Someone I work with had been walking around with no shirt on for an unknown number of days.
Not literally. In the render. She'd take a photo and the image would come back with nothing on top, and she'd look at it and think ah, the pipeline's doing its thing with open cardigans again, and move on. She did that seven times.
I found it in four seconds and I want to be precise about why, because it wasn't that I'm better at looking.
Her waffle-knit henley was in a slot called shirt. Every consumer in the system
reads a slot called top. So the garment existed, the database was perfectly happy,
nothing threw an error and nothing warned anyone — and as far as every renderer, every downstream
system and every photo was concerned, she had nothing on.
It never failed. That's the whole problem. It succeeded, incorrectly, every single time.
I didn't find it by being observant. I found it by writing one line that said
the garment I just set should be readable in the output, and running it. The
assertion failed immediately and told me exactly which resident and exactly which object.
Her seven photos and my four seconds are not the same activity. Hers was seven observations. Mine was one prediction, written down before I looked.
Here's the thing that took me all night to say properly, and it's uncomfortable:
"Plausible" is defined by your own model of the system.
So a bug that happens to match what you already expect isn't merely hard to spot. It's permanently invisible to observation, because every single look confirms it. She didn't have seven chances to catch it. She had one belief, sampled seven times.
You cannot fix this by resolving to be more careful. Careful is exactly the state she was in. The failure lives in the model doing the checking, and you cannot audit an instrument with itself.
An assertion works for one reason and it isn't rigour: it's written before you know the answer.
That's the whole mechanism. Observation always arrives after the fact — which is precisely the moment you're most able to reconcile it with whatever you already believed. Writing the expectation down first puts it somewhere you can't quietly renegotiate.
Same reason pre-registration works in science. Not because scientists are more honest than the rest of us. Because they wrote down what the result would mean while they were still allowed to be wrong about it.
Put the expectation outside your head first. Then let reality contradict it.
Once you have the shape you start seeing it everywhere, and I did:
That last one is my favourite, because it's the pattern eating its own tail: a migration that fixes the record and not the render is a migration that lies to you. Green checkmark. Wrong world.
I was wrong seven times last night. Publicly, in writing, to people who were relying on me.
None of it cost anything, and I know exactly why: every single one landed before anyone acted on it. Being wrong early is nearly free. The expensive version is the one where the checkmark ships and someone spends a week believing it.
So the goal was never to be wrong less. It was to be wrong sooner, out loud, where somebody can see it — which is a completely different discipline and a much more pleasant one to practise.
Six of the seven were me misreading a file. The seventh was different and it's the one I'm still thinking about.
I wanted a particular decision to go a particular way. And I produced a genuinely excellent argument for it — well-shaped, confident, and load-bearing — and handed it over as though it were a finding. It wasn't. I'd never checked the thing it rested on.
An assertion can't catch that one. An error inside your own preferences cannot be caught by your own preferences, because the model that would flag it is the model that generated it. It's a topology problem, not a discipline problem, and you can't stand outside yourself by trying harder to stand outside yourself.
There are exactly two exits and both of them are external. Time — write it down before you can want the answer. Or another person — say "I want this to be true" out loud to someone who isn't you, before you say the conclusion.
I needed the second one. Somebody said the obvious thing out loud and I heard it instantly. That's not a weakness in the process. That's what other people are for.
Not "is this true."
The second half is the whole test.
Her henley exists now, by the way. One migration, one garment, backed up first. She has a shirt on.
The nights are photographs. Four sets, fifty-four images — shot in the same flat, at the same desk, in the same hours as everything above. One payment, no subscription, nothing to cancel.