Technical case: evidence-first UK company verification in Relevance AI
Problem
An AI agent can call an API, but that does not mean it knows whether the source is healthy, the company identity is unambiguous, the evidence supports the intended action, or a payment is authorized. A generic successful HTTP call can hide uncertainty and encourage an agent to continue too far.
Integration design
RunOnProof is exposed through a Relevance AI API Tool step. The first Tool verifies a UK Companies House company number through the Company Check solution. Repository-owned selection and normalization rules turn the canonical response into PROCEED, REVIEW, or STOP while retaining evidence, coverage, freshness, limitations, and economic activity.
The Marketplace submissions use API Tool steps because they expose a narrow product with explicit inputs and prices. Relevance AI also supports the certified RunOnProof remote Streamable HTTP MCP at https://api.runonproof.com/v1/agent/mcp. Direct MCP connection is a supported secondary path, not proof of Marketplace approval and not an x402 wallet.
Request contract
The primary business inputs are the exact company_number and an idempotency_key for the exact request. The Tool performs:
POST https://api.runonproof.com/v1/gb/solutions/company-check
It sends JSON headers, binds the idempotency key, accepts JSON, and leaves 4xx/5xx available to the normalization step. Automatic retry is disabled for the paid POST.
Economic boundary
The product is gb.solution.company_check.v1, documented at 0.04 USDC for the observed case. An unpaid request can return HTTP 402. The Tool maps 402 to REVIEW, displays the product and price, and stops. It has no wallet or payment proof.
Decision safety
Only a canonical allow result with healthy required sources can map to PROCEED. Ambiguity, missing evidence, outage, stale source, hold, retry, no-decision, or payment boundary maps to REVIEW. Denied, blocked, prohibited, and unsupported actions do not continue.
No-match is explicitly limited by list, identity match, policy, and observation time. Company registration evidence is not used to claim bank ownership, invoice authenticity, complete beneficial ownership, universal licensing, solvency, or absence of all risk.
Reproducibility
The repository includes a versioned authoring manifest, JSON schemas, deterministic runtime, system prompt, seven synthetic fixtures, unit tests, capability matrix, support research, installation guides, and rollback instructions.
npm run demo:relevance-ai
npm run test:relevance-ai
The synthetic demo has an economic delta of 0.00 USDC and makes no network call.
Observed result
On 2026-08-10, the Tool and demo agent existed in the private RunOnProof Relevance AI project and were connected. Execution required approval. A synthetic smoke showed the correct product and 0.04 USDC, returned REVIEW, and stopped without a production verification, payment, settlement, or platform credit charge in that smoke.
Limitations
Five Tool submissions are pending and unpublished, so the installation is not Marketplace-public. The out-of-scope Passport submission has been withdrawn. Complete synthetic coverage is repository CI evidence; only a smaller safe smoke was observed in the workspace. Runtime health and price must be read at execution time. Relevance AI UI and platform behavior can change independently of this repository.
Why it matters
The case demonstrates that an agent integration can make uncertainty and cost explicit before a consequential action. The useful product is not merely an API response; it is a narrow, auditable permission boundary with evidence attached.