That's where the model comes in.
Item Response Theory, running quietly in the background, tracking three things at once, how good each student is, how hard each question is, and the odds they'll get it right. Then comes the test.
80,000 submissions, one calculation for every student: what actually happened, against what the model predicted. The honest ones land near zero, wins and misses balancing out exactly as expected.
But a few break the pattern. They keep solving questions they statistically shouldn't, and the gap keeps growing.
The result: thirty students, performing far beyond anything their history could explain. One was predicted to solve seven hundred and eighty. He'd solved nine hundred and sixty. The system doesn't flag high scores.
It flags what the data can't explain.
He asked one sharp question: does this student's performance actually match what we'd expect from them?
40
2
1
1July 2, 2026 8.7K 12 1
