Pricing under confounding

The flagship. Offline data is logged under a behaviour policy that ties the action to a confounder z which also drives the outcome — retention teams gave big discounts to the customers most likely to churn. A controller that fits the action's effect without adjusting for z learns the wrong sign and drives the true plant away from its target; adjusting for z recovers the effect and control succeeds. Both controllers plan with their learned effect and act on the true system.

uv run python scripts/flagship_demo.py

What it printed when this site was built:

true effect b   = +1.000
naive estimate  = -0.200   (confounded, wrong sign)
causal estimate = +1.001   (adjusted for z)
target x*       = +2.000
causal  final x = +1.847   -> reaches target
naive   final x = -20.179   -> wrong way, at the actuator limit
wrote outputs/flagship.png

With matplotlib installed the script also writes outputs/flagship.png; tutorial 1 draws the same comparison and scores it against an oracle. Code: chc.flagship, on chc.causal.