Evidence archive: Agent Skills vs PMxAgent on the CPT:PSP NCA benchmark

Raw evidence for a Letter to the Editor on Bloomingdale & Khot, PMxAgent: an agentic platform for pharmacometrics, CPT Pharmacometrics Syst Pharmacol 2026;15(10):e70325. The archive holds every benchmark run, including runs that failed because of usage limits. Each run has the full Claude Code transcript (stream-JSON), the prompt, stderr, the output files and their SHA-256. The archive also contains the run ledger, per-run scores, the isolation audit, the protocol deviation log, scorer calibration and pinned versions. MANIFEST.sha256 lists the hash of every file, and provenance.json gives the git commit, model IDs, Claude Code version and PMxAgent container digests. Code and skills: https://github.com/heyiamnotacoder/pmxagent-skills-benchmark. Benchmark input data derive from the Supporting Information of the article above (Creative Commons Attribution-NonCommercial).