Built at The Agent Harness Hackathon

TrueDrug

Loom Link: `https://www.loom.com/share/0a63d01d149d4266b82c5dd85c94c764` # truedrug ***truedrug**** is a prototype that estimates the probability an FDA drug application will be approved during its first review cycle.* *It focuses on once-weekly injectable GLP-1–based treatments for type 2 diabetes.* ## What it does The system turns pre-action public FDA evidence into a blinded dossier, then runs independent reviews across three disciplines: - CMC — chemistry, manufacturing, and controls - Pharm/tox — pharmacology and toxicology - Clinical — clinical evidence and safety Those reviewers produce structured, evidence-grounded findings. Deterministic code—not an LLM—then combines findings with a cited base rate and versioned likelihood-ratio table to calculate an approval probability and assemble a draft Complete Response Letter (CRL). ```text Public evidence → blinded dossier → CMC / pharm-tox / clinical reviews → validated findings → deterministic probability + draft CRL → frontend demo + evaluation report ``` ## Why blinding matters Models may already know drug outcomes. To avoid outcome leakage, the system removes drug, sponsor, trial-program, site, registry, application, and identifying URL information before review. The harness never receives outcome labels or blinding maps. ## Project structure - `docs/domain/` — shared contracts, checklists, regulatory context, and leakage rules - `specs/001-data-collection/` — evidence and dossier preparation - `specs/002-frontend/` — planned demo interface - `specs/003-harness/` — reviewer orchestration and prediction pipeline - `specs/004-evals/` — offline evaluation and leak-probe scaffold - `harness/` — harness implementation work - `evals/` — synthetic evaluation fixtures and checks - `splash/` — visual design material ## Local setup ```sh npm ci --ignore-scripts npm run trueforge ``` Open TrueForge at `http://localhost:8790`. TrueForge is the local agent configuration environment. The polished truedrug frontend is planned separately and is not currently served by this repository. ## Principles - Reviewers use only dossier evidence and their assigned checklist. - Each checklist item receives exactly one finding. - Findings must be grounded in dossier evidence. - Model output is validated and may be repaired once; invalid output fails rather than being fabricated. - Approval probability is computed in code. - Held-out evaluation cases are never used to tune prompts or rules. - Evaluation claims include a leak-probe result and are demonstrations, not statistical proof. ## Team **FirstCycle** A project about making first-cycle FDA review reasoning clearer, more traceable, and easier to evaluate.