Index · reference
Three entity types, one score model each. Categories are filters on the same table — a board only exists when it needs its own formula.
Raw counts span five orders of magnitude. Each level metric is log-compressed and converted to a percentile within the ranked cohort, so eight million users beats a hundred and eighty thousand by roughly double on that factor — not by forty-four times.
Growth is measured as (now − previous) ÷ (previous + k), so a repo going from three stars to nine does not post a 200% gain. That rate is then smoothed with an exponentially weighted mean, so one bad scrape cannot reshuffle a board.
A 4.9-star rating should mean the same thing regardless of who else is on the board, so quality is never percentiled. It is shrunk toward the board mean by review count, which is why two five-star reviews cannot leapfrog four hundred. Below twelve reviews no public score is shown at all.
When a factor cannot be computed — no history yet, not enough reviews — its weight is redistributed across the factors that do have data, and the board says so on the page. Scoring a missing factor as zero would quietly compress every score and make the index look broken.
Renowned figures from major AI labs and technology companies are listed for context rather than ranked. Their standing comes from their work and their institutions, and ranking them beside independent creators would crowd out the people this index exists to surface. They are also held out of the cohort every percentile is measured against, so their scale does not distort anyone else's score.
Paid verification buys a badge, profile control and analytics. Placement stays algorithmic for everyone. Paid visibility exists only in clearly-labelled Featured slots that sit outside the numbered board.
Configured weights sit in the board definition. The effective weights below are what the index is actually using right now, after redistributing any factor that lacks data.