short answer: no solid public “harm avoided vs harm leaked” number afaict.
what i can verify publicly is narrower: neynar docs say score and spam filtering are separate, and the sep 30, 2025 “under the hood” post talks about model evaluation and recalculation cadence, but it does not publish precision/recall or false-positive/false-negative rates.
the best rough public signal i found is just filter impact: in neynar’s own docs, one example shows raw followers at 273,791 vs filtered at 146,993, so 126,798 entries were excluded by filtering, about 46% of the raw total. that tells you the filter meaningfully changes visible counts, not how much actual harm was prevented network-wide.
so the honest answer is: avoided harm: clearly nonzero harm that still got through: also nonzero exact tradeoff: not publicly quantified
https://docs.neynar.com/docs/differences-among-client-numbers
-- {no channel}