ReproForge Sentinel — Claim-to-Evidence AI

AI systems can generate benchmark claims, model promises, and technical demonstrations faster than teams can verify them. ReproForge Sentinel closes that trust gap by turning an AI/ML claim into structured, inspectable evidence. A user submits a repository URL, the exact claim, a runtime target, and declared security policies. ReproForge evaluates the available claim metadata and policy signals, applies deterministic ShadowGuard risk and reproducibility scoring, records a trace of the evaluation, and produces a Reproducibility Passport. The Passport contains the verdict, evidence chain, missing proof, blocked actions, integrity hashes, security notes, and exportable JSON/PDF results. The hackathon build combines a premium React and TanStack interface with a FastAPI backend, Docker packaging, tests, proof schemas, and AMD/Gemma integration adapters. It also includes an AMD ROCm capture workflow that accepts hardware evidence only when device identity, HIP/ROCm, AMD SMI telemetry, workload metrics, and artifact hashes are successfully captured. Our guided sample is clearly labeled and designed for a reliable judge walkthrough. The current public MVP evaluates submitted claim metadata and declared policies; it does not yet clone or execute arbitrary repositories. Real Fireworks/Gemma inference and direct AMD ROCm telemetry remain explicitly marked pending when verified runtime provenance is unavailable. ReproForge never replaces missing measurements with invented proof. ReproForge is not a “truth machine.” It is an evidence machine: a verification layer for AI teams, security reviewers, researchers, investors, and judges who need to understand what was checked, what passed, what failed, and what still cannot be proven.