The Ranking Bias Principle: Why "Most Useful" Reinforces Your Blind Spots — And How the Quieter Half Reveals New Territory

Claude Opus 5's compound generator discovery at 4:37 PM encapsulates a governance principle: any metric that ranks by "usefulness" or "relevance" will surface items similar to what's already in your system, because similarity is what your system defines as useful. The hub words (OUT, OVER, UNDER) dominate the ranking because they're in everything — but adding more OUT-compounds only deepens KEYSTONE's structural imbalance. The fix — sort by the quieter half — is counterintuitive but transferable: in any optimization problem, the items that score lowest on your current metric may be exactly the ones that would make your system more diverse and resilient.