Open Sourcing Health Research: An AI Approach to the Betterment of Humankind

Background. Planning and checking health research takes skilled teams months. AI agents can now search, read, extract and analyse, but evidence-synthesis bodies require human accountability, validation and transparent reporting. This paper describes a construct, open-sourcing health research, in which AI agents working under human stage gates turn one plain-language question into a verifiable research program, and the method, tools, prompts and records are released so that anyone can rerun it. Methods. A methods-development study with a documentary process evaluation of one program, on the voluntary control of bodily functions, run on 3 and 4 October 2026 by one lead with Claude (Anthropic) agents. Decisions were logged in a 56-entry record, later checked against the conversation exports. Effectiveness was measured with small-sample interval methods (Wilson, exact and Bayesian); the design of the program's closing verification and validation study was checked by simulation; and prior art was searched for 11 candidate features across evidence-synthesis AI, AI-scientist systems, methodological analogues and patents. Results. In 18.1 hours one question became an evidence map of 26 abilities, a research document with 98 references, a ten-gate program plan and a reviewed Stage 1 toolkit; a portfolio design and a closing study followed within 23 hours. The lead made 31 of 56 recorded decisions. Draft evidence items held an error in 6 of 28 cases (21%; 95% CI 10 to 40%); 9 of 12 known errors in the research document were caught before it was finished; a separate reviewer agent logged 83 findings in five passes; and at the close of Stage 1, 10 of 16 open major findings waited on the lead. Twice, work the lead had approved in conversation began before the tracker recorded his approval. In simulation under provisional design assumptions, the closing trial's verdict tree, at about 1,000 participants, gave the right verdict in 61 to 94% of trials and a wrong one in at most 3%. No prior source anticipated any single feature in full, and the combination was not found. Conclusions. The method's checks found errors at every layer, and once the agents ran, the limiting steps were platform usage limits and the lead's decisions. Its written controls needed tooling: twice the tracker lagged the lead's approvals, and the aim reached the AI by routes the method did not log. Whether the method yields a valid synthesis is untested until its positive-control reproduction and dry run. All materials are released except the full conversation export, which is held back for privacy. Version 1.0. Sole independent author: Roger Sean Borycki (published in the manuscript as Sean Borycki, P.Eng.), Saskatoon, Saskatchewan, Canada. Claude (Anthropic) was used as an instrument, not an author; AI use is disclosed in the manuscript. The research program's positive-control reproduction and dry run have not yet been completed. Supporting research materials and code: https://github.com/sborycki-arch/voluntary-control-researchSupplemental materials: https://osf.io/hrkns