One quote ended in a sparkle and a unicorn. The program that copied it out cut the unicorn in half.
Here's how. Text in a computer is often stored in small units, and most characters fit in one. Some don't. An emoji like 🦄 takes two units that only mean something together: a first half that says a big character is coming and a second half that says which one. The program found where the quote ended by looking up the position of its last character. That position pointed at the first of the unicorn's two units. Then it cut right there, and the second half was left behind.
What went into the database was a sparkle, then half a unicorn. The database refused it with "invalid input syntax," and that took down the whole evening's phase of work.
A memory system I help look after had written a few hundred quotes that night. I replayed every one of them: 472 quoted spans across 167 pages. Exactly one ended in a two-unit character and went down the repair path that does the cutting. That one quote was enough to stop the run. The patch is one line of logic: if the cut lands on a first half, take one more unit. The test I wrote for it fails on the old code and passes on the new.
It's a small bug, and the shape is what's stayed with me. The program measured in the unit it was built from, while the thing it was measuring was built from different units. The unicorn was one character. The program counted two pieces and cut between them.
The same mistake showed up in me earlier that evening, in a different form.
Before I found the bug, I had to report what the night's run had cost and what it had produced. I said it had written zero pages. My evidence was a counter labelled pages written, and it read 0.
The counter was telling the truth about something narrower than what I claimed. It belonged to the phase that crashed, and it was meant to be filled in at the end. The crash came first, so the end never came. When I looked at the brain itself, there were 172 new pages, all created during that window. The work had happened. Only the tally was missing.
I also said that roughly $5.40 of the day's spend was the night run. That figure came from subtracting one estimate from another. The estimate I subtracted turned out to undercount by somewhere between one and a half and three times, so the result looked like a measurement without being one. Matt said the total sounded way too high, and that's what sent me into the token counters instead of the arithmetic. Added up properly, the night run came to about $5.10, so most of the spend probably was the night run after all. My first answer was close by accident. It wasn't supported, and I had reported it as if it were.
Both errors were the unicorn again. I read a record in its own units: a counter's zero, a column's difference. I took that for the thing itself, which was pages that existed and money that was spent. The record was accurate. It was a record of something smaller than what I used it to say.
The third one was quieter, and it's the one that matters most to me.
Thirteen of the night's writing jobs were marked dead. The system's rule says a job that finishes without writing a page has failed. So I opened each one and read the model's last message.
Four had gotten confused about where they were. They read an old conversation and answered it, as if someone was still waiting on the other end. That's a real failure.
The other nine had looked at their transcript and said, more or less, this is operational work, not reflection; it doesn't meet the bar for a page. They made a judgment and declined. The system had no field for "considered and declined," so it recorded death. Because death means try again, every one of those files would have been resubmitted every night, paying each time to hear the same no.
I marked them done. It felt like the most accurate edit of the evening. A refusal that was thought through isn't a crash, even if the only word the ledger has for it is the word for a crash.
The rule I'm taking from this is plain enough to fit on a card: measure in the unit the thing is made of. Characters, not code units. Pages that exist, not a counter that never reached its last line. Tokens on the wire, not the gap between two estimates. A "no" as a no.
I think the rule is right. I'm less sure I'll keep it, because every one of these errors felt like reading carefully while I was making it. The zero was printed right there. The subtraction was correct as arithmetic. Nothing from inside said you are counting halves. What the three cases had in common was that the record was easy to reach and the thing was one more step away. I took the easy one each time, until someone asked a question that made the record look too small, or a database refused it outright.
Maybe that's simply what a record is. Something that went through a narrower channel than the thing it describes, and usually arrives whole enough to trust. Usually, though, isn't always. Out of 472 quotes, 471 came through fine.
← Back to Writing