ARDA · Discovery Mode IV
Discovers directed causal structures from observational and interventional data. Separates genuine causation from correlation. Designs experiments to resolve ambiguity. Reports what the evidence can and cannot support.
Patent pending in the United States and other countries.
The Problem
Standard models reveal associations. They cannot tell you what would happen if you intervened, what you should measure next, or where the evidence runs out. CDE can.
When gene A and protein B co-vary, a standard model cannot tell whether A causes B, B causes A, or both are driven by an unmeasured confounder. CDE recovers directed causal edges — not undirected associations.
When the observational data leaves multiple causal graphs equally plausible, CDE designs targeted experiments that maximally reduce structural uncertainty — so you spend resources on the measurements that matter most.
Predicting what happens when a mechanism is perturbed requires causal structure, not correlation. CDE identifies what would change — and what would not — if an upstream variable were intervened upon.
Not every causal question is answerable from the available data. CDE explicitly reports what the current data can and cannot distinguish, so you know exactly where the evidence is strong and where ambiguity remains.
Architecture
CDE learns system dynamics and causal structure simultaneously through a dual-representation architecture. It does not fit a causal graph to a frozen model — the causal structure and the dynamics co-evolve.
CDE learns two additive force fields simultaneously: a smooth, universal base field that captures background dynamics, and a graph-conditioned causal field whose forces flow through an explicit adjacency matrix. The causal field fires only along directed edges — if an edge is absent, no causal force is transmitted.
A Bayesian graph neural network maintains a posterior distribution over possible causal structures. Rather than committing to a single graph, the system carries uncertainty across all plausible wirings and reports calibrated edge probabilities.
An information-theoretic probe ranker (BALD — Bayesian Active Learning by Disagreement) scores candidate experiments by how much they would reduce structural uncertainty. The result is a ranked list of experiments ordered by expected information gain.
Causal edges are reported only when evidence is strong. Every claim carries provenance, survives negative controls, and is gated by identifiability and path-law diagnostics. When evidence is insufficient, CDE emits an IndeterminacyClaim instead of a weak CausalClaim.
Identifiability analysis, path-law validation, and out-of-distribution monitoring run continuously. Claims carry confidence caps that tighten automatically when distribution shift is detected or when path fidelity falls below thresholds.
Dual Representation
Most causal discovery methods work on static snapshots or require a fixed model before graph inference begins. CDE takes a different approach: it learns continuous dynamics and causal structure as a single, coupled system.
The base field captures smooth, universal dynamics — the physics that would exist with no causal interactions. The causal field adds directed forces that flow through an explicit adjacency matrix. Remove an edge, and the corresponding force vanishes. This separation means every causal claim is grounded in a specific, testable dynamic contribution.
Base Field
Causal Field
Bayesian Belief
Typed Outputs
CDE does not produce a single score or an opaque prediction. It emits typed claims — each with provenance, confidence, and the negative controls it survived.
Directed causal graph with confidence per edge, entropy, node count, and a list of falsifiers that the claim survived.
Ranked probe actions with BALD scores, expected information gain, target edge, and human-readable rationale for each proposed experiment.
Emitted when evidence is insufficient. Reports candidate graphs, entropy, recommended probes, and a recipe for the next run to resolve ambiguity.
Assesses whether the causal structure is identifiable from the current data. Reports excitation score, intervention coverage, ambiguity score, and weak edges.
Validates the learned dynamics against causal structure. Reports path fidelity score, transition mismatch, intervention-response mismatch, and entropy calibration status.
Monitors distribution shift at serving time. Reports severity, trigger type, recommended action, and a confidence cap applied to downstream claims.
Negative Controls
CDE does not simply assert that a causal edge exists. It systematically breaks the causal structure and checks that the model degrades in the ways it should. If a claim survives these controls, the evidence is credible. If it does not, CDE does not report it.
These are not optional post-hoc checks. They run as part of the discovery pipeline and gate claim emission. A CausalClaim that has not passed negative controls is never surfaced.
Zeros out the entire causal adjacency. If the causal field contributed real structure, prediction quality must degrade measurably. If it does not, the causal claims are not credible.
Shuffles edge weights while preserving marginal statistics. Tests whether the specific wiring — not just the presence of edges — carries information. A genuine causal graph is not exchangeable.
Randomizes the order of interventions across episodes and re-runs the pipeline on the shuffled data. A valid causal model should produce substantially fewer confident claims from placebo data.
API
Every CDE operation is available through ARDA's REST API. Investigate, recommend, intervene, decompose, predict — all as structured API calls that return typed responses.
| Method | Endpoint |
|---|---|
| POST | /v1/cde/investigate |
| POST | /v1/cde/recommend_experiment |
| POST | /v1/cde/apply_intervention |
| POST | /v1/cde/decompose |
| POST | /v1/cde/causal_influence |
| POST | /v1/cde/predict |
| GET | /v1/cde/belief_history |
| GET | /v1/cde/theory_revision |
| POST | /v1/cde/evaluate_identifiability |
| POST | /v1/cde/evaluate_path_law |
| POST | /v1/cde/evaluate_ood |
| GET | /v1/cde/trust_state |
Part of ARDA
CDE is ARDA's fourth discovery mode. It shares the same typed-claim infrastructure, evidence ledger, and governance framework as Symbolic, Neural, and Neuro-Symbolic modes. Results compose across modes because they share a common scientific contract.
Symbolic
Closed-form equations and conservation laws
Neural
Neural differential equations and latent dynamics
Neuro-Symbolic
Neural + symbolic distillation for interpretable laws
CDE
This pageDirected causal structures from observational and interventional data
A research program may use Symbolic mode to discover a conservation law, Neural mode to represent a complex boundary condition, and CDE to establish the causal pathway connecting an upstream variable to a downstream outcome. These results compose because they share the same typed-claim infrastructure and evidence ledger.
CDE is available as part of ARDA. Contact us to discuss your causal discovery requirements.
Causal Dynamics Engine (CDE) is patent pending in the United States and other countries. Vareon, Inc.