Cranary: evidence before release.
An R package screener that treats release readiness as an evidence-backed, reviewable decision—not an AI verdict or a single score.
Project status: early implementation / active research prototype. The source repository is currently private. Longitudinal package rankings are not established or externally validated. Cranary is independent of CRAN and cannot certify acceptance.
The problem
A package can pass local checks and still raise documentation, dependency, portability, or reviewer concerns. Equally, noisy rules can label valid package patterns as blockers. Developers need a defensible way to separate confirmed issues from uncertainty and suggestions.
How the decision guard works
Package metadata, source, tests, documentation, local checks, and (when available) external builder results are collected with provenance.
Known mechanical failures are classified using reproducible checks. Existing findings remain authoritative.
Local semantic analysis can flag mismatches between documentation and implementation, request review, or abstain. It cannot turn an unresolved issue into a pass.
The evidence and uncertainty are presented to a maintainer. Even a clean screen does not guarantee CRAN acceptance.
A package can be ready for submission review without being guaranteed acceptance.
Benchmark against real R packages
The goal is not to invent a universal package score, but to test whether findings track defensible evidence and improve when known problems are repaired.
Current CRAN versions
Compare rule findings with contemporary CRAN check evidence and peer packages, without assuming every accepted package is issue-free.
Archive-based comparisons
Pair past source versions with dated check logs and identify confirmed issue/fix transitions. Do scores and findings improve after an actual fix?
Adjudicated controls
Keep confirmed issue cases and defensible clean controls separate from the cases used to tune the rules. Measure false blockers, coverage, review burden, and false clears.
Broader software ecosystem
Extend to diverse CRAN fields, well-used GitHub packages, and Cranary's own source to stress-test whether checks generalize.
What this demonstrates
Cranary combines rule-based engineering, statistical validation, local semantic AI, and release governance. The primary skill is designing a system whose authority is earned from evidence and whose uncertainties remain visible.