2026-09-02
Guidance is not a mechanism
We went looking for a better model and found four bugs in our own prompts instead. The cheapest engine we run then scored 8 of 8 on Juice Shop for 37 cents.
Blog
Engineering notes from building Ryvx
2026-08-28
The bugs that only showed up when it ran for real
The GitHub fix-PR bot had never run. It took four attempts on a real repository, and a CSP fix that passed its own checks while shipping zero JavaScript.
2026-08-11
Withdrawing our benchmark numbers
Why we pulled the DVWA and Juice Shop recall numbers we published on 2026-08-11, and what replaces them.