For years I helped run a large public-service operation — the kind where people’s access to care depended on something showing up at a particular door at a particular time.
At that scale you stop learning lessons about the operation and start learning lessons about numbers. These are the ones that transferred to everything else I do.
The scariest failures report success
Every operational failure I lost sleep over had the same shape: a report that looked fine.
A metric that quietly read zero for months because a formula broke upstream and nothing complained. A performance figure computed slightly off-standard from the definition the regulator actually judges you on — close enough to look right, far enough to be wrong in the direction that mattered. A historical series that got restated upward with no note saying so, six months of numbers silently different than they had been the quarter before.
None of these announced themselves. That is the property that makes them dangerous. A loud failure gets fixed the same week, because a loud failure is annoying and somebody makes it stop. A quiet one compounds until it surfaces in the worst possible room, in front of the people least equipped to hear about it gently.
So the design principle is now stamped on every system I build, professional or personal: failures loud, successes silent. A process that breaks must be physically incapable of looking like a process that worked.
That sounds obvious written down. In practice it means arguing for the unglamorous version of everything. It means a job that succeeds should produce no notification at all, because a notification you receive daily is a notification you stop reading, and the day you stop reading it is the day the failures start arriving in the same channel you’ve trained yourself to ignore. I built the loud version first. Within a week I had taught myself not to look.
Pin every denominator
Most metric arguments aren’t about performance. They’re about arithmetic nobody agreed on.
A rate is a numerator over a denominator. If two parties compute the denominator differently, they can both be entirely right while briefing opposite stories to the same board. I have watched a denominator definition change quietly — not maliciously, not even carelessly, just a reasonable person making a reasonable choice about what to exclude — and flip an outcome that mattered.
So: pin them. Write down exactly what counts and what doesn’t. Write down the exclusions, and then write down why each exclusion exists, because in eighteen months nobody will remember and someone will propose removing one.
Compute rates as sums over sums, never as averages of daily rates. This is the single most common arithmetic error I have seen in operational reporting and it is invisible from the outside: averaging a rate across days weights a quiet Sunday the same as a brutal Monday, and the number it produces is not the number anyone thinks they’re reading.
And before you trust any automated figure, make it reproduce the officially reported one. Not approximately. Exactly. If your computation cannot reproduce the number of record, you do not understand the number of record — and a footnote explaining the variance is not a substitute for understanding, it’s a promissory note you will be asked to redeem at the worst time.
I held that line hard enough to be unpopular about it. Reports went out later than people wanted. I would do it again.
Your vendor’s metric is not your metric
The number a contractor quotes you from their workflow system and the number a regulator judges you on are almost never the same number.
Different windows. Different exclusions. Different denominators — see above. Neither party is lying, and this is the part that catches people, because everyone is looking for the bad actor and there usually isn’t one. There are just two honest systems answering two slightly different questions.
But if you brief one number as the other, eventually you will be lying, whatever your intent was. Reconcile them side by side. Name the basis of each, in writing, on the page where both appear. Never let the two series silently merge into one — separate namespaces, an explicit crosswalk, no exceptions.
The version of this that costs you is subtler than the obvious one. It isn’t the meeting where you cite the wrong figure. It’s the slow drift where a convenient number becomes the house number because it was easier to pull, and two years later nobody can tell you where it came from or what it excludes, and it’s on every board packet.
Freeze what you report
The first version of any reported period gets frozen.
If it later needs restating — and sometimes it honestly does; data arrives late, an error is found, a definition is corrected — the restatement gets recorded next to the original, never written over it. Both versions live. Both are dated. Anyone can see what changed and when.
History that can be silently rewritten isn’t history. It’s a liability with a timestamp problem.
This one has a personal test case. When I rebuilt the maintenance record on a tractor that had been running for eight years on my memory and one messy note, I hit exactly the same temptation: write down what “must have” happened as though it were fact. The rule I settled on was that work I couldn’t date got a question mark and an amber flag, forever, until a real dated entry replaced it. The log refuses to pretend it knows what it doesn’t. Same principle, smaller stakes, and it changed how the whole thing felt to use — when the dashboard says something is current, it is current, not probably fine.
Few numbers, tied to decisions
The last lesson is subtraction, and it took me longest.
A dashboard with sixty metrics is a dashboard nobody reads. It feels like rigor and it functions as noise. Worse, it’s a hiding place — with sixty numbers on a page, there is always one going the right direction to point at.
The discipline is picking the few that actually connect to a decision or a dollar, and then defending them like infrastructure: their definitions, their denominators, their history, their exclusions. Because that’s what they are. A metric nobody would act on is a metric nobody should be maintaining, and every one you keep out of sentiment is one more thing that can quietly break without anyone noticing.
I run a garden, a tractor, and an investment practice on the same five principles. The stakes are smaller. The math doesn’t care.
■