Limitations

CausalRL is an OPE toolkit, not a replacement for careful causal reasoning.

When not to use it

• If you cannot justify sequential ignorability or any sensitivity model.
• If the behavior policy never explores the target policy's actions.
• If you need guaranteed policy optimization, not just evaluation.

Common failure modes

• Poor overlap leads to unstable IS/PDIS estimates.
• Model misspecification can bias DR/FQE.
• Long horizons amplify variance for importance-weighted methods.

Scope constraints

• evaluate currently assumes discrete action spaces; continuous-action OPE is not implemented yet.
• Transition-only logs must include episode_id and timestep to reconstruct trajectories, or you must build a TrajectoryDataset directly.

What to do instead

• Log behavior propensities during data collection.
• Run diagnostics and show them alongside results.
• Use sensitivity analysis when confounding is plausible.

Versioning caveat

This library is still in 0.x. Public APIs follow SemVer and deprecation warnings, but breaking changes may still occur before 1.0.