept graph: queries graph