HyCIF: A Hybrid Explainable Causal Intelligence Framework for Discovering Causal Relationships beyond Correlation and Generating Trustworthy, Actionable Intelligence

Abstract

Analytics pipelines deployed in practice tend to stop at the point of statistical association. A correlation matrix, a regression coefficient, or a feature-importance ranking is handed to a decision-maker as though it licensed intervention, when in fact none of these quantities distinguishes a cause from a proxy, a confounder, or a collider. Causal-discovery algorithms such as PC, FCI, GES, LiNGAM, and NOTEARS were built to close exactly this gap, but each does so under an idealized assumption — causal sufficiency, faithfulness, a linear or stationary generating process — that observational and temporal data in the wild rarely satisfy in full. Explainable-AI tools such as SHAP and LIME face a related but distinct problem: they attribute a trained model’s prediction to its input features, which is not the same claim as attributing an outcome to its cause, so a variable can receive high importance for reasons that have nothing to do with mechanism. This paper works through a framework, HyCIF (Hybrid Explainable Causal Intelligence Framework), that tries to close both gaps at once. Fifteen modules carry data from raw observation through to a decision recommendation: constraint-based, score-based, functional, and temporal discovery methods are run in parallel and reconciled into a single consensus graph; confounders, colliders, and mediators are screened before any effect is estimated; effects are estimated via the do-calculus rather than naive conditioning; and counterfactual predictions are checked for stability before a relationship is trusted at all. The output of this pipeline is deliberately compressed into two numbers a decision-maker can actually use — a Causal Trust Score (CTS) and an Actionability Score (AS) — both defined mathematically rather than asserted, with the CTS's internal weighting fit by calibration against ground-truth synthetic graphs rather than set by hand. Every module is formalized in terms of structural causal models, the do-calculus, and the backdoor/front-door criteria, and a full evaluation protocol is specified — synthetic graphs with known structure, real observational and temporal benchmarks, an eight-baseline comparison, a fifteen-module ablation, sensitivity and robustness testing, and paired statistical significance testing. Consistent with the goal of not conflating a proposed method with a validated one, this manuscript reports no numerical results: every results table is left as a template for values obtained only after the protocol is actually run.