Why marathon finish times spike at 3:30 and 4:00, and what that says about every threshold you've trusted
Dan Luu's piece on 'suspicious discontinuities' shows people warp their behavior around sharp cutoffs — in 9.8 million marathon times, in tax brackets, in exam scores. In the thread, one commenter called the opening story fabricated, then publicly retracted it.
Dan Luu's 2020 essay maps the spikes that appear wherever a hard threshold sits: 9.8M marathon finishes clump at round times, Polish exams bunch at the pass mark, drug charges pile up exactly at the sentencing cutoff.
The thread adds fresh cases — chess ratings, AWS gaming its own P90 latency target — and one self-correction: a commenter called the lead example bogus, realized it predated a policy change, and withdrew it. The marathon spikes turn out to be social, not just vanity: pace-runners with flags.
The pattern, with numbers
Dan Luu's "Suspicious Discontinuities," resurfacing six years after he wrote it, is about one idea: wherever a system draws a hard line, people bend their behavior to land on the favorable side, and you can see the bend in the data. The examples are specific. Across 9.8 million marathon finishes, times spike at half-hour marks and round numbers as runners push to beat a psychological target. Poland's language exit exams show an abnormal pile-up exactly at the passing score of 30 — teacher leniency — while the math exam, harder to fudge, stays smooth. After the 2010 Fair Sentencing Act, cocaine charges bunched at precisely the 280-gram threshold. His proposed fix is dull and correct: replace sharp cutoffs with slow phase-outs.
The thread does science to itself
The best moment in the thread is a commenter modeling the behavior the article is about — updating on evidence. He opened by calling the lead anecdote fake:
"The opening story is fabricated and/or bullshit … there was no limit on income for health insurance subsidies … [EDIT: Since the undated article is actually from Feb 2020, predating the expansion of federal health insurance subsidies, I withdraw everything I said below.]" — a commenter who corrected himself in-line
Others brought new instances Luu didn't have. A quantum-computing scientist noted the identical shape in chess ratings on Lichess — players straining to cross a multiple of 100 and refusing to drop below. And the cleanest engineering example: AWS targeting P50/P90 latencies created the same clumping, as engineers gamed the metric, nerfing the worst cases to squeak under the target fence post.
The correction the article needed
One reply genuinely improves on the source. Luu reads the marathon spikes as individual psychology; a runner explained they're largely social infrastructure:
"Many (most?) marathons have pace runners who run the course, hewing to each 30 minute and 15 minute finish. So there'll be a 3:30:00 pacer, a 3:45:00 pacer … super hard to miss as you're running. By the end of the race they'd acquired a small army of marathoners." — a runner explaining the pacer effect
That's the upgrade. The discontinuity is real, but the mechanism isn't only vanity at the finish line — it's that the race organizers literally provide a person with a flag to cluster around. Which sharpens Luu's own point: thresholds don't just tempt individuals to game them, institutions build the scaffolding that makes the gaming effortless.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
A novelist explains why he still writes books in a 1978 word processor. The thread says it's not nostalgia — it's defense
Robert J. Sawyer's case for WordStar is that the program obeys the writer instead of the reverse. The thread turned it into something sharper: a 1996 essay about software became an argument about reclaiming attention from everything that pings.