VETO — Verification Engine
Multi-model epistemic verification engine that validates scientific claims through adaptive jury deliberation, achieving 90% cost reduction over brute-force approaches.
Overview
Why Verification Matters
In the era of AI-generated content, the ability to verify scientific claims at scale is no longer optional — it is infrastructure. VETO addresses this by providing a transparent, cost-efficient, and model-agnostic verification pipeline that treats every assertion as a hypothesis to be tested.
Unlike generic fact-checking tools, VETO is purpose-built for scientific and technical domains, where nuance, citation integrity, and multi-source triangulation are essential.
Core Capabilities
- Cost Pyramid: intelligent claim triage (SKIP → CITATION_ONLY → STANDARD → CRITICAL) routes 90% of claims through lightweight verification, reserving expensive multi-model deliberation for high-stakes assertions only.
- Adaptive Jury: starts with 3 judges and escalates on disagreement — not 9 models by default. This keeps costs predictable while maintaining rigorous consensus on contested claims.
- Source Fabric: hierarchical retrieval engine capable of searching across 10,000+ sources with deduplication and semantic ranking.
- Convergence Loop: iterative deliberation with 5 circuit breakers prevents runaway costs and infinite loops.
Technical Specs
Architecture Deep Dive
- ClaimClassifier: local 8B-parameter model for claim-level triage — zero API cost for classification.
- CitationVerifier: DeBERTa NLI model for citation verification — no LLM calls required for evidence grounding.
- HybridReflector: template-first reasoning engine that invokes LLMs only on conflicts, reducing token consumption by up to 80%.
- Governor: budget-aware orchestrator with real-time cost tracking and graceful degradation policies.
- Storage: SQLite backend with full CLI, SDK, and API integration for seamless deployment.
Deployment & Integration
- Docker: multi-stage Dockerfile with docker-compose for one-command deployment.
- CI/CD: GitHub Actions workflow with lint + test on Python 3.11/3.12.
- API: FastAPI-based REST interface for programmatic access.
- Extensible: 6 model connectors (Anthropic, OpenAI, Google, Groq, DeepSeek, Mistral) with plug-in architecture for custom providers.
Quality Assurance
VETO ships with 1,053 automated tests covering 48 modules, 8 verification rubrics, and all integration paths. The benchmark suite includes 50 seed claims for correlation analysis and regression testing.
Project Partners
No partners listed.
Timeline
Core verification engine with Cost Pyramid and Adaptive Jury
Source Fabric with 10K+ source hierarchical retrieval
API deployment and Docker containerization
Real-world LLM benchmark and error correlation validation
Gallery
No media available.