nomue

Give research agents a separate statistical verification call

nomue gives research agents a separate place to check supported scientific and numerical decisions. Welch is the first working vertical slice of a much larger capability architecture.

Limited Release 1: Approved recipients can submit data and required scientific declarations for a supported Welch calculation, or submit a claimed result with structured evidence for checking, through authenticated MCP and HTTP. The service returns scoped outcomes, reasons, evidence, and next actions. Access and release scope. Public self-registration is not available. Hosted results are not yet public Records for local replay.

Install and run · Add the MCP server · verifier on npm · MCP on npm · See examples

Public release candidates · verifier 0.2.1-rc.1 · MCP 0.2.0-rc.1 / stdio · Linux / macOS / Windows · Node.js 20 / 22

Product development since limited Release 1

Since limited Release 1, nomue development has expanded candidate calculation range, improved completion of difficult calculations, clarified what was checked, and added historical-result handling.

Development update: integrated candidates and completed internal milestones are not a new hosted release or an expansion of public verifier support.

Making verification results more useful to research agents

Available now

Verify a Release 1 Welch Record locally

Anyone can install the public @licklider/nomue-verifier package from npm and run nomue verify locally to check a conforming Release 1 Record for independent two-group continuous outcomes under the two-sided Welch two-sample t procedure. It recomputes the covered numerical quantities and returns a machine-readable report of the scoped checks without calling a nomue server after installation.

Check the exact bundle

Confirm that the Record uses the published Release 1 profile, procedure, and version.

Recompute covered quantities

Recalculate the covered numbers and compare them using the published rule.

Return structured evidence

Produce scoped pass, fail, refusal, or unsupported outcomes in a JSON report.

npm install --global @licklider/nomue-verifier@0.2.1-rc.1
nomue verify ./record.json --format json

# Or run without a global install
npx --yes @licklider/nomue-verifier@0.2.1-rc.1 verify ./record.json --format json

After npm installs the package and its dependencies, Record verification stays on the local machine and does not call a nomue server. The package path is exercised in CI on Linux, macOS, and Windows with Node.js 20 and 22.

For an agent workflow, this is a concrete verification-call boundary: provide a versioned Record to a separate command and carry its scoped JSON result back into the workflow without asking the research model to judge its own calculation.

The same bounded verification is available to MCP clients through the exact@licklider/nomue-verifier-mcp@0.2.0-rc.1 local stdio server. It exposes only verify_nomue_record and does not add a broader verdict.

The same CLI provides canonicalize and digest utilities for local Record processing. Their exact arguments and output behavior are published before installation in the CLI reference.

Install and run the verifier · Read the CLI reference

What nomue is

nomue is Licklider's agent-callable scientific verification product. It moves important scientific decisions and numerical checks out of the research model and into a separate, inspectable capability.

nomue's Welch verification is implemented: it keeps the agent from guessing required scientific facts, follows the supported decision path, runs fixed numerical checks, and returns structured outcomes and next actions.

nomue asks a precise question: given these research conditions and this calculation or result, what exactly can be checked, and what does that check establish?

The result tells the agent what to do next: execute a supported path, ask for one missing scientific fact, route an unsupported request elsewhere, preserve a failed check, or stop safely.

A deliberate division of roles

Research agent

Understand the task, gather context, explain required declarations, orchestrate the call, and explain the returned evidence and reason codes.

nomue

Apply the exact capability and version boundary, protect scientific declarations from silent inference, execute the registered deterministic path, and return scoped machine-readable outcomes.

Researcher

Supply scientific facts that cannot be inferred, resolve experimental meaning, and remain responsible for the research judgment.

How a verification call works

Name the check. The agent identifies the exact question and version instead of asking whether the whole analysis is “correct.”
Supply the facts. Required scientific information comes from an authorized source. If it is unknown, it stays unknown instead of being guessed.
Choose the next action. nomue decides whether the check can run, needs clarification, falls outside current support, or must stop.
Run the check. A supported numerical answer comes from a fixed calculation and its published rule, not generated prose.
Return the evidence. The outcome includes the scoped result, evidence, versions, next action, and the exact boundary of the check.

Read the verification-call guide

Product and roadmap

Welch is the first working, publicly checkable vertical slice, not the product boundary. The adopted product sequence expands both the platform beneath each call and the scientific methods available through it.

Use today

Public npm package, Protocol, and local verifier

Install @licklider/nomue-verifier 0.2.1-rc.1 from npm and run it locally against the nomue Protocol Release 1 Public Draft support target for independent two-group continuous outcomes under the two-sided Welch two-sample t procedure.

How to use the public verifier

Use today — release candidate

Local MCP server

The exact @licklider/nomue-verifier-mcp@0.2.0-rc.1 release candidate is public on npm and in the official MCP Registry as nomue Record Verifier. It exposes the method-neutral verify_nomue_record tool over local stdio; the current supported scientific scope remains Release 1 Welch Record verification. It delegates to @licklider/nomue-verifier@0.2.1-rc.1 and has passing package-path CI across Linux, macOS, and Windows. It requires no account, API key, environment variable, or Licklider-hosted service. The first npx launch may download npm dependencies; after installation, verification runs locally. This release candidate supports stdio only: it is not a hosted HTTP endpoint and does not add paired-t, Wilcoxon, Mann–Whitney, method selection, raw-sample calculation, or an overall scientific verdict.

