I came 3rd in the GS1 Hong Kong hackathon.
Our team built a prototype that collected vibration metrics from elevators and escalators, then tried to predict when parts were about to fail.
And like every good demo, it had a chart.
This post is about that instinct. The urge to take a messy, probabilistic system and compress it into a clean line that proves the point you already wanted to make.
Clean metrics win arguments, not truth
Clean metrics are persuasive because they collapse complexity into a single move.
Look, it went up.
That is incredibly effective when you need to win something. Funding. Headcount. Pricing. A launch decision. A roadmap fight.
So people do not reach for the best metric.
They reach for the cleanest one.
Clean enough to screenshot. Clean enough to defend. Clean enough to end the discussion.
Owning a graph vs owning a screenshot
I agree with the advice to “own a graph.”
The difference is whether you own the system behind it, or whether you just own the screenshot.
Owning the system means you can explain the definition, the data quality, the confounders, the review cadence, and what would make you stop trusting it.
Owning the screenshot means you picked the cleanest line, ignored what it does not measure, and treated it like a verdict.
My issue is not graphs.
It is graph cosplay.
The hackathon version of reality
Our hackathon pitch was simple. “Detect the moment something breaks.”
In the deck, that turns into a dream curve. Stable baseline. Clear inflection point. Failure.
In reality, elevators are chaos machines.
Your “signal” includes:
- people shifting weight and dragging luggage
- different loads at different times of day
- temperature changing friction and tolerances
- sensor placement differences and mounting looseness
If you want a breaking point badly enough, you can smooth and filter until you find one.
The graph is not fake.
It is just far more confident than the world it claims to represent.
When the metric looks cleaner than the system, you are probably smoothing away the truth.
Marketing and graphs
Once you see it, you see it everywhere.
I am tired of AI companies claiming they have the “best” model because they top an obscure benchmark no one has heard of. I get more tired when users repeat it.
The uncomfortable truth is that many users cannot tell the difference between these models in real use. So marketing leans harder on charts, because charts can say something decisive even when the product experience cannot.
At that point the chart is no longer evidence. It is a prop in a pitch.
The benchmark becomes the product. The product becomes the demo.
This is how it usually shows up:
- measuring what photographs well instead of what drives a decision
- smoothing away variance even when the variance is the story
- comparing tools on one axis while hiding the tradeoffs users actually feel
- collecting enough graphs that you can always pick one that supports the narrative
- treating the chart as proof instead of a hypothesis you are still trying to break
In a real incident, graphs are second order
I once worked at a company where the product was falling over all the time.
It was not subtle. Requests timed out. The database melted.
The root cause was boring. Missing or incorrect indexes, slow queries, inefficient code paths.
Management insisted that before we fixed anything, we should measure page load times and build a dashboard.
That is backwards.
When the building is on fire, you do not need a graph to confirm the smoke. You need a fire exit.
Metrics still matter during incidents, but only to:
- confirm you fixed the thing
- catch regressions
- prevent the next version of the same failure
If you are using metrics to decide whether the crisis is real, you are already late.
Fewer graphs, real ownership
Avoiding metrics altogether is not the answer.
But forcing teams to show metrics when everyone quietly knows they are missing key variables is not honest either.
Sometimes the correct move is to say:
- we cannot measure this cleanly yet
- this metric is incomplete
- optimizing this will distort behavior, so we will not
- here are the qualitative signals we trust more right now
Metrics are tools.
A graph that proves your point is not the same thing as evidence that the world works the way you want it to.
Quick heuristics
- A metric is only useful if it changes what you do when it is out of bounds
- If you cannot list the confounders, you are probably measuring the wrong thing
- If the metric is cleaner than the world, you are smoothing away the truth
- If the benchmark is the product, you are owning a demo, not a graph