The 1-in-6 number, and why I stopped trying to make it zero

I've mentioned before that I override FounderFlow's "Needs Review" flag about 1 in 6 times. Someone asked me last week why I haven't tried to drive that number to zero, and it's a fair question, so here's the honest answer.

Early on I treated every override as a bug to fix. Then I actually started logging what I was overriding and why. Most of them weren't the model being wrong, they were the model being right about the data and wrong about the context, a cafe supplier invoice that looked like an anomaly but was just a seasonal order, a "dormant" lead that was actually just someone on vacation. The kind of thing that isn't in any dataset, it's just in my head from running the business.

That's when I realized 1 in 6 isn't a bug rate, it's the size of the gap between "what the data shows" and "what I know." Chasing it to zero would mean the system pretending to know things it can't, which is exactly the failure mode I built the confidence grading to avoid in the first place.

So now I track it differently: not "how do I make this smaller" but "is this the same 1 in 6 kind of miss, or a new kind." New kinds get fixed. The familiar kind just gets logged, because that's the part that's supposed to stay human.

Curious if anyone else building AI tools has landed on a number like this, one you stopped trying to optimize away because it turned out to be a feature of the boundary, not a defect in the system.