ducksemantics 0.1.0 is the first release of the DuckDB-native semantic graph, lexical grounding, dense retrieval, exact late-interaction, structured judgment, and benchmark APIs for R.
The release review corrected OBO qualifier parsing, imported alternate identifiers, decoded escaped definitions, made graph writes transactional and idempotent, preserved indexes after graph projection and closure replacement, deduplicated lexical candidates, assigned unique identifiers to ambiguous candidates, enforced grounded structured judgments, isolated ColBERT blocks by their full identity, registered S7 methods at load time, and made embedding caches atomic, content-aware, resumable, and safe to refresh.
The complete HPO 2026-06-23 ontology was loaded and queried. The validated graph contains 19,836 active nodes, 50,029 aliases, 24,378 direct edges, no dangling qualified edge targets, and 202,740 materialized is_a closure rows. The full native model run persisted 19,836 256-dimensional EmbeddingGemma vectors and 182,884 128-dimensional LFM2.5-ColBERT token vectors. Stored DuckDB MaxSim agreed with Rbebelm::colbert_maxsim() to an absolute delta of 0.0000021 on the checked result.
The source package passes standard R CMD check with status OK, all 107 tinytest results pass, the pkgdown site builds cleanly, URL checks pass, and the release lint profile reports no findings. GitHub R-CMD-check passed on Linux, macOS, and Windows: https://github.com/sounkou-bioinfo/ducksemantics/actions/runs/29833840274. The --as-cran run has no errors or warnings and one expected incoming-feasibility note because optional Rbebelm and GitHub Remotes are distributed through the project R-universe rather than mainstream CRAN repositories.
The attached evidence manifest records source/model checksums, timings, dimensions, row counts, memory use, and check status. The two attached logs contain the complete graph-scale and model-scale runs.