Published
nomue launches limited Release 1 for agent-callable Welch verification
nomue now gives approved recipients a way to call bounded Welch statistical verification from compatible AI agents and HTTP clients.
nomue's first agent-callable Scientific Verification Capability is now available in limited Release 1.
Approved recipients can connect a compatible AI agent or HTTP client to nomue and use a hosted verification service for a defined statistical scope: independent two-group continuous data analyzed with a two-sided Welch two-sample t procedure.
Public self-registration is not yet available, and access is currently limited to approved recipients.
What the release can do
Limited Release 1 provides two verification paths.
Compute and verify allows an agent to submit data and the scientific information required for the supported Welch analysis. nomue checks whether the request falls within the supported scope and, when it does, runs the registered deterministic calculation and returns a structured result.
Verify a claimed result allows an agent to submit a claimed Welch result together with structured evidence. nomue then checks that claim against the registered execution and numerical rules.
In both cases, nomue returns the result together with the relevant reasons, evidence, and suggested next action.
If important scientific information is missing, nomue asks for clarification instead of filling in the missing meaning itself. Requests that are outside the supported scope are also kept separate from calculation or system failures.
The capability is available through authenticated MCP and direct HTTP interfaces.
Why this matters for research agents
AI agents can help researchers organize tasks, interpret information, and coordinate workflows. Statistical verification, however, should not depend on an agent inventing missing experimental meaning or treating its own generated answer as the authority for a calculation.
nomue provides a separate verification step for those bounded decisions.
For supported requests, the numerical result comes from nomue's registered deterministic engine. When a request cannot proceed, the response indicates whether the agent should ask for more information, route the request elsewhere, or stop.
This allows an AI agent to carry a scoped verification result back into a research workflow without confusing a successful API call with a successful scientific check.
Limited Release 1 is the first operational release of this approach for nomue. It does not make claims about the relative performance of AI models; evaluation of frontier models remains a separate program.
This release is limited to the defined Welch scope and does not assess the scientific validity of a study as a whole.
Before release, the hosted capability was tested through its registered MCP and HTTP paths using synthetic fixtures. These checks covered both supported capability paths, authentication failures, credential replacement, service recovery, and the deployed runtime configuration.
These are operational release checks. They are not an independent security audit and do not guarantee future service availability.
Public documentation for the nomue Protocol and the locally runnable Release 1 Record verifier remains available through the nomue documentation.
Compare the available nomue paths · Product and Protocol roadmap
The hosted capability does not yet emit public Records for independent replay through the public Record verifier.