OpenEinstein

OpenEinstein is a local-first AI research platform for reproducible theoretical-physics campaigns.

The core platform is domain-agnostic and safety-enforced. Domain specifics live in Campaign Packs under campaign-packs/.

Quickstart

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

pytest --tb=short -q
ruff check src/ tests/
mypy src/openeinstein/ --ignore-missing-imports

openeinstein --help
openeinstein config --validate --path configs/openeinstein.example.yaml
openeinstein eval list
openeinstein pack list

Core Architecture Rules

• Route model selection through logical roles: reasoning, generation, fast, embeddings.
• Route tool execution through ToolBus abstractions (no direct MCP/subprocess calls in agent/campaign logic).
• Enforce machine policy through configs/POLICY.json at gateway boundaries.
• Keep physics-subfield logic out of core modules; put it in Campaign Packs.

First Campaign Pack

• Pack path: campaign-packs/modified-gravity-action-search/
• Includes: campaign config, skills, templates, known-model eval fixture, literature seed, and docs.
• Validate and dry-run:

openeinstein config --validate --path campaign-packs/modified-gravity-action-search/campaign.yaml
pytest tests/integration/test_modified_gravity_pack.py --tb=short -q

Documentation

• Architecture overview: docs/ARCHITECTURE.md
• Canonical subsystem docs: docs/canonical/
• Trust model: docs/trust-model.md
• Configuration reference: docs/configuration-reference.md
• Campaign Pack authoring guide: docs/campaign-pack-authoring.md

Repository Layout

• src/openeinstein/: platform code (routing, tools, gateway, agents, campaigns, persistence, tracing, evals, CLI)
• tests/: unit, integration, and eval coverage
• campaign-packs/: installable campaign-specific research packs
• configs/: runtime configuration and machine-enforced policy files
• docs/: architecture, canonical references, audits, and build plans

License

MIT