What's New

What changed in each release. The notes for the two most recent versions are the same texts shown on the store listings; each version links its own DOI, so a release can be cited exactly. Installed copies update automatically through their store.

0.23.0 — 2026-08-31

DOI: 10.5281/zenodo.22199867

The catalog stays at 108 validated tools. This release is about trust: check what the app prints rather than believe it.

Help → Verify this installation recomputes pinned reference checks on your own installed copy — the same numbers validated before release — and writes a report; it runs headlessly too, with --verify.

Every result now names its data: a dataset fingerprint in each output's footer, beside the active filter, weight and split. File → Export Reproducibility Bundle packs journal, fingerprint and versions into one file; ChakataStat re-runs it against your data and confirms the numbers match in one step.

The methods are in print: every procedure page in the guide gains a "How this is computed" section, each line linking its convention — 48 previously implicit definitions written down. Two defaults move to the conventional choice, Cox ties to Efron and mixed models to REML; saved syntax without those options prints slightly different numbers; one option restores the old behaviour.

Independent evidence: every Python-referenced pin re-derived in R (267 of 267, each divergence a recorded convention), and the full pin set compared across Windows, macOS and Linux; the agreement report is in the guide.

And the file boundary is hardened: a crafted or corrupted SPSS, Stata, SAS, Excel, Parquet or Arrow file reads as an error message, never a crash — proved by continuous fuzzing and an 84-case battery of degenerate-data edges, each pinned to a number or a named error.

0.22.0 — 2026-08-26

DOI: 10.5281/zenodo.22112257

The catalog stays at 108 validated tools. This release deepens what they draw, export and print — and was checked against live services and captured screens before shipping.

Finish a chart without leaving the app. The chart editor gains axis titles, axis limits, reference lines, marker shapes and sizes, a gridline toggle, a figure height, a text-size slider that scales every label at once, and logarithmic axes labelled at the decades — a log axis is refused, naming the value, rather than silently dropping a zero. Every edit is written back into the chart's own syntax command, so a styled chart re-runs styled.

Then export it for publication. Any chart saves as SVG or PDF — real vector, with labels that stay selectable text — and PNG export gains a resolution chooser.

Underneath, accuracy. Normal and chi-square p-values were returned as exactly zero past about eight standard deviations and are now accurate far into the tail, on a dedicated erfc checked against a 50-digit reference. Matrix inversion is scale-invariant and the GLM families fit through an orthogonal decomposition.

Small things you will see: chart fills clear the 3:1 contrast bar in both themes, Variable View labels end in an ellipsis, scrolling dialogs fade at the edge, and charts repaint when you switch theme.

The assistant streams its replies, caches its prompt and shows the cached share; ask it about any result in your output log and it explains what you see, using no tools. Both backends verified live.

Earlier releases

Releases before 0.12.0 predate this page; their records are in the project's release documents. Versions from 0.20.0 on carry DOIs — the full list is on the How to Cite page.