Validation
"Validated" is a checkable claim here, not a marketing word. The repository contains a versioned reference-case suite (tests/validation_cases.json) that runs on every CI build. Each case states its inputs, its expected values, the tolerance, and where the expected values come from.
External anchors (public domain sources)
| Case | Source | What must match |
|---|---|---|
capability-nist-michelso |
NIST StRD, dataset Michelso | certified mean 299.8524 and certified std 0.0790105478190518, to 1e-9 |
capability-nist-numacc1 |
NIST StRD, dataset NumAcc1 | exact certified mean and std on a catastrophic-cancellation probe |
ewma-nist-handbook |
NIST/SEMATECH e-Handbook 6.3.2.4 | the published EWMA limits and, in the unit tests, the full published EWMA series |
Derived cases
The remaining cases are synthetic datasets whose expected values are fully derived in the case file itself (formula by formula, with literature references: Montgomery 8th ed. for charts and capability, Nelson 1984 for the rules), at tolerance 1e-9. Synthetic-with-published-derivation means the suite is redistributable without copyright concerns; commercial worked examples (e.g. from the AIAG manuals) are checked privately and never shipped.
Constants are computed, then tested against tables
d2, d3, and c4 are computed from their defining integrals (closed forms for
n = 2, numerical integration otherwise). The published tables serve as test
oracles, not as the source: tests/test_constants.py checks the computed
values against Montgomery's Appendix VI to table precision and against the
exact closed forms where they exist.
Policy
- Every new analysis ships with at least one reference case in the suite.
- Reference values are never adjusted to make a test pass; if a case fails, the code is wrong or the case's derivation is wrong, and either fix is a reviewed change.
- Wording discipline: results are "validated against NIST reference data" and published formulas. No compatibility claims about commercial products.