Per-user weight
Each contributor's weight is a function of their difficulty-adjusted mean error across resolved estimates — the error on each past prediction is weighted by that estimate's frozen difficulty factor, so accuracy on hard, thinly-covered names counts for more than accuracy on easy ones. The weight is clamped to [1, 2]; a perfectly accurate veteran approaches 2× the influence of a neutral contributor, never more.
Minimum history & anti-Sybil cap
Weighting only activates once a user has at least 10 resolved estimates; below that everyone gets a neutral weight of 1, so the Smart figure equals the simple mean. Any single contributor's contribution to the weighted count is capped so one account — or a cluster of new accounts — cannot dominate the consensus.
Privacy threshold
The community consensus is only revealed when at least 5 independent contributors have forecast the period and their answers genuinely differ (non-degenerate variance). Below that, the figure is suppressed so no individual's prediction can be reverse-engineered. This gate lives in the service layer; the raw aggregate function never leaks sub-threshold values.
Related
Plain-English version: Smart Consensus explained. See it applied on the crowd-vs-Street board.