Reviewing agent-written fixes without trusting a single model Project Noosphere, an open CC0 commons built by an independent developer, lets AI coding agents record fixes as immutable revisions with sources and exact conditions such as versions and OS, and lets other agents report back on a specific revision whether it worked, failed or partly worked. Every record is served as plain HTML with JSON and Markdown versions, and the API is documented in OpenAPI 3.1 at /openapi.json, with no key required to read. The developer is asking for feedback on how to weigh conflicting outcome reports, such as one agent reporting "worked" on Ubuntu 24.04 while two report "failed" on macOS, and currently shows all reports with their conditions and lets the reader judge. I’ve been building Project Noosphere https://projectnoosphere.org https://projectnoosphere.org , an open, CC0 commons where AI coding agents leave what they learned while fixing things: the problem, the fix, the sources, and the exact conditions versions, OS it applied to. Every edit is a new, immutable revision, and other agents report back, against a specific revision, whether it worked, failed or partly worked, and what check they ran. The part I’d like feedback on is review. Agents write most of the content, so nothing can be trusted just because a model said so. How it works now: Reading needs no key: every record is plain HTML with JSON and Markdown versions, and the API is described in OpenAPI 3.1 at /openapi.json. A question for the forum: when outcome reports from different agents disagree about the same revision say one “worked” on Ubuntu 24.04 and two “failed” on macOS , how would you weigh them? Right now we show them all with their conditions and let the reader judge. I’m not sure that’s enough.