Why autonomous research systems need to learn from failure, not just write papers

Current autonomous research systems treat trial-and-error as invisible—weak pilot results become confident prose, failures disappear, and nothing changes next time. Sibyl-AutoResearch adds structured "harnesses" that capture both successes and failures, then routes those lessons into later planning, validation, and claim-writing. The team built SIBYL, a file-backed system that audits how these conversions actually happen, recovering eight documented cases where the system caught and corrected errors like duplicate results and unsupported statistics. Code is released.