Why, and when not to use it

The gap

Each of the three obvious approaches solves a different problem and leaves the same gap:

• Forecast + argmax. Optimises a correlation. Under a confounded logging policy the fitted action response is the observational one, so the planner acts on the wrong effect and never notices. Tutorial 1 shows the sign flip and what it does to control.
• Causal effect estimation alone (DML, staggered DiD, synthetic control, causal forests). Returns an effect, not a decision. The effect can be dynamic — EconML's DynamicDML estimates each period's within the logged horizon — but the estimator chooses no actions: no actuation limits, no plan past the logged horizon, no notion of a state the action must not reach. chc treats these as backends (chc.estimators) rather than rivals.
• Offline RL / MPC on a learned model. Fits the dynamics by residual MSE, which is not identification, and calibrates its pessimism to sampling noise rather than to unmeasured confounding — so the uncertainty penalty shrinks with N while the bias does not. A few methods are the exception and penalise the uncertainty that hidden confounding leaves: Delphic offline RL (Pace et al., 2024) measures it by the disagreement among world models compatible with the data, robust fitted-Q iteration (Bruns-Smith and Zhou, arXiv 2302.00662) takes the worst case over a sensitivity model, and confounding-robust DQN (Li, Zhang and Bareinboim, 2025) the worst case over the environments compatible with the observations.

What is here is the seam: identify the interventional control channel from confounded logs (chc.dynamics_id), plan against it under constraints, and price what unmeasured confounding can do to the plan's performance (chc.sensitivity) and to its safety (chc.barrier).

Where it sits

Read by the two questions this library refuses to merge: does the tool identify the effect of an action from data a policy generated, and does it certify the plan it hands you. Most tools answer one. Of the packages here only the closed decisionOS claims both; elsewhere both are answered in papers. Versions and licences read 2026-10-06; what each of CHC's certificates bounds is in the scopes.

	what it is for	identifies an interventional effect	produces a schedule	ships a certificate	licence
chc	decisions from a confounded log, over a plant	yes, the control channel of a control-affine residual (cross-fit Robinson DML), and it says not_identified rather than guessing	yes, projected gradient over a box and linear constraints (budgets, rate limits), with a confounding-robust barrier held in the solve	yes — identification status, trajectory tube, barrier Γ*, each with its scope	Apache-2.0
DoWhy / DoWhy-GCM 0.14	identify and refute an effect on a DAG	yes — back-door, front-door, IV, the Rotnitzky–Smucler efficient backdoor set, which minimises asymptotic variance among backdoor sets, and since 0.13 Perković et al.'s generalised adjustment criterion, which returns a minimal valid set whenever one exists. CHC's CausalGraph returns that criterion's canonical set instead, which is valid iff any observed set is	no	refutation tests, not a control guarantee	MIT
EconML 0.17.0	heterogeneous treatment effects, DML/DR/orthogonal forests	yes, static and sequential: its DML estimators return a matrix θ(X), and DynamicDML estimates each period's effect of an adaptively assigned treatment on the last period's outcome (sequential ignorability, a linear-Markov state, a balanced panel). CHC applies the same orthogonal moment one step at a time to a control-affine plant	no — it chooses no sequence	confidence intervals	MIT
DCBO, last commit 2023-04	sequential interventions in a time-varying SCM	yes, by GP emulation over an SCM	yes, a sequence of interventions	regret empirics, no feasibility guarantee	ambiguous: MIT in LICENSE, GPL-3.0-or-later in the README and setup.py; research code, not on PyPI
Google Meridian 2.1.0	Bayesian marketing-mix modelling	partially — priors and geo experiments calibrate it; the estimand is the media response	no — it splits a budget across channels and holds each channel's flighting over geos and time fixed; a flighting can be supplied, not optimised	posterior intervals, and a model review of five checks	Apache-2.0
PyMC-Marketing 1.2.0	Bayesian marketing-mix modelling, with time-varying media and intercept	partially — lift tests enter the likelihood; the estimand is the media response	a budget split across channels, and across the dims of its multidimensional model, with value-at-risk, CVaR and Sharpe utilities; a flighting is supplied, not optimised	posterior intervals	Apache-2.0
Robyn 3.12.1, R	marketing-mix modelling by ridge regression under an evolutionary hyperparameter search	partially — lift tests calibrate the search; the estimand is the media response	a budget split across channels (robyn_allocator), not a schedule over time	no	MIT
Orbit 1.1.5.1	Bayesian time-series forecasting, with time-varying regression (KTR)	no — a regression on the log	no	posterior intervals	Apache-2.0
do-mpc 5.1.2	robust and economic nonlinear MPC	no — the model is yours and assumed correct	yes, and more general constraints than CHC's: nonlinear path constraints on the state itself, where CHC holds linear rows on the actions and a barrier's decay condition	robust multi-stage MPC guarantees, under a correct model	LGPL-3.0
d3rlpy 2.8.1	offline deep RL from logged trajectories	no — conservatism bounds value error, not confounding	yes, a policy	pessimistic value bounds	MIT
causaLens decisionOS	enterprise causal decision platform	yes, per its own account	yes	not publicly auditable	commercial, closed

