RootCause learns the causal structure over your ontology, at enterprise variable counts, in hours. Bring your own hypotheses and test them against the same data. Every edge is inspectable, attributable and versioned.
Promo spend
List price
Weatheroutside the business
Lead time3 lags
Web traffic
Stock cover
Unknown influencenot in your data
Carrier
Units sold2 lags
On-time rate
Support load
Repeat rateGross marginChurn rate
Three relationships here carry an open circle at both ends. The data settled that they are related and could not settle which way the arrow runs, so the model keeps them marked. One inferred common cause sits outside the measured columns and is labelled as such.
The methods are decades old and well validated. They are also expensive. Conventional causal discovery scales badly in the number of variables, which is why the technique has historically been reserved for a handful of carefully framed questions per year, run by specialists.
Enterprise data does not arrive as a handful of carefully framed questions. It arrives as thousands of variables across dozens of systems.
RootCause runs discovery at sub-quadratic scaling, which changes what is practical. Thousands of models can run at once. Each one rebuilds when the data behind it moves, so the graph in front of a planning meeting matches the business the meeting is about.
Built for parallel scale, with thousands of models kept current as their inputs change.
Both routes carry equal weight. Automated discovery covers the whole ontology. Hypothesis testing suits a team that already has a view of how the business works and wants the data to arbitrate it.
Promo spend
List price
Web traffic
Units sold2 lags
Gross margin
Unknown influenceSupport loadChurn rate
The association survives every conditioning set tested. Reversing the edge fits the noise structure worse, so the direction is settled and the edge enters the model.
Click any edge and see what put it there. The conditional independence tests that survived, the conditioning sets, the estimated strength and its interval, the rows involved, and the alternative structures that were considered and rejected. Where orientation could not be determined from data alone, the edge says so. Where a relationship depends on an assumption, the assumption is named on the edge.
Discovery uses SPARC-fast, our causal discovery engine and a Screening, Pruning, Agreement and Reconciliation Cascade. It is built around a conditional independence test with O(n log n) complexity in place of the quadratic tests conventional approaches rely on. That single substitution is most of the scaling result.
Orientation within Markov equivalence classes uses noise structure that idealized formulations discard. Real-world data propagates noise through causal pathways asymmetrically, and that asymmetry carries orientation information. Where the classical treatment says a direction is unidentifiable, real data frequently identifies it. Where it does not, the edge stays undetermined.
The metadata layer feeds this directly. Known seasonality, monotonic constraints and inferred structural properties enter as priors, so discovery spends its budget on the structure it has to learn.
The axes carry no values. Runtimes and skeleton accuracy are published with the benchmark configuration and the dataset list, because a chart without its configuration is worth nothing to anyone qualified to check it.
Three limits, stated here so you do not have to find them yourself. Each one is reported on the affected edge inside the product.
A lever that has never been moved has no observable effect to estimate. The model says so and leaves the effect unestimated.
RootCause flags where a common cause is likely and marks the affected edges underdetermined. No direction is assigned to them.
A relationship visible at weekly grain can vanish at monthly. The ontology records grain, so the model reasons about it explicitly and reports when a result is grain-sensitive.
The graph itself is a research artifact. What a planning meeting needs is the interventional answer, and that answer is computed on the graph.
The interesting result is usually the one that fails.