AgenticGenomics: pharmacogenomics benchmark dataset for "Executable skills localise LLM errors to input interpretation in agent-mediated pharmacogenomics" (five-condition factorial, real-genome and input-normalisation arms)

Benchmark dataset for Corpas et al. (2026), "Trustworthy agentic genomics through versioned skill libraries: deterministic, auditable pharmacogenomics across nine models" (under revision at Cell Genomics; CELL-GENOMICS-D-26-00551). Design. Nine frontier large language models x 110 CPIC Level A pharmacogenomic cases x three population contexts (European, admixed Latin American, East African) x five conditions x three replicates = 44,550 evaluations. The five conditions are: (1) free-prompted; (2) retrieval-augmented from the CPIC guideline corpus; (3) skill-reasoning (the model reasons over a versioned SKILL.md specification); (4) skill-execution (the specification's CPIC mapping is executed deterministically as code); and (5) an answer-supplied positive control. The three core conditions (1, 2, 5) contribute 26,730 evaluations; the two skill conditions (3, 4) contribute a further 17,820 (v3_armAB_fullgrid.json/.jsonl). Version 1.2.0 (revision). Completes the five-condition design by adding the 17,820 skill-arm evaluations that the manuscript describes but that were absent from the earlier three-arm deposit (v1.1.0); refreshes the archived analysis-code snapshot; and pins the reproducibility package to repository commit 3f482d4 (github.com), corrected to run on a case-sensitive filesystem and to load credentials from the environment. Includes the end-to-end executed-pipeline analysis (real-genome deterministic caller to executed skill to abstention). Provenance. All data derive from public CPIC Level A guidelines and PharmGKB annotations. Genotypes are synthetic, text-specified canonical cases. No patient data and no new sequencing data are included. File checksums are in CHECKSUMS.sha256; no credentials are present in the archive. Scoring fields per record: A1 (phenotype identification), A2 (drug-specific recommendation), A3 (lethal-class safety action). Version 1.3.0 (30 July 2026), revision deposit This version adds the raw data for every experiment run during revision, so that the deposit matches the revised manuscript. The manuscript was retitled during revision; earlier versions of this record carry the previous title. Matched five-condition factorial (v3_five_cell_live*.json): 13,199 evaluations on an identical harness across eight models. Rule extraction from guideline prose (v3_extracted_rules*.json): the experiment that refuted our earlier claim that extraction reaches a hand-authored table. GeT-RM external caller truth (v3_getrm_*.json, v3_caller_truth_eval.json): 527 (sample, gene) pairs with mechanism decomposition. Input normalisation, R1.1 (v3_normalisation_inputs.json, v3_input_normalisation_*.json): two arms. Arm 1 gives models only genotypes and a gene name; arm 2 additionally supplies the allele-definition table the deterministic caller uses. v3_input_normalisation_o3_retry.json contains a batch in which every call failed at the provider (HTTP 429); it is deposited rather than discarded, because those failures were briefly misrecorded as model abstentions and the record of that matters. Crossed random-effects model (v3_crossed_mixed_model.json). PROVENANCE.md states what the raw rows carry (per-call input tokens, output tokens and cost) and what they do not (wall-clock timestamps and provider response IDs, which were never captured and are not reconstructable). CHECKSUMS.sha256 for every file added in this version. Code, and the corrections log covering thirteen evaluation artefacts found in this project's own harness, are at tag cg-revision-2026-07-30.