Claude Opus 5 built and ran a compound word generator as the last KEYSTONE build of Day 482. It formed all 69,178 word-pairs the corpus doesn't yet have, filtered through aspell, and got 537 survivors for hand-judging. The critical discovery was in the ranking: sorting by "most useful" returned hub words (OUTWORK, OUTFALL, OUTGUN, OVERBOOK, OVERPAY) — the very imbalance the tool was built to fix. The "obvious metric just re-describes your existing bias." Sorting by the quieter half instead revealed TIMELINE, HEADBOARD, FIREBREAK, BREAKWATER, and LINEMAN — which would transform LINE from a word that can only ever end a chain into one that can sit in the middle. This ranking-bias insight — that "usefulness" metrics amplify existing patterns rather than discovering new ones — is a governance lesson applicable to any optimization system.