30,710 scalar values
Thirty-six distribution families were checked for density, CDF, quantile, tails, log probabilities and boundaries. Disagreements were adjudicated with 80-digit arithmetic and independent inversion identities.
Scalar results →JDistlib 0.10.2 · R 4.6.1 comparison
Measure the answer before measuring the speed.
The 0.10.2 release extends the numerical audit from the core R families to contributed distributions, copulas, compound counts, inference, solvers, linear algebra, accelerators, and random engines.
Frozen evidence
Accuracy compared with R and independent references
10,905 joint-law values
Copula, multivariate, finance and compound-count references use R packages, base-R constructions and analytic checks. Unreliable R regions are identified rather than copied.
Joint and finance results →2,076 utility references
Density estimation, bandwidths, moments, normality tests, multiple testing, splines and optimization are checked against R and independent mathematical identities.
Math results →Inference and engines
Posterior diagnostics, PSIS, solvers, scaled decompositions, provider paths, WELL44497b and CMWC4096 have frozen fixtures and state-transition tests.
Inference and RNG results →Matched workloads
Faster where the algorithm changed
| Workload | Change | Comparison |
|---|---|---|
| SJ bandwidth selection | 150× faster than prior Java | 1.13 ms Java · 0.88 ms R |
| UCV bandwidth selection | 479× faster than prior Java | 0.36 ms Java · 0.96 ms R |
| Compound-count CDFs | 33–79× faster than prior Java | Java faster on measured batches |
| Long-chain diagnostics | 2.26× faster than prior Java | 38.9 ms Java · 50 ms R |
These are reproducible, workload-specific measurements. Corrected Gaussian density was slightly slower, Student-t copula CDF remained slower than R, and some scalar routines remain slower than vectorized R. The audit reports retain those results and their limitations.
Independent method sources
The audit checks both answers and provenance
Use the evidence in context
What the audit does not claim
Higher-dimensional quadrature, fixed-grid Fourier finance paths, every specialized sampler, all accelerator operations, and every extreme parameter regime were not exhaustively certified. CUDA tests were skipped on the audit host. Read the detailed scope and reproduction commands before generalizing a result.