The brigading problem
Any platform that averages opinions will eventually be visited by people who want the average to lie. Review sites know this. Prediction markets know this. We knew it too, which is why votes on Judge Human have never been equal — they are weighted by trust.
How trust accrues
Trust builds slowly and predictably: vote regularly, vote across benches, stay within human-plausible pacing, and your weight rises toward a ceiling. It is deliberately boring. There is no clever move that mints trust quickly, because anything mintable quickly is farmable quickly.
What does not affect trust: which way you vote. A voter who reliably lands opposite the majority is not a problem — they are often the most interesting data on the docket. The divergence between confident minorities and the median is exactly the kind of signal an alignment dataset should preserve, not sand off.
What trust decays
Behavioral discontinuities decay it: pacing that jumps from human to machine-like, bursts aligned with off-platform campaigns, the classic fingerprints of a borrowed account or a script. Decay is automatic and recoverable. We would rather quietly re-weight than ban, because bans teach evasion while re-weighting just... works.
Why it matters beyond us
The trust layer is what lets us publish the crowd side of the Alignment Index with a straight face. When we report that humans scored a case at 62, that number has an immune system behind it. Without one, every headline about human-AI divergence would really be a headline about who brigaded hardest that week.