- Sources: arXiv 2608.12440
- Summary: The reported task was making a streaming generation survive the closing of its UI panel and reattach to the same live stream, across 189 files of a 717,725-line TypeScript application, 288 files and 34,770 insertions across both commits, in three days at USD 2,430. The stated protocol is a specification written by the agent, 14 refinement cycles auditing the specification against the source, atomic implementation, a compile and test feedback loop, then 17 verification cycles auditing the code against the frozen specification, with 201 defects corrected across 31 audit passes before any human ran the program. There was no pre-existing test oracle and no human code review, and the convergence criterion was two consecutive verification passes returning zero findings from the same agent auditing its own output. This is n=1 and self-reported by the author, with the evidence published as more than 1,500 pages of French session logs. It is carried for the protocol, not as a measured success rate.
- Why it matters: The audit-cycle protocol is the transferable part, but with the agent auditing its own output against no test oracle the reported outcome measures self-agreement rather than correctness.
send feedback on this story