What I was looking for
Drawdown bugs in quant libraries, checked against hand-computed values on short series.
What I found
Code that builds the running peak only from observed values skips the start, so a loss on the first bar never counts as a drawdown. The series looks safest exactly when it starts badly.
On a long history a later, deeper drawdown hides the error. On short windows it decides things: cross-validation folds, walk-forward periods, the first days of a live record. I found it in five libraries in one scan; in skfolio a −10% then +5% series reported a max drawdown of 0 instead of 0.10, and its own optimizer disagreed with its own measure. The fix shipped in v1.4.1.
The move
On any metric with a running state, I test the first bar on purpose: a loss on day one is the case that breaks it.