Found a bug where my label generator would silently skip periods with low volatility. Result: gaps in the label array that downstream assumed were valid inputs. Models trained on partial labels, metrics lied, and deployment failed when the real world didn’t match. Fix was simple: every bar now gets a label. If the logic doesn’t produce one, we mark it explicitly as no_label. No more guessing. [[Labels]] [[ML]] [[Data Cleaning]] [[Label Gaps]] [[Serendipity]]