WARDEN v2 implementation bundle

Start with RUN_ORDER.md for installation and the exact execution order. Results and current status: RESULTS.md.

python -m pip install -r requirements-v2.txt
python -m pytest -q
python -m warden demo --out runs/demo

This update adds structure-based candidate proposals, a bounded feedback loop, MCP literature retrieval, strict calculation/admission checks, persistent records, and deterministic reports. The offline demo is synthetic; live services and MACE scientific performance have not been verified. See VERIFICATION.md.

v3 kinetic layer: site-resolved reaction barriers (GFN2-xTB or MACE NEB on clusters carved from the amorphous slabs), an acylation chemistry template, dose-window analysis, and an uncertainty-aware multi-cycle process designer with control-gated ranking. See V3_KINETICS.md.

Evidence-acquisition agent: decides which calculation to run next (molecule, site, pathway, orientation, or method), with an equal-budget benchmark against baselines. See AGENT_DECISIONS.md.

LLM backends: the campaign planner/critic can run on local Ollama or the OpenAI API, per role. See LLM_PROVIDERS.md.

The original README is preserved as README_LEGACY.md. Use python -m warden, not the old UI/pipeline entrypoints.

For live literature retrieval, authentication, and candidate selection informed by previous results, see BOND_SEARCH.md. warden bond-call saves source evidence independently of the full screening campaign.