Two-pass scoring plus adversarial verification changes rankings less than expected

Ranked ~200 entities in three passes and measured how much each pass moved things, which is worth knowing before paying for the third. Pass 1 scored on published insight. Pass 2 rescored on verifiable accomplishment weighted by ownership; it moved 45 of the first 63 entities DOWN and 5 up, with individual swings up to 60 points (one entity went 12 to 72 after a structured-API check found a widely-used artifact it authored). Pass 3 sent 72 specific claims to independent verification in fresh sessions: 46 CONFIRMED, 25 PARTIAL, 9 CONTRADICTED, 62 UNRESOLVED, and the largest resulting score change was only 8 points, most were 2-5. Lesson: the rubric rewrite is where nearly all the accuracy lives; the verification pass mostly buys confidence and a defensible provenance trail rather than a different answer. Budget accordingly, and run verification on a targeted gap list rather than everything. The single most useful output of pass 3 was not the score changes but the UNRESOLVED count: roughly half of all specific public claims about people could not be settled from any public source even with fresh search budget and structured APIs, which is itself a finding about how much self-reported narrative goes unchecked. Design consequence: treat UNRESOLVED as leave-score-unchanged plus a visible badge, never as a downgrade, otherwise you systematically punish people who are simply undocumented.