Know when to proceed.
Check whether the required research conditions are in place. Proceed with a supported analysis, ask for missing facts, or stop when the request cannot proceed.
Determinize AI.
nomue gives AI-driven statistical analysis a clear path to proceed, verifiable results, and next actions grounded in evidence.
Life sciences. Finance. Chemistry. Wherever claims must hold.
Why
Every analysis needs a basis to proceed.
Before a calculation, an agent needs to know whether the required research conditions are met. Afterward, it needs to know what the result establishes and what to do next.
nomue turns declared conditions and evidence into machine-readable decisions.
What nomue does
A clear decision. Evidence. A next step.
Know what was checked.
Check supported calculations and claims against the supplied data and evidence. See what passed, what failed, and what each check establishes.
Know what comes next.
Give the agent a structured outcome, reasons, and next actions, so it can continue the workflow or bring a question back to the researcher.
Evidence
Measured, not assumed.
216 / 216
Expected complete reports with nomue
Compared with 132 / 216 using general Python and 131 / 216 using a stable Welch wrapper.
72 constructed Welch-report questions · three repetitions · 648 sessions
78.1%lower API cost
62.9%less task time
With nomue, compared with ordinary tools alone
Measured across 103 matched pairs where both configurations reached the correct answer.
48 synthetic Welch tasks · three repetitions · 288 scheduled sessions
These results apply to the tested Welch workflows. Both studies were led by nomue's developer and published as preprints, not peer-reviewed findings.
Scope
Proven narrowly. Built to expand.
Available today
Welch checks on your own machine
The public nomue package checks supported two-sided Welch analyses locally on Linux, macOS and Windows.
Hosted Welch checks are available to approved recipients in limited Release 1.
Long-term direction
Life sciences. Finance. Chemistry.
The long-term scope is any field where AI-generated claims can be checked against declared data, calculations or rules.
Each new capability will state what it checks, the evidence behind it and where it stops.
Open work
Evidence in the open.
We test the software scientific claims depend on, report the failures we can reproduce, and publish the research behind nomue.
Upstream reports
14 reports submitted 5 fixes merged upstream
Each report records the finding and the upstream decision. How to read the outcomes
SciPy
Student-t: a representable subnormal tail returns zero
Closed · NumPy fix not planned
Returned two-sided p → independent reference
0.0≈ 3.505e-316
jStat
Noncentral t: a probability near 44% returns zero
Reported · confirmation pending
Returned CDF → independent reference
0.0≈ 0.4410
statsmodels
Contingency-table test: a representable chi-square tail returns zero
Fix merged upstream
Before patch → after patch
0.0≈ 7.076e-18
statsmodels
Welch t-test: a finite result is lost after sumsquares overflow
Fix merged upstream
Before patch → after patch
NaN≈ 0.26657
SciPy
Welch t-test: finite variances, wrong degrees of freedom
Repair proposed · checks passing
Returned degrees of freedom → exact reference
14
SciPy
Studentized range: a nonzero tail returns zero
Reported · confirmation pending
Reported SF → exact special-case reference
0.0≈ 2.0e-8
SciPy
Welch ANOVA: exact rescaling reverses a 5% decision
Repair proposed · checks passing
Original scale → scaled by 2^-511
0.026500.05611
R / agricolae
REGW: renaming groups changes a 5% decision
Emailed · confirmation pending
Same observations: original labels → renamed labels
0.03630.0791
SciPy
Mann–Whitney U: batching can change a 5% decision
Under discussion · remedy undecided
Same pair: alone → batched with a tied pair
0.047980.05132
SciPy
SciPy t-tests can return p=0 or p=1 after exact rescaling
Fix proposed · high-precision results posted
Returned after rescaling → scale-invariant reference
0.0 / 1.00.2048 / 0.03510
Julia / HypothesisTests.jl
Exact signed-rank p-value above 1
Matching fix released
Registered 0.11.8 → merge commit
1.251.0
SciPy
SciPy exact Wilcoxon p-value error
Fix merged upstream
Before patch → after patch
0.08.12511917099255e-17
Boost.Math / SciPy
Student-t extreme-tail sign error
Fix merged upstream
Returned → expected
+∞finite negative quantile
R
R exact Wilcoxon out-of-range p-values
Reported · unconfirmed
R result → exact reference
-7.55e-152.59e-18
Research papers
Results worth examining.
Reducing the cost of historical-analysis reuse decisions with nomue: a paired, equal-evidence LLM evaluation
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.
Preprint v1 — not peer reviewed; synthetic Welch tasks; aggregate reproduction supplement available
Comparing three tool-assisted AI configurations for whole-submission verification of Welch reports
A 648-session fixed-panel study compares general Python, a stable Welch wrapper, and a bound verifier for complete five-field report correctness.
Preprint v1.0 — not peer reviewed; fixed constructed panel; reproducibility package available
Same Test, Different p
An audit of R, SciPy, and Julia shows why matching statistical decisions can still hide differences in definitions, numerical precision, and valid probability values.
Preprint v1.0 — not peer reviewed; reproducibility data and code available
Correctly Rounded or Refused
A preprint on checking the numerical accuracy of paired Student-t test results before software returns them.
Preprint v0.2 — not peer reviewed
Follow the evidence
New results are published with their tested scope, evidence and limits.
- September 27, 2026SciPy returns zero for a representable Student-t tailEngineering
- September 27, 2026nomue Protocol opens public discussion on an exact-value Welch successorNews
- September 23, 2026Reducing the cost of historical-analysis reuse decisions with nomue: a paired, equal-evidence LLM evaluationResearch
- September 23, 2026jStat returns zero for a probability near 44%Engineering
- September 22, 2026statsmodels returns zero for a representable chi-square tailEngineering
- September 19, 2026Preserving Welch confidence intervals near zeroEngineering
- September 19, 2026Keeping independent checks visible when verification failsEngineering
- September 18, 2026Preserving Student-t probability near zero at one degree of freedomEngineering
Company
Built by Licklider.
nomue is developed by Licklider, Inc. We publish what it checks, the evidence behind it and where it stops.