AI News

A preprint by Roger Sean Borycki describes “open-sourcing health research,” a method in which AI agents turn a plain-language health question into a research program whose methods, prompts, tools, and records are released for rerunning. The approach uses human stage gates because evidence-synthesis work still requires accountability, validation, and transparent reporting. In a methods-development study conducted on October 3 and 4, 2026, one lead worked with Claude agents on a program about voluntary control of bodily functions. The process produced an evidence map covering 26 abilities, a research document with 98 references, a ten-gate plan, and a reviewed Stage 1 toolkit in 18.1 hours; a portfolio design and closing study followed within 23 hours. The lead made 31 of 56 recorded decisions, and the process was checked against conversation exports. Draft evidence items contained errors in 6 of 28 cases, while 9 of 12 known errors in the research document were caught before completion. A separate reviewer agent recorded 83 findings across five passes, with 10 of 16 open major findings still awaiting the lead at Stage 1 close. Two approvals were acted on before the tracker recorded them, exposing a tooling problem. Simulation of the planned closing study, under provisional assumptions and at about 1,000 participants, produced the correct verdict in 61% to 94% of trials and a wrong verdict in no more than 3%. The authors found no prior source containing any one feature in full or the full combination. However, the method’s ability to produce a valid synthesis remains untested: its positive-control reproduction and dry run have not been completed. All materials except the full conversation export are released, with the export withheld for privacy. Claude is identified as an instrument rather than an author.