Documented failures in deployed AI systems consistently trace to data infrastructure rather than model design, yet the four dimensions most central to data-level accountability—integrity, fairness, synthetic data, and provenance—are typically managed as independent technical controls. The components themselves are individually mature: integrity validation, fairness auditing, synthetic-data generation, and lineage recording each have substantial dedicated literatures. What is missing is the operational coupling between them. This article’s contribution is that coupling: a four-layer governance framework, the Data-Centric Trust Pipeline, that specifies inter-layer dependencies, conflict-resolution protocols for the irreducible trade-offs between integrity and fairness or between fidelity and demographic balance, and a decision-level provenance record structure that captures why and by whom each transformation was authorized rather than only what changed. We implement the framework as open-source Python and exercise it on two canonical fairness benchmarks (Adult Census Income, <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(n = 30{,}162\)</EquationSource> </InlineEquation>; COMPAS recidivism, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(n = 6{,}130\)</EquationSource> </InlineEquation>) across 32 dataset-generator-seed configurations covering 26 independent pipeline executions, spanning three synthesizer families (Gaussian Copula, CTGAN, TVAE) and ten random seeds. The empirical evaluation corroborates—without claiming as novel—the established observation that naive distributional synthesis tends to preserve or worsen fairness rather than correct it [<CitationRef CitationID="CR1">1</CitationRef>, <CitationRef CitationID="CR2">2</CitationRef>]: across all 20 multi-seed configurations the pipeline’s fairness audit triggered a rollback. What the evaluation does contribute is a structural finding about which protocols carry the diagnostic burden: distributional fidelity (P1) and the fairness delta (P2), operating jointly, account for the discriminative work in this evaluation; cryptographic provenance and domain plausibility (P3, P4) are record-level invariants whose role is to surface pathologies that the well-behaved generators in this evaluation did not produce. The governance contribution—operational coupling, conflict-resolution, decision-level provenance—is what we propose; the empirical evaluation calibrates what that governance structure can and cannot demonstrate.

A data-centric trust pipeline:an empirical framework for trustworthy AI in sensitive domains

期刊论文
发表日期：

2026年7月6日

查看全文

Carlos Diego Cavalcanti Pereira

摘要

Documented failures in deployed AI systems consistently trace to data infrastructure rather than model design, yet the four dimensions most central to data-level accountability—integrity, fairness, synthetic data, and provenance—are typically managed as independent technical controls. The components themselves are individually mature: integrity validation, fairness auditing, synthetic-data generation, and lineage recording each have substantial dedicated literatures. What is missing is the operational coupling between them. This article’s contribution is that coupling: a four-layer governance framework, the Data-Centric Trust Pipeline, that specifies inter-layer dependencies, conflict-resolution protocols for the irreducible trade-offs between integrity and fairness or between fidelity and demographic balance, and a decision-level provenance record structure that captures why and by whom each transformation was authorized rather than only what changed. We implement the framework as open-source Python and exercise it on two canonical fairness benchmarks (Adult Census Income, \(n = 30{,}162\) ; COMPAS recidivism, \(n = 6{,}130\) ) across 32 dataset-generator-seed configurations covering 26 independent pipeline executions, spanning three synthesizer families (Gaussian Copula, CTGAN, TVAE) and ten random seeds. The empirical evaluation corroborates—without claiming as novel—the established observation that naive distributional synthesis tends to preserve or worsen fairness rather than correct it [1, 2]: across all 20 multi-seed configurations the pipeline’s fairness audit triggered a rollback. What the evaluation does contribute is a structural finding about which protocols carry the diagnostic burden: distributional fidelity (P1) and the fairness delta (P2), operating jointly, account for the discriminative work in this evaluation; cryptographic provenance and domain plausibility (P3, P4) are record-level invariants whose role is to surface pathologies that the well-behaved generators in this evaluation did not produce. The governance contribution—operational coupling, conflict-resolution, decision-level provenance—is what we propose; the empirical evaluation calibrates what that governance structure can and cannot demonstrate.