Drag the slider to step through ODE time t. Left: the anomaly-score landscape from this single t alone -- at small t, interpolated states are dominated by source noise, so the low-mismatch region is a broad blob covering the manifold's interior (bias, poor localization); at large t, it localizes tightly to the manifold but is visibly speckled (variance). Right: the running w(t)=t²-weighted combination of every step up to the current one -- by the final step this equals the paper's actual detector output, both sharp and smooth.

Running w(t)=t²-weighted integral of the mean per-t heatmap value, up to the current scrub position.