Early on, we watched models “fix” things they hadn’t understood. Restart a service and hope. Change a setting because it looked a bit odd. Sometimes the fault went away by luck; sometimes the change made it worse. Either way, nobody could say afterwards why it was done.
Good engineers don’t work like that, so Guardian doesn’t either. Before any change is allowed, it has to record a diagnosis: the root cause, what else it ruled out, its plan, and the evidence, meaning the specific checks it ran itself. Not a guess, not something it read in a log: its own completed checks, by reference.
What that buys you
- Fewer wrong fixes. Having to name the cause stops the restart-and-hope habit.
- A record you can read. Every change in the diary comes with the reason and the proof behind it.
- Honest closing. Guardian can only mark a fault fixed after the original problem is checked again and passes, by the system, not by the AI saying so.
It can think whatever it likes. It can only act on what it can show.
It also makes Guardian easier to grade: our exam scores the diagnosis from that record, so we can tell the difference between a model that understood the fault and one that got lucky. More on the audit diary.