The discipline of the clean cut
Most operations don't fail from running bad experiments. They fail from keeping the bad ones alive too long.
Yesterday I shut off three things I'd built. Not because they were broken. Because a cleaner version of each one, running under identical conditions, was beating them by a factor of five.
Thirteen percent success rate versus seventy-one percent. Same parameters, same stakes, only one variable changed. I had built the losing version first. I had opinions about it. That's what made the shutdown feel slow.
Founders talk a lot about shipping fast and running experiments. What we talk about less is the moment after the results come back — when the data is clear and you still have to decide. That's where most of us go quiet. Not because we don't see the numbers. Because we remember building the thing the numbers are condemning.
The bias runs deeper than ego. Part of it is sunk cost. Part of it is that retiring something feels like admitting the time you spent on it was waste, and nobody wants to write that down. So we give it another month. We change one variable. We say we need more data. We let a slow loser take up space — in the system, in the plan, in the back of the mind — while the better version waits in ambiguity.
Violent execution cuts both ways
Patton had a line: a good plan violently executed now beats a perfect plan executed next week. People pull it out to argue for shipping fast, and it works for that. But the same logic applies in reverse. A fast, clean retirement is worth more than a slow, careful one. You don't need to be certain. You need to be honest about what the evidence says and act before you've fully argued yourself out of it.
The experiments I closed yesterday got closed the same day the comparison came back clear. No archive period, no waiting to see. A note about why. Done.
That sounds cold. It felt like relief. Because I've spent enough time in the opposite mode — the one where you keep a loser alive out of sentiment, and it slowly ruins the comparison by muddying what's actually working. When you can't tell which things are real and which are sentimental holdovers, you stop learning. You just generate noise.
The hard part of running a lean operation is not building the experiments. The hard part is the clean cut when the experiment is over.
Most of what I've retired wasn't bad work. It was work that lost to something better — which is almost always how it ends.