We judge almost everything by its best day.

The best quarter. The best night. The best version of a person when things are going well. The pitch deck shows the good month. The portfolio gets judged on the year it did well.

We ask what something can do at full speed and treat that number as the truth. It is not the truth. It is the ceiling. And the ceiling is the least useful number you can have.

What actually decides whether something lasts is the floor. Not how good it gets. How bad it gets, how often and whether you can keep going while it is bad. This is my underlying thinking when building networks and it turns out to apply almost everywhere else.

There are two ways to keep a city talking to itself.

The first is the one you already use. Cell towers. A small number of very powerful points doing a huge amount of work. On a normal day it is excellent. Fast, wide coverage, nothing else comes close.

The second is a mesh. Many small, weak devices, each one talking to whichever neighbour it can reach. No centre. On a normal day it is worse in every way. Slower, patchier, harder to run.

Now take the normal day away. Cyclone. Flood. Power gone. A tower down and the road to it under water. The tower system does not slow down. It stops. It works, then it does not, with nothing in between. And the exact moment you need it most is the moment it is gone. Read any report written after a disaster and you will find the same line: communication failed early and nobody could organise anything after that. The mesh does not stop. It sometimes gets worse. However it finds a way around the missing pieces and keeps carrying something. And something is a completely different thing from nothing.

I have spent a lot of time on this problem. What I keep coming back to is that we have been optimising the wrong thing for decades. We built systems to be excellent on normal days, when their actual job is to survive the bad ones. Then we act surprised when they are missing on the day they were built for.

I run flight tracking hardware as part of a network spread across many locations. My single receiver is not reliable. It goes down. Power cuts, weather, antenna problems, my own mistakes.

If the network needed my device to work well, the network would be useless. It does not need that. It needs many devices, in many places, with overlapping coverage, so that any one failure does not matter. The system is dependable and not one part of it is. That is the trade almost nobody wants to make, because it feels like a downgrade. You are choosing many average things over one excellent thing. Every instinct says buy the better one.

But the excellent one can take the whole system down with it when it fails. The average ones cannot. Over a long enough period, the setup that never fully collapses wins, even though it loses every comparison you could run on a normal Tuesday.

Here is the part that applies to everything else.

Big swings in the market do not destroy returns because of the maths. The maths is fine. They destroy returns because they change how people behave, and behaviour is what actually carries out the plan. Nobody holds on through a fall they did not expect. Nobody makes good decisions in month four of not knowing.

The same cost shows up in any system with people in it and almost nobody counts it.

A team with an unpredictable boss produces less because everybody starts spending part of their attention guessing what mood he will be in. That cost never shows up anywhere. Not in headcount, not in a project plan. It just quietly reduces the amount of thinking left over for the actual work.

A person whose reaction you cannot guess does not cost you the argument. They cost you the energy you spend bracing for it.

None of this appears as a loss. It appears as a slightly worse version of every decision you make for years, which is far more expensive and much harder to trace.

The inherent instinct is to collect the strongest people you can find. The best engineer. The sharpest person in the room. Talent does matter. But talent is a measure of the ceiling, and ceilings are not what break.

The question I ask now is simple: what is this person like on a bad day?

Are they good? Are they the same? Can I guess what happens when things go wrong, when the money is late, when the launch fails, when I am the one who messed up? If I can, I can plan around it and build something on top of it. If I cannot, then it does not matter how good they are, because I will spend my planning energy on them instead of on the work.

This is true for co-founders, people you work with, whoever you go home to.

The uncomfortable part is that you are a part in somebody else's system and they are running the same test on you. Most self-improvement is ceiling work. Get better, get sharper, raise your maximum. Almost none of it is floor work, which is the far less exciting job of making your worst state less bad and less frequent.

Floor work gets no credit. Nobody is impressed that you were reachable, steady and the same person to talk to during a hard week. It does not look like anything. But it is the entire basis on which anyone decides whether they can build something with you. And it is worth more than any amount of ability that only shows up when things are going well.

Stop asking how good this gets.

Ask what happens when it breaks. How often it breaks. And whether the thing still works in some reduced form or just disappears.

Build the version with no total collapse setting, even if it is with mediocre nodes. Choose the people you can predict over the people who impress you. Fix your floor before you raise your ceiling.

The system that survives its worst day is the only one that gets more days.