philoweeb/Vergil-PolySafe

PolySafe (Vergil)

Research code, documentation, and evidence artifacts for PolySafe, a drug–drug-interaction (DDI) side-effect prediction system with a certified selective-abstention layer. This is a final-year BTech research + patent-development project (VIT Vellore).

Research use only. PolySafe predicts pharmacovigilance associations, not clinical safety or prescribing advice. Nothing here is clinical software.

What is in this repo

• src/polysafe/ — the library: models (GINE pair encoder, additive/matched family), reliability machinery (selective risk, surfaced-set FDP, endpoint-disjoint unit construction, certificate capacity), features, data pipeline.
• services/api/ — a FastAPI serving layer (predict / alternatives / explain) with an honesty-gated abstention contract.
• apps/web/ — a Next.js demo instrument and a /research evidence ledger.
• scripts/ — every experiment driver (RQ-SC* score-comparability, RQ-G* certification gates, RQ-SEE* selective-emission, RQ-FEAS1 certificate capacity, RQ-MARG1 marginal decomposition, symmetry audits, …).
• tests/ — unit + conformance tests.
• research/ — evidence log, notes, and (public subset of) the decision/preregistration trail.
• experiments/ — immutable per-experiment report.json / manifest.json and the score arrays needed to recompute results.
• data/processed/ (small, committed) — e.g. ChemBERTa embeddings metadata, ATC map.

What is NOT here (deliberately)

• Heavy raw data and checkpoints (TWOSIDES combo export, trained .pt/.ckpt, 15 GB of experiments/ intermediates) are gitignored/regenerable and not included. See data/raw/MANIFEST.json and scripts/acquire_data.py for provenance.
• Patent-strategy documents (research/PRIOR_ART.md, docs/IP_READINESS.md, research/DECISIONS.md, research/PREREGISTRATION.md, research/submission/patent-counsel-disclosure.md, the invention assessment) are excluded — they remain private during the provisional-patent window.

Reproducing

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.lock.txt
# dataset/checkpoint acquisition: see scripts/acquire_data.py + data/raw/MANIFEST.json
.venv/bin/python -m pytest tests/ -q

Every headline number traces to an experiments/*/report.json and a docs/PAPER_KIT.md claim row. See research/ for the append-only evidence log.

License

Code and artifacts are released for research and educational use only. No clinical use. No warranty. The underlying TWOSIDES/BioSNAP data carry their own upstream redistribution terms — see data/raw/MANIFEST.json.