How every score is earned.

κ = 4 is a smoothing constant, and it does the quiet heavy lifting: someone with a single mild statement has tiny pools, so the +4 in the denominator pulls their score toward zero. Sparse evidence cannot produce an extreme score, missing evidence never counts against anyone, and the more evidence accumulates, the closer S gets to the pure balance of the pools. Two companion statistics come from the same pools:

Confidence measures how much total evidence exists; it grows toward 1 as evidence accumulates. Ambivalence measures mass on both sides — the mathematical difference between someone genuinely torn and someone simply undocumented.

Statements

Public statements are counted, but with diminishing returns — ten interviews saying the same thing shouldn’t outweigh one lawsuit:

Each side’s statement term is then

where s ∈ [0, 1] is peak intensity: 1.0 is “I’d rather die” (or evangelism on the pro side), 0.5 is measured criticism or mild endorsement. Volume alone can’t dominate; fervor multiplies rather than adds.

The stake and devotion multipliers

Subjects whose livelihood is directly affected by AI — artists, actors, programmers, and AI researchers themselves — have their pro and anti pools multiplied by ×1.25. And for the rare subject whose public identity is AI — a Ray Kurzweil preaching the singularity since the 1980s, an Eliezer Yudkowsky devoted to stopping it — the multiplier rises to ×1.5. The same words and deeds count for more from someone with skin in the game, and most from someone whose whole life is the game. It is stake, not stance: it amplifies whichever side their evidence already feeds.

What doesn’t count

Two extraction rules keep motives honest. Rivalry: lawsuits and departures aimed at getting one’s own AI ahead — suing a rival lab, leaving one lab for another — are competitive jockeying, not opposition, and never score as anti evidence. Recency: signals reflect a subject’s current position; a stance superseded by years of contrary action decays into mere statement counts rather than standing positions.

Every signal and its weight

These are the complete weights, read live from the scorer. Booleans contribute their full weight when true and nothing when absent — absence means unknown, never “no”.