How CausalDynamics compares

CausalDynamics.jl owns one niche: identify → intervene → shared-U counterfactual trajectories for discrete-time structural models. It composes with, rather than competes against, the packages below.

Need	Reach for	Why
Learn a DAG from data (PC, OCE)	Associations.jl	Optional weakdep bridge (`infer_pc_graph`, `infer_oce_temporal_spec`); see Associations integration
Learn a DAG from data (legacy Julia PC)	CausalInference.jl	`pcalg`; we use CausalInference for d-separation and backdoor primitives
Solve ODEs / SDEs / UDEs	SciML, UniversalDiffEq.jl	We supply structure and `do(·)` semantics, not integrators — see SciML recipes
Doubly robust effect estimation on tables	TMLE.jl	We identify the adjustment set; TMLE estimates (integration)
Scalable Bayesian inference	RxInfer.jl	Variational backdoor head is a weakdep extension
Tabular causal simulation / estimation	CausalTables.jl	Table-centric; we are trajectory-centric over occasions `t = 1:T`
Python causal workflows	DoWhy	Broad four-step workflow; no discrete-time trajectory counterfactuals with shared `U`

What is distinctive here

Trajectories, not single outcomes — simulate returns a CDMTrajectory with the realised exogenous draws retained
Shared-U counterfactual paths — counterfactual replays the same creative advance under a different intervention, giving unit-level alternative histories rather than population contrasts
Time-indexed identification — unroll_temporal_dag turns a lag specification into a static DAG so ordinary backdoor logic covers lagged confounding
Soft interventions — Policy supports treatment rules that react to state
Lean core — hard dependencies are Graphs and CausalInference only

What is deliberately absent

Full causal discovery implementations, differential-equation solvers, and full symbolic do-calculus. Discovery runs in Associations.jl; identification and trajectories stay here. See Scope.

The CDCS book is the long-form narrative companion for these APIs.