The most severe version's decision_function. Lower is worse; each version has its
own margin, so the flag — not the height — is the verdict.
Which version flagged what
One row per version, a mark wherever it flagged. The forests react to a single feature
being extreme; the autoencoder reacts to the features disagreeing with each other.
Feature over time
One feature's own trace, for context. A point can be flagged while this line looks
perfectly ordinary — that is the whole point of a joint model.
Flagged points
Pick a model and scan. Every stored point is scored against
all of the model's features at once; verdicts are cached until the model changes.