AgenticGenomics: pharmacogenomics benchmark dataset for "Trustworthy agentic genomics through versioned skill libraries"

Controlled three-arm comparison of nine frontier large language models on pharmacogenomic interpretation: Claude Opus 4, Claude Sonnet 4, GPT-5.2, GPT-4.1, o3, o4-mini, Gemini 2.5 Flash, DeepSeek V3 and Mistral Large 2, evaluated on 110 CPIC Level A cases across three population contexts (European, admixed Latin American, East African) under three conditions (free-prompted; retrieval-augmented from the CPIC guideline corpus; constrained by a versioned SKILL.md-format specification executed as a contract), three replicates per cell, 26,730 evaluations.

The archive contains raw model responses, baseline and clinical-equivalence rescorer outputs, the locked merged three-arm dataset (26,730 rows), the forward and reverse adversarial scrambled-specification experiments, the (gene, drug)-keyed chunking control, the drug-recommendation regression classification, and figure-input tables. Records are scored on phenotype identification (A1), drug recommendation (A2) and lethal-class safety action (A3).