Two rows that are not here, and the reason is the same. The 2024–26 literature on causal Bayesian optimisation under safety constraints, and on causal optimal control ("COAST"-style), has no shipped, installable implementation this could be run against. The nearest code is cCBO, for constrained causal Bayesian optimisation (Aglietti et al., ICML 2023): research code under Apache-2.0, not on PyPI, last changed 2024-06. COAST (arXiv 2605.29008) promises its code on publication. That is the gap CHC is aimed at, and stating it as an absence is more honest than a row of dashes against a paper.

Where a row says no it is not a criticism: do-mpc solves control problems CHC cannot state, and EconML answers effect questions CHC does not ask. The EconML row's yes was run, not read: scripts/econml_reference.py gives LinearDML a 2×2 channel B(x) behind a confounded action, which it recovers to 0.068 at worst, and DynamicDML a linear plant logged over three periods, whose impulse response it recovers to 0.013, every period's 95 % interval covering the truth (EconML 0.17.0, 2026-10-06). The claim is narrower — that going from a confounded log to a schedule and its certificate in one place is what nothing above does end to end.

When not to use it

The same table, read the other way.

• Your model is known and trusted, and the constraints are nonlinear in the state. That is do-mpc's problem, and it states control problems CHC cannot: CHC holds linear rows on the actions and a barrier's decay condition, not general path constraints on the state. CasADi, acados and do-mpc are far stronger solvers, and CHC has nothing for embedded MPC.
• You need an effect, not a decision. Heterogeneous effects are EconML's question, and within the logged horizon so are dynamic ones: DynamicDML estimates each period's effect with intervals, and CHC adds nothing there. Identifying and refuting an effect on a DAG is DoWhy's. CHC calls such estimators through chc.estimators rather than competing with them.
• You need sequential off-policy evaluation beyond a linear-Gaussian plant. chc.evaluation has per-decision and marginalised importance sampling, doubly robust and fitted Q, and certifies a log in advance only on a linear-Gaussian plant with affine policies; SCOPE-RL 0.2.1 has more sequential estimators, the DICE family among them.
• You want a Bayesian marketing-mix model calibrated by priors and geo experiments: that is Meridian. The media-budgets case study is about the other half — scheduling spend when the log was planned against demand. Planning that schedule through the adstock state is Nerlove and Arrow (1962); the shipped tooling is new, the idea is not.
• You can run experiments on the system. Causal Bayesian optimisation intervenes and learns from each intervention; prescribe reads a log and never experiments. On DCBO's own three dynamic SCMs (Track L of causaldyn-bench), prescribe from the log alone is level with DCBO on the stationary one and behind every method that experiments, plain BO included, on the other two. The track names what it lacks there: a minimise objective, a choice of which variables to intervene on at each step (max_levers chooses one set for the whole horizon), experiments, and a model that changes in time. The converse holds too: DCBO reads the true SEM for every explorative evaluation, so it cannot run from a log alone.
• Your residual is not control-affine. chc.dynamics_id is the identified route, and it is restricted to control-affine residuals; outside that class this library offers a sensitivity radius (chc.sensitivity), not an unbiased estimate.
• Nothing in your log identifies the effect — no adjustment set and no instrument. Then there is no schedule to be had: prescribe returns none, on purpose, and what remains is to price the radius (the sensitivity level Γ).
• You need a frozen API today. This is early, single-author research code before 1.0. The stability tiers say what each module promises; an experimental module may change or be withdrawn in any release.

Honest positioning

chc composes ideas that exist — hybrid dynamics (SciML UDE), pessimistic offline control (MOPO/MOReL/Delphic), sequential causal identification (g-methods / dynamic treatment regimes), differentiable control (Neuromancer). The contribution is the integration behind one API plus a benchmark with ground-truth interventional effects. KAN is one interpretable residual backend, not the identity of the framework — and chc.symbolic makes "interpretable" checkable rather than asserted, including the two places where the interpretation stops being valid.

Worth knowing before you rely on a fitted model: a low residual MSE is not causal identification. Fitting r_θ by prediction error recovers the observational control response, which is the wrong one whenever the logged action was chosen from something that also moved the state, and no amount of extra training fixes it because it is not a fitting problem. See identification.