Data-Centric Trust Pipeline

Four mature dimensions, none talking to each other

Integrity, fairness, synthetic data and provenance are managed today as independent technical controls. Each is individually mature — what is missing is the operational coupling between them.

In practice: an organization can pass integrity checks at ingestion, satisfy fairness audits at deployment, and still produce biased predictions — because the integrity check never assessed demographic completeness, and the fairness audit cannot trace which records were excluded during validation.

An upstream problem — with a moral structure

In healthcare, public administration, financial services and criminal justice, AI fails not because models are poorly designed but because the data feeding them are incomplete, poorly documented or demographically unrepresentative. The problem is upstream, not downstream.

It is also institutional: the people affected rarely had any visibility into the data decisions that shaped their outcomes, and institutions rarely retained the lineage needed to answer for them afterward. Trustworthiness is not only a technical property of the model — it is a property of the practices that produced it. Recent regulation (EU AI Act, ISO/IEC 42001, NIST AI RMF) shifted the framing toward organizational accountability; a necessary but insufficient step: it says what to document, not how the layers connect.

A single, auditable control system in a recursive cycle

The Data-Centric Trust Pipeline couples four interdependent governance layers in a recursive cycle, formalizing the dependencies between them, the conflict-resolution protocols for irreducible trade-offs, and a decision-level provenance — recording why and by whom each transformation was authorized, not merely what changed.

• ›Three conflict protocols: Fairness–Synthesis, Integrity–Fairness and Provenance–Integrity — which do not resolve the trade-off automatically but document and escalate it to a responsible human.
• ›Five design principles: recursive verification, explicit handoffs, human escalation, decision-level provenance and synthesizer-agnostic validation.
• ›Open-source implementation in Python (~1,560 lines, 5 modules), with deterministic reproduction scripts.

Making the invisible operationally diagnosable

The goal is not a new generator or a new fairness criterion, but the structure that connects them: turning four problems handled in parallel into a single control system that produces auditable evidence — provenance graphs, per-batch outcomes and rollback decisions — and that empirically calibrates which protocols carry the diagnostic burden.

What the empirical evaluation showed

Exercised on two canonical fairness benchmarks — Adult Census Income and ProPublica COMPAS — across 32 configurations (dataset × generator × seed), three synthesizer families (Gaussian Copula, CTGAN, TVAE) and ten random seeds:

• ›P1 (fidelity) alone would admit all ~80 batches without noticing fairness was worsening — it is P2 operating jointly that exposes the regression.
• ›Three generator families produce qualitatively distinct failure modes: no single checker suffices.
• ›An auditor can answer why, by whom and on what evidence each decision was made — which conventional lineage systems do not capture.

Where the research goes next

• 01Replicate with diffusion-based tabular synthesizers — a generative family not yet evaluated.
• 02Scale to industrial pipelines, with more protected attributes and intersectional groups.
• 03Empirically exercise the Integrity–Fairness and Provenance–Integrity conflicts, specified but not yet triggered.
• 04Characterize the conditions of governable augmentation — the single accepted-synthesis case (TVAE–COMPAS) calls for more observations.

Open fronts for collaboration

• ›Operational counterpart to regulation — connect the EU AI Act, NIST AI RMF and ISO/IEC 42001 to the mechanics of generating the evidence.
• ›Multivariate P1 metric — capture joint-dependency fidelity, not only marginals.
• ›Category auditing — self-reported vs. administrative race, with the disagreement recorded in provenance.
• ›Production-grade P4 rules — co-occurrence constraints derived from domain experts.
• ›Anti-“rubber-stamp” governance — processes that turn authorization records into a real evaluation of human judgment.
• ›Sector applications — healthcare, public administration, financial services and criminal justice.