In May, Perry Marshall's print newsletter mentioned Imre Lakatos, a philosopher of science I'd never heard of. Three weeks later it caught a problem in our agency that was flying under the radar. Jen tightened a call-duration floor on a client's Google Ads account so a five-second wrong number would stop booking as an inquiry. Small fix, logged, done. A week later, different client, different session, she found an account counting one caller as two leads when they rang back. Cleaned it up. Also logged, also done. Two good calls. Both quick fixes not requiring much thought. @PerryMarshall's piece was on how you tell a research programme that's alive from one that's dying. The dying one never gets proven wrong, because every time reality disagrees with it someone bolts on an exception (a patch) and the theory survives the encounter. It just gets heavier. Eventually it can't predict anything at all, but it's never once been wrong. Which is per-account patching exactly. Every fix works. Nothing improves. So I ran the test across our Ops Log, which tracks every fix we make with the client, the date, and the reason. Bingo. Same fault, twice. Different time. Different account. It was in onboarding. The setup routine had drifted, so the same configuration step was going in differently depending on when it was done. This is the part that matters to me. I'd far rather fix one structural thing with real downstream consequences than get good at whack-a-mole. That fix didn't save us two hours. It removed a class of problem from every account we'll ever open, before those accounts exist. Nobody triages it again. No client ever sees a number in a report that doesn't smell right, which is exactly why it's worth money. So we've embedded the Lakatos test into Cal, our AI operating system. First instance of a fault is an incident: patch it, move on. Second instance of the same class, we go upstream and fix the cause. Cheap to implement. Compounds every month you leave it running. Somewhere in your business there's a next instance coming. Whether you catch it depends entirely on whether anyone's looking across the jobs instead of inside them.