Every result below is a running experiment next to its source: the claim, the equation, the conditions, our solver, the reference values with their page or table, the difference and a verdict against a tolerance fixed before the first comparison. Where the reference could not be read, or the result misses, it stays flagged.
How it was run
One Claude Opus 5.5 session as lead, with at most one helper agent at a time, in four roles. The reader opened the sources and wrote down values with their table or page and whether it read them itself or found them quoted. The builder wrote the solvers, fixed every tolerance, then ran each experiment. The reviewer recomputed every number with its own scripts and audited the labels; its changes split the consistency checks out of the matched count. The interface role built this page. Every step was logged with a timestamp, and the ring at the top replays that log line by line: .