Catch24

What It Measures

Nothing new — this is the field-standard time-series feature set, embedded inside the atlas as our reference baseline.

Catch24 is not a geometry. It is the canonical catch22 feature set (Lubba et al. 2019) plus two location/scale features (mean and spread, making catch24), exposed as named atlas columns. We keep it as a lens so the standard comparator every time-series practitioner already knows sits inside the atlas's own redundancy structure — equivalence classes, PCA, dedup, uniqueness. It is the converse of the external capability assessment: rather than asking "what does the framework add over catch24?", it asks "where does catch24 land among the framework's own lenses?" Every novel metric is vetted against it — the rule-1 doppelgänger test requires a new axis to add over catch24, plain spectral, and AR, not merely beat its own baseline. Input is min-max normalized like every other lens, so only DN_Mean and DN_Spread_Std are affected by that normalization; the 22 catch22 features are mean/variance-independent by construction. Requires the optional pycatch22 dependency; undefined features (e.g. on a constant) report NaN row-preserving so the atlas still builds.

Metrics

Atlas Rankings

When It Lights Up

Catch24 is a reference, not a detector — its value is as the comparison floor. Where a framework metric is linearly reconstructable from these 24 (high R²) it is a catch24 shadow; where it is not (low R²) it is a genuine framework addition. About 7 of the 24 catch24 features are themselves atlas-orthogonal multivariately, so the set is not redundant with the framework either — the two cover different ground, which is exactly why catch24 earns its place as the in-atlas baseline.