Debugging

The intern swore the code had worked yesterday. Priya believed him — that
was the trouble. He wasn't the type to lie about something so easily checked, and
he looked more confused than defensive, which in her experience was the tell of
someone telling the truth about something inexplicable.

She pulled up the git log on the shared monitor, scrolling back through the commit
history of the internal moderation tool that flagged customer complaints for follow-
up. There it was: a commit timestamped 3:14 AM, authored by Devon, the twenty-
three-year-old intern who'd been with the team six weeks and had, until this
morning, distinguished himself mainly by asking good questions in standup and
terrible ones in Slack.

"You committed this at three in the morning," she said, not quite a question.

"I couldn't sleep. I kept thinking about the false negative rate on the sentiment
classifier. I figured I'd just take a look. I didn't mean to actually change anything."

Priya pulled up the diff. It was small — deceptively small, the kind that made her
stomach drop, because small diffs either did nothing or did everything. Somewhere
in the classification logic, in a conditional that checked whether a flagged
complaint should be escalated to a human or auto-closed with a templated
response, Devon had changed a single equals sign to a double equals sign.

In that language, that was the difference between assignment and comparison. The
original line had been quietly setting a flag to true every time it ran, overriding
whatever the sentiment analysis had actually determined. Which meant that for as
long as the bug had existed — a quick scan put it at fourteen months — every
complaint that reached that line had been automatically marked as needing human
review, regardless of what the classifier upstream concluded.

Which explained a lot. It explained why the support team had grumbled for over a
year about how much of their queue felt unnecessary — complaints that were
clearly routine, clearly low priority, that the system kept insisting needed a human
look. Management had chalked it up to an overly cautious classifier. A whole
quarter's OKR had been built around "reducing unnecessary escalations." Nobody
had thought to check whether the model's output was even being used.
It wasn't. For fourteen months, the actual classification hadn't mattered at all.
Every complaint, happy or furious, urgent or trivial, got the exact same treatment.

"So I broke it," Devon said, watching her face for a verdict.

"You didn't break it. You fixed a fourteen-month-old bug at three in the morning
because you couldn't sleep." Priya leaned back and started laughing, the kind with
more exhaustion than humor in it. "The classifier's going to start actually filtering
things now. Half the team's queue is about to shrink by sixty percent."

"Is that good?"

"It's what the system was supposed to do the whole time." She thought it through
further. "But it also means every metric anyone's tracked for over a year has been
measuring the wrong thing."

Devon looked slightly ill. "Should I revert it?"

Priya thought about that seriously. Reverting was the safe answer — the one that
didn't require explaining to leadership why a fourteen-month-old assumption had
just evaporated, the one that let everyone keep their dashboards and quarterly plans
intact. It was also, she realized, the answer that meant deliberately putting the bug
back — choosing to keep telling the system to lie to itself because the truth was
inconvenient.

"No. We're not reverting it. But we're writing a very careful incident report, and
you're helping me write it, and we're going to be extremely clear this was a genuine
bug fix and not you unilaterally changing production behavior at three in the
morning without approval — which is also a conversation we're having, separately,
about process."

"That's fair."

"But Devon." She looked at him properly. "You found something a team of
engineers, a product manager, and an entire quarterly review missed for over a
year. Because you looked at a conditional and asked why it was there instead of
assuming someone smarter had already checked."

He didn't say anything, but he sat up a little straighter.

By the following week the queue really had dropped, and the support team sent a
message somewhere between confused and grateful, wanting to know what had
changed. Priya wrote back with the honest version: a bug had been quietly doing
everyone's job for over a year, badly, and now it wasn't. It took the team a month to
fully trust the new numbers.

Devon got three sentences in the next engineering newsletter, buried under
"notable fixes." He printed it out and stuck it to his monitor, where it stayed long
after he'd stopped being the intern and started being just another engineer who
occasionally looked too closely at conditionals nobody else had thought to
question.