Verity: MCP-Based Fact-Checking Layer for Local AI Systems

Signal

Verity: MCP-Based Fact-Checking Layer for Local AI Systems · ICCL · 2026-05-14

Verity is an open-source implementation of a Model Context Protocol (MCP) layer designed for local AI systems. It functions as a fact-checking intermediary that reviews model-generated answers and citations against trusted sources before the system treats them as reliable. The tool introduces a verification gate into the inference pipeline, aiming to reduce hallucination propagation by enforcing source grounding at the runtime level.

Context

Verity addresses the reliability gap in local autonomous workflows where models operate without immediate cloud-based retrieval or verification services. By leveraging the MCP standard, it decouples the verification logic from the base model, allowing the fact-checking layer to be swapped or updated independently of the inference engine. The architecture positions Verity as a middleware component that intercepts output streams, performs citation validation, and flags unverified claims, thereby introducing a structured accountability checkpoint for deterministic and probabilistic agents alike. This approach treats factuality as a protocol-level concern rather than a prompt engineering variable.

Relevance

This entry stabilizes the pattern of runtime verification as a distinct infrastructure layer rather than relying on post-hoc corrections or static guardrails. As local AI systems gain autonomy, the cost of factual errors increases; Verity demonstrates a shift toward "verification-by-design" where factual integrity is enforced through executable tooling. This supports the broader transition from open-weight models to trustworthy operational agents by providing a reusable mechanism for source grounding that aligns with governance requirements without sacrificing local execution. It exemplifies how MCP can be used to compose governance primitives into the agent's tool execution path.

Current State

Verity is available as an open-source project distributed via the ICCL digital data channel. It is implemented as an MCP integration that can be attached to local agent runtimes. The current iteration focuses on answer and source review, providing a structured output indicating verification status. Deployment involves integrating the layer into the agent's workflow, where it acts as a pre-processing or post-processing gate depending on the runtime's MCP support capabilities. The tool is designed to operate within constrained environments, maintaining data sovereignty while enforcing verification policies.

Open Questions

• How does Verity handle conflicting sources or ambiguous citations within local knowledge bases?
• What is the latency overhead introduced by the verification step in real-time agent loops?
• Does the MCP implementation support bidirectional feedback, allowing the agent to refine queries based on verification failures?
• How does the layer manage privacy when verifying against external trusted sources from a local environment?
• Can the verification criteria be dynamically adjusted based on the risk profile of the agent's task?

Connections

Verity intersects with the agent governance infrastructure by providing a concrete mechanism for output validation, complementing broader policy frameworks that define constraints but lack runtime enforcement. It extends the policy-as-code paradigm by encoding factuality checks as executable MCP tools, enabling dynamic verification rather than static rule application. This aligns with the operationalization of governance primitives seen in enterprise toolkits, demonstrating how open-source components can deliver production-grade reliability features for local agent deployments. The tool's MCP-based design also supports interoperability across different agent frameworks, reducing vendor lock-in for verification capabilities.