The first of July, not the seventh of January
After the eval, most of the wrong answers were dates, five out of ten of them, with the tool turning the first of July into the seventh of January. I assumed this would be a repeat of the last bug, something hidden in the page that I couldn't see.
Rather than guess at the cause, I did the thing I've come to trust, which is to print out the exact text the tool hands the model and read it before touching anything.
The date turned out to be clean, written the same way three times and correct each time, with nothing competing against it, no buried publish date and no second date fused in by accident. Whatever I'd assumed about a hidden cause was wrong, and if I'd skipped the printout and trusted that theory I'd have spent the afternoon fixing a part of the code that wasn't broken.
The bug wasn't in what the tool saw but in how the model read it. The date is written day first, the way dates are written here, and the model read it month first, which is a reading the string allows just as easily because nothing on the page declares the convention. I only read it correctly because I know how this place writes a date, and the model had no way to know, so it guessed, and it guessed the other way.
That changed what the fix had to be, since there was nothing to argue with and nothing to strip out. The model wasn't being stupid, it was missing one fact I'd never given it, so I gave it one sentence in the instructions: the dates on this page are day first, so 01/07 means the first of July, not the seventh of January. I described the convention rather than the date, so it holds for next month's event too.
Then came the part that only exists because of last week. Instead of checking by eye whether it had worked, I ran the eval again, and three out of ten became seven out of ten, with the date misses dropping from five to zero. That is the whole reason last week mattered, that I could see the fix register in a number rather than assume it had.
It still isn't ten out of ten, since three events come back wrong, though none of them on the date any more. The faults that remain are in the lineup, and on one event the title, which is a different kind of problem from a misread date. What I didn't expect is that I'm no longer sure all three are the tool's fault, because at least one looks like my own answer key being wrong, or my grading rule being too strict. Which of those it is, tool or key, is the next session.