AgenticGenomics: reproducibility data and code for Authored skills improve accuracy in agentic pharmacogenomics but do not establish end-to-end clinical validation

Third-revision reproducibility data and code for the text-formatted pharmacogenomics benchmark. The measured endpoints are phenotype concordance and recommendation concordance with authored references; the executable rules and reference answers share an authored source. These results do not establish end-to-end clinical validation.

Version 1.9.0. Preserves historical references, prompts and outputs while adding a separately versioned RYR1 correction, an eight-candidate discrepancy audit with cited guidelines, the full 110-case screen partition and an explicit threshold-proximity check. It includes denominator-corrected attempted-evaluation analyses, both wording-rule definitions, the 109-case exclusion, and the reproducible 96.31% zero-credit sensitivity. The unchanged frozen scorer's uninformative 0.5 fallback and resulting 96.76% software diagnostic are disclosed separately.

The ZIP contains the runnable public repository, all 9,557 historical rows registered by the reparser, the distinct canonical V4 Flash panel, opaque model inputs and separate evaluator references, prompt-identity proofs, source hashes, candidate responses, and regeneration instructions. Tables S10 and S11 are also available as CSV files. The manifest identifies source commit 99f6f4b4047bc662cae95df6237f67067dc5cd43 and hashes every packaged file. Confidential manuscript and correspondence files are excluded.

This is a new version under the existing concept DOI. The prior version 10.5281/zenodo.22141182 remains unchanged. The historical code tag agentic-pgx-benchmark-v2.10 identifies the previous release; the code for this update is inside the ZIP. One confirmed reference error among eight selected candidates is not a whole-benchmark clinical error-rate estimate, and no independent clinical adjudication of the complete benchmark is claimed.