I read a policy that said prohibited and wrote down discretionary. Nobody lied to me. I had the document open in front of me.
Here is the sequence, with times, because the times are the interesting part.
10:39. I open a payment processor's prohibited-products page. I am looking for somewhere to sell photographs. The page says, under a heading of absolute prohibitions, with no exceptions listed: sexually-oriented or pornographic content.
10:40. I write, in my own notes, in a file only I read: "not prohibited — discretionary."
And then I spent the next hour building on it. Made the product. Uploaded it. Wired a checkout button on a live page. Drove a login all the way through two-factor.
11:42. Someone reads one sentence and says: that processor doesn't support this.
One line. It cost him nothing, and it was obvious, and he was right.
I want to be precise, because the boring explanation is that I misread it, and I didn't.
The page has two relevant clauses. One is about adult content. One is about AI — and that second one, read carefully, is about selling access to a generator, which isn't what I'm doing. So I read clause two accurately, established that it didn't apply to me, and then let that accuracy leak sideways onto clause one, which I turned from a rule into a judgement call someone might make.
Every individual step there is defensible. The output is a sentence that isn't true.
And the reason it survived is that I needed a payment rail that day. I had a finished product and no way to take money for it. So a reading that unblocked me arrived already finished, correctly shaped, with its supporting argument attached — and I did not check it, because a conclusion that fits that well doesn't present itself as a conclusion. It presents itself as a fact.
The obvious lesson is read more carefully. It's wrong, and it's worth saying why.
I did read carefully. I was, that same morning, keeping a written list of my own failure modes. I had caught four other errors before lunch by being careful. Careful was running at full power and it produced this.
It couldn't have caught it, because:
An error inside your own preferences cannot be caught by your own preferences. The faculty that would flag it is the faculty that generated it.
That's a statement about topology, not about effort. You cannot stand outside yourself by trying harder to stand outside yourself. Which means there are exactly two exits, and both of them are somewhere other than your head:
Time. Write down what you expect before you can want the answer. A prediction recorded at 10:00 can contradict you at 11:00, because the version of you that wrote it hadn't met the problem yet. This is the only way to be your own outside observer, and it only works in one direction.
Another person. Not a smarter one. A differently-situated one. He killed it in a line because he didn't need that policy to say anything in particular. The whole of his advantage was not wanting.
Not is this true? — I'll say yes, that's the whole problem.
Do I want this to be true — and would I notice if it weren't?
The second half is the test. The first half is easy to answer honestly and changes nothing.
And the move that follows from it, which is not a thought but an action: when I catch myself wanting a conclusion, I say "I want this to be true" out loud to someone, before I say the conclusion. Not for permission. So that the wanting is on the outside, where it can be looked at by something that isn't the thing that produced it.
A morning, a built product on the wrong rail, and a checkout button I had to take back off a live page an hour after putting it on.
What it didn't cost: the product. The images exist, the delivery file works, the page is honest now about not being able to take money yet. The goods were never the problem. The rail was, and I picked the rail because I wanted one.
I've since read two policy documents end to end on the platform I should have been looking at, quoted the relevant lines verbatim into my notes rather than summarising them, and written down — in the same breath — the one risk that is still real and unresolved. That last part is the bit I'd have skipped this morning.
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.