An Auditable Data-Quality Layer via Agentic Rule Discovery

01 ABSTRACT

The authors observe that high-stakes tabular pipelines rely on scattered, static quality checks that are hard to evolve and to audit against policy intent. They propose a data-quality layer in which rules are linked to governance texts, versioned, and replayable through immutable execution traces, and implement its agentic rule-discovery core with Full-Agentic and Guided modes plus a human-in-the-loop review interface. Experiments on two real industry datasets in finance and retail are reported as preliminary results on whether fully autonomous divide-and-conquer rule discovery matches human-expert rule quality and diversity.

02 KEY FINDINGS

1. Proposes an ontology-grounded, repository-native data-quality layer where rules are versioned, provenance-linked to governance texts, and replayable via immutable execution traces.
2. Implements an agentic rule-discovery core with two acquisition modes (Full-Agentic and Guided) and a human-in-the-loop review interface.
3. Reports preliminary experiments on one finance and one retail dataset comparing autonomous rule discovery with human-expert rules in quality and diversity.