A data-scarce decision support architecture for auditing high-rise indirect costs: Decoding estimator volatility via dual-engine XAI (A corporate case study in Thailand)

Controlling indirect cost overruns in high-rise construction is fundamentally bottlenecked by the inherent volatility of estimator heuristics and fragmented corporate data, leading to extreme financial risk (CV = 59.0%). This study engineers an automated bias-correction and risk-auditing Decision Support System (DSS) architecture, explicitly designed for small-sample corporate data integration as an enterprise-level proof of concept. Using a highly homogeneous dataset of 15 completed projects' tender baselines from a Thai general contractor, the system employs a decoupled dual-engine computational framework. Support Vector Regression (SVR) serves as the high-dimensional anomaly-detection core ( R 2 = 0.8838 via Nested LOOCV, reducing auditing error by 34.6% compared to the enterprise's historical naïve markup), while a parallel auditing layer (Ridge regression) integrated with exact linear SHAP extracts monotonic, rule-based logic for management dashboards. Results reveal that within these specific typologically controlled boundaries, tender-stage financial proxies—specifically, budgeted worker accommodation and utilities—emerge as the primary sentinel variables for estimating bias, strictly dominating the variance of the predicted cost structure and superseding macro-physical volume. Ultimately, this architecture establishes a transparent auditing baseline, offering a conceptual foundation for management to embed rule-based welfare safeguards directly into the algorithmic auditing loop, demonstrating the potential to prevent excessive cost-cutting and support ESG compliance without compromising tender competitiveness.