Validation, with limits in view

Beta / provisional. Independent statistical review and the release checklist are pending. NIST reference data do not certify this software. Tests check selected cases, not universal accuracy.

Independent numerical references

Checked-in fixtures were generated using SciPy 1.17.1 (not the JavaScript engine): ANOVA, balanced and unequal-size Tukey intervals and probabilities, mean Student-t intervals, and median-centered Levene. R was unavailable and no R run is claimed. The analytic smoke case yields F=3 and p=0.125. Detailed test commands and results are recorded in the repository validation document.

Dose-response references use SciPy 1.17.1 least_squares with direct physical parameters and a finite-difference Jacobian, independently of the browser implementation. Tests cover increasing/decreasing 4PL, fixed-bottom/fixed-top 3PL and 95%/99% intervals, with additional unit-conversion, technical-replication and failure guards. Selected reference agreement is not proof that every nonlinear fit or interval is reliable.

The comparison engine is checked against 48 SciPy 1.17.1 design/sidedness/alpha/adjustment cases and five additional Welch stress cases. Exact hand-computed cases independently check Holm adjustment. These checks do not establish design suitability or certify all inputs.

Known failures and guarded outputs

The lower-difficulty NIST SiRstv test passes. Higher-difficulty SmLs07 fails the certified-value tolerance: binary64 parsing loses decimal information at its large common offset. Centered sums cannot recover information already lost in input conversion. That failure is retained in the test suite; the software is not validated for that precision regime.

The original jStat F-tail and low-df studentized-range implementations failed the initial accuracy checks. The beta continued fraction was tightened and independently retested. Two-group Tukey uses the exact Student-t/F relationship. For three or more groups, probabilities require residual df≥max(k,6), df≤9998 and q≤10. Quantiles outside that df domain are withheld. Adjusted tails below 10⁻⁵ are unavailable; no unvalidated bound or exact zero is invented. Near-alpha results within the numerical error envelope are marked uncertain.

Tests use a 2×10⁻⁷ absolute CDF envelope in the supported studentized-range domain, combined with relative checks for adjusted tails and separate interval tolerances. F tails use relative checks, including very small values. Quantile inversion is bracketed; it does not make the underlying CDF exact. No claim of four-decimal accuracy is made for all possible inputs.

Release status

Consult docs/numerical-validation.md, docs/progress.md and the browser test report for the exact executed checks, known defects and supported environments. No scientific validation badge or research-ready certification is claimed.