Deterministic Bias Detection for NYC Local Law 144: Why Reproducibility Matters More Than Accuracy

This technical report presents a reproducibility-first architecture for detecting linguistic bias in job descriptions under New York City Local Law 144. While most bias detection tools rely on probabilistic machine-learning models, this work argues that regulatory compliance requires deterministic systems capable of producing identical, verifiable results over time.

The paper details a deterministic bias detection engine built around rule-based pattern matching, version-controlled lexicons, reproducible scoring, and cryptographic evidence generation. It explains why probabilistic AI models struggle to meet evidentiary and auditability requirements, and how deterministic architectures enable replayable analyses, tamper-evident audit trails, and legally defensible documentation.

The report covers system architecture, implementation trade-offs, performance characteristics, known limitations, and deployment patterns, with specific focus on compliance obligations imposed by NYC Local Law 144. It is intended for compliance officers, legal counsel, auditors, HR technology leaders, and engineers building or evaluating AI systems used in employment decision workflows.

This work positions deterministic bias detection as a preventive compliance control rather than a substitute for statutory bias audits, and emphasizes transparency, reproducibility, and evidentiary integrity as foundational requirements for responsible AI governance in regulated hiring contexts.