The bug I couldn't see by looking
In the last entry I said the tool ran and returned the fields, but whether they were the right fields was a different question. Here's where that question got interesting.
I pointed it at a real event page from work and checked the result. Everything looked fine except the date. It came back as a day in April. The event is in June.
My first instinct was the obvious one: the tool is wrong, the model misread the date, I need to fix the extraction. I nearly started rewriting the prompt then and there.
But the date on the page is perfectly clear. I can see it with my own eyes. So why would the model, looking at the same page, get something I find obvious so badly wrong?
That gap is the whole lesson, and it took me a while to understand it. The model and I are not looking at the same thing. Not even close. I see the finished page in a browser: the event date, big and clear, and nothing else date-shaped anywhere on it. But the page also carries a second date I never see, the day the listing was published, sitting in the HTML behind the page where only a machine reads it. It's not on the page. It's under it.
And the tool reads what's under it. Before the text reaches the model, my code pulls in the raw HTML and strips it down to plain words. That stripping drags the hidden publish date out into the open and drops it right next to the event date, with nothing left to tell them apart. So the model gets two dates where I only ever saw one.
So I did the thing I should have done first. I stopped guessing and printed out exactly what the tool was feeding the model. And there it was: two dates, side by side, one of which doesn't appear anywhere I'd ever looked. The model picked the wrong one. Reasonably, even. It chose the one written out in full, formal form over the one that looked like a fragment.
The fix, once I understood it, was not to argue with the model. It was to stop handing it the confusing input. The publish date always sits in the same place in the HTML, so I remove it before the model ever sees it, by its pattern, not by deleting that one specific date, so it works on next month's event too.
The lesson I'll keep: when a tool gets something obviously wrong, my instinct is to assume the tool is stupid. Usually it's working perfectly on input I can't see. And sometimes that's literal, the thing that broke it was a date sitting in the page's code that no visitor ever sees. The bug isn't where you're looking. It's in what the machine is actually being handed, which is almost never what you think you handed it.
I only found it because I stopped trusting the screen and printed the raw input. That's a habit now.