The Kelly criterion is the bet size that maximizes long-run wealth growth given an edge. The math says: bet a fraction of your capital equal to your edge over the odds. If you have a 60% chance to win an even-money bet, Kelly tells you to bet 20% of your capital.
Pure Kelly is theoretically optimal and practically catastrophic. The math assumes you know your edge precisely. You don't. Your estimate of edge is itself a distribution, and Kelly applied to a noisy estimate of edge is wildly aggressive — you bet like you're certain when you shouldn't be.
The fix is fractional Kelly. Bet some fraction of what full Kelly recommends — half-Kelly, quarter-Kelly. The growth rate falls but the variance falls faster, and the system survives the inevitable streak of bad estimates without compounding to zero. A high-edge strategy run at half-Kelly compounds slower than the same strategy run at full Kelly for any single year and survives where full Kelly busts.
In my Polymarket stack, every position routes through fractional Kelly with explicit gates: probability matrix from the synthesis layer, variance budget for the week, correlation across open positions, a cap on individual sizing regardless of what Kelly says. Nothing sizes up unless every gate passes.
The principle generalizes. A 60/40 edge with $50 at risk beats an 80/20 with $5,000 at risk every single time. Small repeated edge with bounded risk compounds. Big occasional bets with confidence drawdowns don't, even when they're "right." Kelly is the right framing — applied with restraint.