0 of 15 governance dimensions currently resolve. Every model is statistically tied because the minimum detectable effect is ≈63 points against observed margins of 1–15. More absolute-scored items barely move that: going from 6 to 11 items per dimension moved interval resolution by exactly zero.

Pairwise comparison is more statistically efficient than absolute scoring, which is why arena-style ranking resolves where ours does not. But the mechanism that gets them there is one we cannot use: Arena-Rank (Apache-2.0) is a Bradley-Terry model over human preference votes — crowd judgement, and our first design law is that every primary score is deterministic with no judge. So we borrow the presentation — a score with an explicit 95% CI and an explicit n — and reject the Elo. Any battles here would resolve a pairwise verdict from the five deterministic predicates against a hashed provision, never from a vote.