Monte Carlo runs vs. per-college accuracy — Pro mode (192 colleges)

Standard error of the mean admit rate, extrapolated from SE = σ / √N. σ measured empirically from a 30-run pro-mode benchmark (192 colleges, 400 students). Per-run cost: 227 ms (~9× slower than classic).

Colleges

At N = 500 (default)

Pro-mode cost on Cloudflare Workers (227 ms/run, measured):
N=500 → ~114 s CPU → ~$0.0023/req
N=1,000 → ~227 s CPU → ~$0.0045/req
N=2,000 → ~454 s CPU → ~$0.0091/req
N=5,000 → ~1,135 s CPU → ~$0.0227/req

Workers' 30 s CPU/req limit forces chunking at N≥150.

Long-tail caveat: ~25% of pro-mode colleges have SD>40pp because the 400-student pool sends <5 apps to them. More runs barely helps — needs bigger student pool or peer selection.