#reinforcement learning
Open access
cvar-sf-repro: reproduction package for "A CVaR Safety Floor for Offline Reinforcement Learning under Unobserved Confounding: Duality and a Finite-Sample Validity Regime"
Sep 2026 · Zenodo (CERN European Organization for Nuclear Research)
Abstract
Reproduction package for the WAIMLAp 2026 paper A CVaR Safety Floor for Offline Reinforcement Learning under Unobserved Confounding: Duality and a Finite-Sample Validity Regime. It ships the confounded medical benchmark and its offline-data generator, the agents (CVaR-TRPO with the Safety-Floor acceptance gate, a CQL-style pessimistic baseline and instrumental-variable value iteration), the Bayesian S-SCM ensemble and its Delphic uncertainty, the experiment configurations, and the evidence files behind every table and figure of the final version, together with the scripts that regenerate them. Policy seeds 0 to 9, each generating its own logged dataset; the estimator-level study uses seed 42.