How victorzhong0110/llm-research-os works
LLM Research OS is a model-, training-backend-, and compute-provider-neutral research control plane. Researchers define and validate research specifications, compose experiments, and may use AI to propose or contest plans; authorization gates sensitive actions before runs execute on simulated, local, or remote Workers. The system records events and artifacts, then supports evidence import, metrics, reports, lineage, and research decisions. The README documents these major capabilities; sampled source confirms the shared application façade, append-only event store, Worker client, and reviewed-execution constraints, but not every subsystem’s wiring.
Main entry points: storage/store.py Append-only event store · artifacts/store.py Content-addressed artifacts · service.py Command service · control.py Run control
Components 23
Research authoring
- Research specification
src/llm_research_os/spec/io.py - Experiment blocks
src/llm_research_os/blocks/registry.py - Research decisions
src/llm_research_os/research/control.py
Execution control
- Command service
src/llm_research_os/application/service.py - Plan and dry run
src/llm_research_os/execution/planner.py - Authorization gate
src/llm_research_os/execution/authorization.py - Reviewed execution
src/llm_research_os/execution/native_reviewed.py - Run control
src/llm_research_os/runs/control.py - Worker control plane
src/llm_research_os/workers/plane.py
Facts and evidence
- Event catalog
src/llm_research_os/events/catalog.py - Append-only event store
src/llm_research_os/storage/store.py - Replay projections
src/llm_research_os/projections/replay.py - Content-addressed artifacts
src/llm_research_os/artifacts/store.py - Evidence import
src/llm_research_os/evidence/control.py
Compute and analysis
- Worker client
src/llm_research_os/workers/client.py - Execution runtimes
src/llm_research_os/execution/kernel.py - Training adapters
src/llm_research_os/training/ms_swift.py - Model providers
src/llm_research_os/providers/control.py - Metrics and reports
src/llm_research_os/report/render.py
Other components
- Researcher
- AI adviser
- Research CLI
src/llm_research_os/cli/parser.py - Remote Worker
Connections 28
- Researcher → Research CLI: submits commands
- AI adviser → Research decisions: proposes or contests
- Research CLI → Command service: dispatches commands
- Research CLI → Research specification: validates documents
- Command service → Append-only event store: records facts
- Command service → Run control: executes run commands
- Command service → Experiment blocks: loads manifests
- Research specification → Plan and dry run: supplies experiment
- Experiment blocks → Plan and dry run: supplies block registry
- Plan and dry run → Authorization gate: plans actions
- Authorization gate → Append-only event store: records decisions
- Reviewed execution → Append-only event store: verifies authorization facts
- Reviewed execution → Content-addressed artifacts: verifies execution material
- Run control → Append-only event store: records run facts
- Worker control plane → Append-only event store: records Worker facts
- Worker control plane → Content-addressed artifacts: serves artifacts
- Worker client → Worker control plane: polls and reports
- Worker client → Content-addressed artifacts: uploads artifacts
- Remote Worker → Worker client: runs client
- Run control → Execution runtimes: dispatches execution
- Execution runtimes → Training adapters: uses training backend
- Append-only event store → Replay projections: replays facts
- Evidence import → Content-addressed artifacts: stores evidence
- Evidence import → Append-only event store: records imports
- Append-only event store → Metrics and reports: supplies run facts
- Content-addressed artifacts → Metrics and reports: supplies result material
- Model providers → AI adviser: enables model calls
- Research decisions → Append-only event store: records decisions