Validation

R-level correctness, in one citable binary

Quanta's engine is pure Swift on Apple's Accelerate framework, with zero R or Python at runtime. R and Python are used only as validation oracles during development: we compute reference values with them, match our output to many digits, then ship none of their code.

Live three-way head-to-head

One dataset, eight statistics, run through ReliCheck (Swift), R 4.6 (base), and Python 3 (scipy / statsmodels). All three reading one identical CSV, printed to ten decimal places.

StatisticReliCheck (Swift)R 4.6Python 3
Mean7.05000000007.05000000007.0500000000
SD (n−1)1.46808145481.46808145481.4680814548
Pearson r0.92315371240.92315371240.9231537124
Welch t2.74109566452.74109566452.7410956645
One-way ANOVA F21.966666666721.966666666721.9666666667
OLS slope0.10157419920.10157419920.1015741992
OLS R²0.34490579730.34490579730.3449057973
Cronbach alpha0.93344262300.93344262300.9334426230

All three agree to ten decimal places on every statistic.

Per-method validation record

Every engine is validated against an independent reference, not just internal consistency.

MethodReference oracleAgreement
Univariate summaryNIST StRD (certified)mean 12+ digits; SD ~8 (cancellation-safe)
Linear regressionNIST StRD Norriscoefficients, SEs, R² to 9-10 digits
t-tests / ANOVA / ANCOVAR t.test, aov, carexact / matched
Correlation & nonparametricsR cor; hand-computedexact / closed-form
Logistic / Poisson / ordinalR glm, MASS::polrcoefficients to 1e-3
Reliability (alpha, omega)R / Python; closed-form10 digits / 1e-5
Factor analysis & PCAindependent Python; closed-formrecovers simple structure
Repeated-measures ANOVAR afex / ez + SPSS GLMF, partial η², Mauchly, GG/HF matched
MANOVA / discriminant / canonicalR manova, MASS::lda, cancortest statistics matched
Multiple imputation (MICE)R mice + Rubin's rulespooled estimates, variance, FMI matched
SEM / CFAR lavaan (Holzinger-Swineford)χ² exact; CFI/TLI/RMSEA/SRMR to 3 figs
Complex-survey SER survey (svymean)mean/SE to 1e-6 / 1e-4

Real-file import, proven

The SPSS and Stata readers are pure Swift. They are tested by round-tripping genuine binary files (SPSS .sav, Stata .dta releases 117-119) written by a ReadStat-based writer: variable names, rows, variable labels, and value labels read exactly. Value labels are exposed but never substituted into cells, so statistics are computed on the values you recorded.

Why this matters

Reviewers accept results from software they can verify. This record is the evidence. Quanta ships as one version-stamped, notarized binary with a method file, so the version flows straight into your Methods section: R-level accuracy with no R or Python to install, plus honest-stats guardrails the big packages do not give by default.