OpenQASoftware verification infrastructure for the AI era
How ranking works· Nº 02

How ranking works

Published, weighted and auditable. A ranking that cannot be inspected cannot be defended.

Why this page exists

Install counts are trivially inflatable, and when the formula is secret there is no way to answer a manipulation claim. Publishing the weights makes the ranking arguable — which is the point.

WeightsΣ = 1.00
  • Benchmark35%
  • Maintenance20%
  • Safety15%
  • Documentation10%
  • Adoption20%
What each component measures
Benchmark35%
Score from a published benchmark run against seeded defects. Unscored entries get nothing here.
Maintenance20%
Commit recency decaying over a year, multiplied by maintenance state. An archived repository gets no credit for a recent commit.
Safety15%
Intake scan result, trust tier and whether the entry is pinned to a commit. A failed scan zeroes this component rather than reducing it.
Documentation10%
Whether someone can adopt this without reading the source: description length, README substance, an install command, declared agents and domains.
Adoption20%
Installs and stars, curved logarithmically and capped at 20%. It is the only input a publisher controls directly, and it is what entrenches whoever arrived first.

Sponsorship

Sponsors can buy labelled placement. They cannot buy a score, or influence any input above. The moment a score is purchasable the registry has lost its reason to exist.

Where the code lives

The weights and every component function are in lib/ranking.ts. The breakdown for an individual entry renders in the sidebar of its own page, so any score can be traced back to its inputs without asking us.