Skip to main content
Every causal graph is a set of variables connected by edges. Add one with the + (Add variable) button in the toolbar; clicking a node opens its inspector, with three tabs: General, Review, and Evidence.

​
General

  • Name how the node reads on the canvas.
  • Description one line on what this variable measures.
  • Value type numeric, count, or yes/no.
  • Measurable? on when you have data for this variable. Turned off, the node is a concept: something you reason about but don’t measure.
  • Unit e.g. “packs/day”, “people”, “KRW”.
  • Value range e.g. [0,120], {yes, no}.
Modeling advice that still applies no matter how you build:
  • Prefer measurable variables. “Customer satisfaction” is vague; “NPS score” is measurable.
  • One variable per concept. Don’t merge two quantities into one node.
  • Name in the direction of measurement. refund_rate, not no_refund.

​
Review

Each node carries a review state:
  • Approved (live) retrieval and verification treat it as live. Demote with To proposal.
  • Proposed staged; approve it (or approve the batch with the toolbar’s Approve all button) to make it live.
Allow auto-edit controls whether LLM expansion or re-runs may overwrite this node. Lock it when you’re confident in its current form.

​
Analysis session badges

While a relationship analysis session is open with a cause and effect picked, the inspector shows a role badge for that node:
  • T (Cause) the treatment/cause node of the current session.
  • Y (Effect) the outcome node.
  • M (Bridge / mediator) the effect passes through this node.
  • C (Watch together / confounder) a hidden common cause of both cause and effect. Include it in the analysis.
  • K (Collider) a node where both the cause and effect converge. Filtering by it invents a relationship that isn’t there.
No badge shows when no analysis session is active.

​
Evidence

  • Origin hand-built, or auto-built from a refinery run (with origin and confidence for review).
  • Search this subtree creates a retrieval preset scoped to the chunks attached to this node and its descendants.
  • Evidence chunks when a pipeline ingests documents into this graph, the attach pass surfaces the source chunks here.

​
Deleting

Open the node’s ⋯ menu → Delete node.