ProofGrid is an open, sovereign claim-to-evidence auditing system built for Hack Apertus.
It addresses a practical failure mode in regulated, public-sector, procurement, compliance, and policy workflows: fluent AI answers often do not show whether each claim is actually supported by the source material.
ProofGrid converts source documents and review claims into a deterministic evidence matrix. It combines line-aware document chunking, SHA-256 source fingerprints, deterministic retrieval, and an Apertus-based evidence judge that is restricted to retrieved source text. Every decision is normalized to supported, partial, or unsupported, and the system fails closed when evidence is insufficient.
Core components:
- deterministic line-aware document chunking and hashing
- deterministic lexical retrieval for reproducibility
- Apertus evidence adjudication over bounded retrieved context
- line-level provenance and source hashes
- audit-manifest fingerprinting
- FastAPI API and compact browser UI
- evaluation metrics for decision accuracy, citation coverage, and unsupported-claim precision
The architecture is deliberately compatible with sovereign deployment. Apertus can run behind an OpenAI-compatible endpoint using vLLM or the competition inference environment, while the rest of the stack can run locally without sending source documents to third-party SaaS services.
A local prototype has been implemented and verified with deterministic tests. The next build stage is live Apertus integration, challenge-specific evaluation, public packaging, and the final Hack Apertus submission.
Log in or sign up for Devpost to join the conversation.