nx_analyst_causal.nx

nx_analyst_causal.nx -- F1005: the CAUSAL layer, and mostly the honesty about not having one. WHY THIS EXISTS. F1004 gave the analyst partial correlation and wired ai_confound_scan into the driver claim -- it controls for whichever candidate WEAKENS the relationship most. That is a defect, and this organ is the fix. "Control for everything, especially whatever moves the number" is a known statistical error, because a third variable's ROLE decides whether controlling for it is right, useless, or actively harmful, and no amount of correlation arithmetic can recover that role from the numbers: CONFOUNDER (z -> x, z -> y) : you MUST adjust. Not adjusting leaves a spurious association. MEDIATOR (x -> m -> y) : you must NOT adjust. Adjusting removes the very effect being measured. COLLIDER (x -> c <- y) : you must NOT adjust. Adjusting MANUFACTURES an association from nothing -- the gate demonstrates two independent columns acquiring a strong negative partial correlation purely by conditioning on their sum. All three look identical to a correlation scan. The only sound source of a role is a DECLARATION from whoever knows the domain, so that is what this organ takes -- and with no declaration it returns ASSOCIATION_ONLY and refuses to dress an argmax as a cause. SCOPE, STATED PLAINLY: this implements the backdoor rule for a DECLARED graph -- adjust for declared confounders, refuse declared colliders and mediators, refuse when nothing valid exists. It does NOT do causal discovery (learning the graph from data), instrumental variables, or sensitivity analysis to unmeasured confounding. Those are still GAP. license_tier: ORIGINAL No hardware writes (Rule 26).