Installation and tool boundary · npm · Official Registry · Source

Limited Release 1

Welch verification inside agent workflows

Use the npm-published Release 1 verifier, the local stdio MCP server, the Protocol, and their machine-readable documentation today. The agent-facing Welch capability is now available in limited Release 1 to approved recipients through authenticated MCP and HTTP interfaces; public self-registration is not available. The hosted capability does not yet emit public Records for replay through the local verifier.

Read the release announcement

Planned platform evolution

Independently verifiable Records

Carry exact capability, engine, Protocol, and result evidence into a Record that another party can check without trusting the product runtime.

Agent-native Project state

Let an authorized agent return to the same research task without rebuilding canonical context from chat history.

Persistent resumability

Reconnect and continue long-running research work while keeping project truth separate from conversational memory.

Generalized capability kernel

Stabilize the common contract, result, refusal, evidence, and version surfaces after materially different methods prove what is shared.

Planned scientific capability families

Independent multi-group · Paired two-group · Repeated measures · Factorial and interaction · Nonlinear and dose response · Nonparametric rank-based · Categorical outcomes · Correlation and linear models · Survival time-to-event · Count outcomes

Method expansion does not require every app capability to wait for a Protocol release. A method family can first ship with app-level deterministic verification after its own contract, independent-oracle, cross-implementation, property, adversarial, and activation reviews close. It can later gain Protocol-backed independent verification through an exact released bundle.

Protocol progress

Release 2: reviewed paired-t candidate

The Release 2 paired-t candidate now has an independently reviewed formal decision-readiness packet. It assembles the D1–D6 decision ledger, numerical and execution evidence, structural candidates, review dispositions, Release 1 safeguards, and the required coupled landing order. The Steward decisions, authoritative issuance, support activation, and release remain open.

Paired-t public RFC

Release 3: independent-group public discussion

Release 3 public discussion is open on independent groups and multiple comparisons. The proposal makes design, comparison families, result meaning, and error-control questions explicit across 49 catalogued procedures. Its evidence scope is limited to supplied originals; method adoption and numerical support remain separate decisions.

Scope and evidence · Join the discussion

Unequal variances · Control and best-treatment targets · Testing graphs

Release 4: two-factor public discussion

Release 4 public discussion is open for a balanced two-by-two fixed-factor proposal. Its unissued numerical, report and controlled-execution candidate has reached independently reviewed final readiness, without establishing Protocol support. A separate amendment discussion covers strict binary64 comparison and completed indeterminate results; it changes neither the current verifier nor the original RFC clock.

Scope and evidence · Join the discussion

Release 5: shared design and timing evidence

Release 5 public discussion is open on a common evidence view for declared study design and selection timing across analysis families. The proposal covers versioned mappings, timing declarations, explicit limits on what a passing check means, and a shared report view. All three candidate families require separately accepted successors; no new verification capability is available.

The intended benefit is a common way for researchers and agents to inspect finalized declarations across supported families. Method recommendations remain a separate product concern.

Scope and evidence · Join the discussion

See the full roadmap · Follow Latest · RSS

The public trust layer

Anyone can inspect the public meaning of a supported Record without relying on Licklider's product service. The nomue Protocol defines public Record and scoped verification semantics, and the open-source verifier checks the covered properties locally and offline.

Release 1 publishes source code together with public fixtures, a signed source archive, checksums, a snapshot manifest, detached signatures, and the public release key.

The product layer remains separate. It can manage agent-facing intake, clarification, execution, and operational behavior without making every product implementation choice part of the royalty-free public Protocol.

nomue Protocol
nomue verifier on npm
nomue verifier

Agent-readable by design

The documentation states when a verification call applies, when it does not, which facts must not be inferred, the possible non-success outcomes, and the limits of a successful result. Clean Markdown pages and machine-readable examples expose the same underlying boundary to agents and researchers.

This documentation helps an agent select a capability, supply the right facts, and interpret the returned outcome. The versioned runtime schema and public Protocol remain the exact execution and verification contracts.

Agent-readable documentation

What a successful nomue result means

A successful result means that the named checks passed under the stated capability, procedure, evidence, and version. It gives the research agent and researcher a checkable basis for that bounded decision.

The following questions remain with the researcher or another capability:

the truth of input data or researcher declarations;
the overall correctness of a research project;
the truth of a scientific or causal conclusion;
validity for a method outside an explicitly supported scope;
publication, regulatory, or clinical acceptance;

Keeping these questions separate makes the positive result useful: each passing check says exactly what can be relied on and why.

Less work to reach a historical-analysis reuse decision

Adding nomue reduced API cost by 78.1% and task time by 62.9% on historical-analysis reuse decisions the ordinary-tool comparator also answered correctly. 103 matched pairs where both configurations reached the correct six-field decision with execution evidence, from a 48-task synthetic Welch study using gpt-5.6-sol.

The preprint evaluates assigned use of nomue’s historical-analysis recheck capability. This product evaluation is separate from the public local Record Verifier MCP package and does not expand its released support.

Read the study and its limits · Preprint v1 and supplement (not peer reviewed)