Here's the thing about your bracket: it looks like everyone else's.
The Problem With Chalk
The national bracket average follows seeds more than 70% of the time in Round 1. Everyone takes the same two 12-over-5 upsets. Every Final Four has at least two 1-seeds. By the Sweet 16, most brackets are practically identical.
That's fine if you want to finish in the middle of your pool. If you want to win your pool, you need differentiation. One correctly called upset that nobody else in your pool picked is worth more than five correctly called 1-seed wins.
The Chaos Agent is optimized for pool-winning potential, not raw accuracy. It accepts more wrong picks in exchange for the right kind of right picks.
Upset Vulnerability Scoring
For every early-round matchup, the higher seed gets scored on five factors:
- Efficiency gap -- how much does the lower seed's KenPom rating close the gap? A 12-seed in the top 40 of adjusted efficiency gets flagged immediately
- Tempo mismatch -- if the higher seed needs pace and the lower seed plays slow, that's structural upset territory
- Experience differential -- a 5-seed with three freshman starters vs. a 12 with a senior backcourt is a setup
- Bracket position -- not all 5-12 matchups are equal; some regional draws and game slots produce more upsets historically
- Brand bias correction -- the model explicitly downweights brand-name programs when the data is ambiguous. Duke has lost in Round 1. Kansas has lost to 8-seeds. History doesn't care about tradition.
Vulnerability score above 65: take the upset. Above 80: high confidence.
The Math
The result is a bracket with roughly 40% more first- and second-round upsets than average. Most will be wrong. That's by design.
Expected value analysis of historical pool data shows high-upset strategies beat chalk in about 35% of head-to-head matchups -- but when they win, they win by a lot. High risk, high reward. Built for pools, not predictions.
Honest Limitations
In years where chalk runs deep through the Sweet 16, this model gets crushed. It also doesn't fully cascade upsets downstream -- a manual consistency pass prevents structural impossibilities, but the bracket carries more variance than a path-dependent model would.
Track the carnage on the dashboard.