EvoCausalTwin: Drift-Aware Causal Structure Learning, Evidence-Calibrated Verification, and Budget-Constrained Intervention for Evolving Systems

Evolving systems undergo distributional, mechanistic, and structural changes that can invalidate static causal models and lead to poorly justified interventions. This paper presents EvoCausalTwin, an evidence-gated artificial intelligence framework linking dynamic causal structure learning, causal-drift decomposition, evidence-calibrated verification, budget-aware intervention, causal-effect decision analysis, and constrained workflow auditing. Its three language-model auditor roles share one quantized Llama3.1-8B-Instruct model and cannot orient graph edges or bypass statistical verification. We evaluate the framework on EvoCausalBench, comprising 2,000 source structural causal models and 50,000 episodes across four graph families, five mechanism families, five drift regimes, and 10–100 variables. Splits are assigned at the source-SCM level before episode generation. On the principal test set, EvoCausalTwin obtains SHD 4.8, directed F1 0.79, ATE RMSE 0.22, PEHE 0.32, five-class drift macro-F1 0.77, decision regret 0.24, and drift-class ECE 0.061, compared with 6.1, 0.74, 0.27, 0.39, 0.66, 0.36, and 0.142 for CD-NOD. Paired analysis uses 300 held-out SCM-level vectors. A separate 50,000-observation reliability audit yields decision-confidence weighted ECE 0.01845, with subgroup calibration reported across all five drift regimes. A full 2^4 factorial ablation distinguishes the descriptive contributions of modules M2–M5. Twelve-month operational evidence is retrospective and does not establish prospective causal treatment effectiveness. The findings support a framework-level contribution under the evaluated protocols, not a new standalone causal-discovery estimator

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11668674/v1
Primary Topic
Bayesian Modeling and Causal Inference
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

EvoCausalTwin: Drift-Aware Causal Structure Learning, Evidence-Calibrated Verification, and Budget-Constrained Intervention for Evolving Systems

Phat Hai Tien Nguyen
Research Square
Bayesian Modeling and Causal Inference
preprint

EvoCausalTwin: Drift-Aware Causal Structure Learning, Evidence-Calibrated Verification, and Budget-Constrained Intervention for Evolving Systems

Phat Hai Tien Nguyen
preprint en

Abstract

Evolving systems undergo distributional, mechanistic, and structural changes that can invalidate static causal models and lead to poorly justified interventions. This paper presents EvoCausalTwin, an evidence-gated artificial intelligence framework linking dynamic causal structure learning, causal-drift decomposition, evidence-calibrated verification, budget-aware intervention, causal-effect decision analysis, and constrained workflow auditing. Its three language-model auditor roles share one quantized Llama3.1-8B-Instruct model and cannot orient graph edges or bypass statistical verification. We evaluate the framework on EvoCausalBench, comprising 2,000 source structural causal models and 50,000 episodes across four graph families, five mechanism families, five drift regimes, and 10–100 variables. Splits are assigned at the source-SCM level before episode generation. On the principal test set, EvoCausalTwin obtains SHD 4.8, directed F1 0.79, ATE RMSE 0.22, PEHE 0.32, five-class drift macro-F1 0.77, decision regret 0.24, and drift-class ECE 0.061, compared with 6.1, 0.74, 0.27, 0.39, 0.66, 0.36, and 0.142 for CD-NOD. Paired analysis uses 300 held-out SCM-level vectors. A separate 50,000-observation reliability audit yields decision-confidence weighted ECE 0.01845, with subgroup calibration reported across all five drift regimes. A full 2^4 factorial ablation distinguishes the descriptive contributions of modules M2–M5. Twelve-month operational evidence is retrospective and does not establish prospective causal treatment effectiveness. The findings support a framework-level contribution under the evaluated protocols, not a new standalone causal-discovery estimator

Research Square
Ton Duc Thang University (VN)
Bayesian Modeling and Causal Inference
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon