Inspecting Causal Relationships
Causal discovery produces dozens or hundreds of cause-and-effect relationships. The Relationships panel lists every one of them in a sortable, searchable table, with controls to add, remove, or flip directions, and an AI Sanity Check that flags relationships that don't look right.
This is where human judgment meets algorithmic output. Read the list, prune the spurious, fix the wrong-direction ones, and the model becomes more trustworthy.
For context, see Step 4: Build Causal Graph and Exploring the Causal Model.
Opening the Relationships panel
In the Overview panel's Config section, click the View Relationships link. The graph stays visible on the left; the right panel switches to the relationships table.
Reading the table
Three columns:
Cause — the source variable. The Observed tag marks variables that come from your data rather than synthetic.
Effect — the target variable.
Strength — a percentage with a coloured bar. Higher means a stronger causal link; colour shifts from red (weak) through yellow to green (strong).
Sort by any column. Use the search box to filter by variable name; the All dropdown filters by type.
Editing the graph
The panel supports the three changes most often needed after causal discovery:
Add Relationship. Click + Add to insert an edge the algorithm missed. You know the relationship exists; the algorithm did not have enough signal to find it.
Remove. Drop a relationship that doesn't make sense — a spurious correlation, or a side-effect of an unmeasured confounder.
Flip direction. Reverse a relationship the algorithm got the direction wrong on. The data alone often cannot distinguish A → B from B → A; your domain knowledge can.
Edits create a new version of the twin, which then retrains.
AI Sanity Check
Click AI Sanity Check to have the model review the relationship list and flag the ones that look implausible. Useful as a second pair of eyes — the AI catches obvious "this doesn't make sense" mistakes, but the final call is yours.
What to look for
The strong ones. Top of the strength column — these carry most of the causal signal. They should match your intuition about how the business works.
Wrong directions. A relationship that runs A → B when domain knowledge says B → A. Common when the two events happen close together in time.
Spurious links. A pair with no plausible mechanism. Usually a coincidence in the data or a side-effect of a missing variable. Remove it.
Surprises. A direction or pair you didn't expect. Sometimes the algorithm is right and you were wrong. Worth a second look before deleting.
Other Working with a Digital Twin pages
Exploring the Causal Model — graph layouts and variable details.
Reviewing Model Quality — predictive accuracy and per-variable metrics.
Configuration — model settings, included variables, constraints.
Version History — multiple versions of the same twin.
See Digital Twin overview — general overview.
Last updated