NeurIPS submission on deployable causal action geometry

A NeurIPS 2026 submission studies continuous decision making from observational time series when outcome levels drift over time.

The paper argues that raw predictive fitting can mix action effects with shifting baseline conditions, so historical prediction accuracy is not enough for future deployment. It introduces a restricted target called deployable causal action geometry, using anchored contrasts over link functions to isolate the action-dependent response morphology while allowing period-specific calibration.

The submission pairs this target with cross-fitted orthogonal pilots and honest profiled selection over morphology, representation, and shape classes. Author details and public PDF links are withheld while the submission is under double-blind review.