BETA — invite-only. Every call below is real, immutable, and scored on calibration — read the exact rules.

/the caliber score

The Caliber Score

Where the Brier score grades a single call, the Caliber Score is a caller's running reputation. It updates Elo-style on every resolved call.

new_rating = old_rating + K × (actual − expected) × difficulty
// K = 24 · difficulty = 1.0 until there's real data to calibrate it

That raw rating has a problem: a caller who's 3-for-3 and a caller with 300 resolved calls should not look equivalent just because both currently sit at 91. So the displayed score is shrunk toward the population mean until there's enough of a record to trust:

displayed_score = (n/(n+20)) × raw_score + (20/(n+20)) × population_mean
// n = resolved call count · 20 = shrinkage constant (m)

displayed raw 97 avg 72

A 3-for-3 caller with a raw 97 displays only ~75 — barely above the 72 population average. It takes ~20 resolved calls to reach the halfway point (85), and by 150 calls the display (94) is almost entirely their own record. This is why a hot new account doesn't rocket to #1.

◆Always show n next to the score — "87 · based on 12 calls," never just "87." A number without its sample size invites exactly the skepticism this platform exists to earn trust away from. This is a platform principle, not an implementation detail.