Anomaly detection framework prioritizes causal consistency over temporal patterns

CAAD doesn't just detect when time series deviate from learned patterns; it flags when the causal dependencies between variables break down. The key insight is treating anomalies as structural failures rather than distributional ones, which means a system can look statistically normal while its internal logic is already failing.

This work sits alongside the constraint-steering paper from this batch (Partial Dependence) in a broader shift toward embedding domain structure into model learning rather than bolting it on afterward. Where that paper uses expert knowledge to guide what networks learn, CAAD assumes the domain structure (causal relationships) is already known and uses it as the detection signal itself. Both reject the idea that anomaly detection or model behavior should emerge purely from data; both assume mechanistic understanding is available and should be operationalized. The ImputeViz dashboard work also shares this ethos of making hidden reasoning transparent, though CAAD applies it to failure diagnosis rather than missing-data handling.

If CAAD outperforms standard autoencoders on industrial datasets where failures are known to involve causal decoupling (e.g., sensor networks where a single broken dependency cascades), but performs similarly on datasets where anomalies are purely distributional shifts, that confirms the method's value is conditional on problem structure. Watch whether industrial maintenance teams adopt this or treat it as a research artifact.