Statistical validation
HypoNum · Status as of 28 September 2026 · Pre-release assessment
HypoNum’s validation combines published statistical methods, independent numerical comparisons, and checks of boundary conditions and invalid inputs. The aim is to make each assessment traceable: what was calculated, which reference was used, how closely the results agreed, and where the conclusion applies.
The completed September 2026 reference assessment contains 149 distinct input-and-setting cases. All 149 met their specified acceptance criteria. This assessment concerns selected calculations in version 0.1.0 on Windows 11. The total counts cases, including some expected refusals and boundary outcomes; it is not a count of independently validated methods or a product-wide pass rate.
Evidence established so far
The customer validation report documents selected comparisons in these areas:
| Area | Examples of assessed calculations |
|---|---|
| Statistical tests and method comparison | Grubbs’ single-outlier test, normal Anderson–Darling calculations, and Deming regression coefficients and jackknife intervals. |
| Process monitoring | Selected Shewhart, CUSUM, EWMA, moving-average and multivariate control-chart calculations, including specified limits and signal decisions. |
| Capability and measurement | Selected capability indices and uncertainty calculations, Type 1 gage measures, supplied-parameter Johnson transformations, normal tolerance factors, and bounded Fieller intervals. |
| Sampling, design and optimization | Binomial acceptance probabilities, tolerance-factor sample-size planning, and the importance-weighted desirability objective, including its zero boundary. |
Coverage applies to the stated examples, outputs and settings within each area. A checked chart calculation does not establish every chart’s false-alarm performance; a checked optimization objective does not establish a global optimum.
How results are checked
References include published papers and worked examples, the NIST/SEMATECH e-Handbook of Statistical Methods, and independently maintained statistical software. Selected comparisons use R packages including mcr, tolerance, nortest and twopartm.
The assessment records inputs, method conventions, reference versions, expected and actual results, and acceptance criteria. Numerical tolerances reflect the calculation and the reference’s precision. Counts and specified decision outcomes require exact agreement where applicable. Differences caused by rounding, alternative conventions or implementation errors are investigated and documented.
Validation in progress
Further reviews cover regression, survey methods, reliability and agreement, multivariate analysis, survival, bootstrap inference, meta-analysis, power and simulation. These reviews have produced additional successful comparisons and identified corrections requiring verification. Their results are not added to the 149-case assessment or presented as completed validation of those method families.
Broader assessment of forecasting, graphical output, prediction uncertainty and application workflows, together with release-specific verification, remains outstanding. Validation conclusions apply to the assessed implementation and settings; they do not automatically extend to later changes or every available option.
The detailed statistical validation report provides the method-specific examples, references, comparison criteria and limitations behind this summary.