Two layers of AI checked the work. Neither one did the math. What caught the error was a gut feeling about the shape of a number, and gut feelings don't scale. The fix fits in one file.
AI output quality silently degrades as your context window fills, and by the time you notice, the session is already cooked. The fix isn't a bigger context window — it's a single line in your status bar and a rule: stop at 60%.
The Ralph Loop doesn't care which direction you point it. When you flip it (or run Ralph in reverse), it doesn't build, it reads. And what it reads from an existing system is something most teams can spend months wishing they had discovered, a specification of what the code actually does, extracted from the code itself.
