# Section 1: Introduction
<!-- Registry references: SD-001, PL-001, S2-010, S2-011, S3-001–003 -->
<!-- Citation files: federal_data_evolution_arc.md, core_argument.md, nsf_norc_landscape.md -->
Federal statistical agencies have spent two decades making their data accessible to machines. Beginning with the machine-readable mandates of the late 2000s and accelerating through structured APIs, metadata catalogs, and master data registries, the investment has been substantial and real. The Census Bureau's API, the Bureau of Labor Statistics' data retrieval tools, and the standardized metadata schemas across Commerce Department statistical assets represent a mature infrastructure for data access. The syntax layer — how data is structured, formatted, and transmitted — is largely solved.
The semantics layer has followed a parallel trajectory. Variable descriptions, concept classifications, table schemas, and geographic hierarchies are documented, standardized, and published. This metadata infrastructure enables both human researchers and automated systems to identify which data products exist, what they measure, and how they are organized. Recent federal initiatives under the banner of "AI-ready data" have extended this work, recognizing that machine learning systems require well-structured metadata to function effectively.
The emergence of large language models has changed the equation in an unexpected way. Models trained on broad corpora that include statistical documentation, methodology reports, and data dictionaries behave as if they have internalized much of this semantic infrastructure. They can translate natural language questions into domain-appropriate queries, identify relevant variables, resolve geographic entities, and retrieve data through APIs — tasks that previously required specialized training or purpose-built search interfaces. The syntax and semantics layers, painstakingly constructed over two decades, are now partially encoded in model training data.
This creates a new problem. When a user asks a language model for the poverty rate in a small county, the model can successfully retrieve the correct estimate from the Census API. But it cannot assess whether that estimate is reliable enough to use. It does not know that the margin of error may exceed the estimate itself, that the coefficient of variation renders the figure unsuitable for most analytical purposes, or that the five-year period estimate represents a 60-month weighted average rather than a point-in-time snapshot. The model delivers the number confidently. A non-expert user has no basis to question it.
This failure mode is not a knowledge gap in the conventional sense. The model is not missing information that could be retrieved from a document or looked up in a database. It is missing expert judgment about fitness for use — the kind of assessment that a senior statistician makes reflexively when evaluating whether a particular estimate is appropriate for a particular purpose. This judgment is rarely stated explicitly in documentation. It lives in the professional practice of experienced practitioners, accumulated through years of working with the data and its limitations.
We call this missing layer *pragmatics*, drawing on Charles Morris's 1938 semiotic framework that distinguishes syntax (the formal structure of signs), semantics (the relationship between signs and what they denote), and pragmatics (the relationship between signs and their interpreters — the contextual judgment required for appropriate use). In the context of federal statistical data, pragmatics is the expert assessment of fitness for use that transforms a data retrieval into a statistical consultation.
This is not a new concept imposed from outside statistical practice. The Federal Committee on Statistical Methodology's own data quality framework (FCSM 20-04) codifies characteristics — relevance, accuracy, timeliness, accessibility, coherence — that are fundamentally pragmatic in nature. They describe not what the data *is* but whether the data is *appropriate* for a given purpose. These quality characteristics have been the standard for decades. What has not existed, until now, is a mechanism to deliver this expert judgment computationally, at the point where a user or automated system is interpreting statistical data.
The current federal landscape reflects this gap. The National Science Foundation recently solicited proposals to measure how well language models understand federal statistical data, seeking empirical evaluations of LLM accuracy, relevancy, and explainability on government data assets (NCSES, 2025). This and similar benchmarking initiatives share a common focus: measuring how well models perform on statistical tasks. They diagnose the problem. They do not treat it.
This paper introduces pragmatics as a defined, implementable concept for federal statistical AI systems and provides empirical evidence that it works. We present a knowledge representation study comparing three conditions with identical data access: a control with no methodology support, retrieval-augmented generation (RAG) using document chunks from authoritative source material, and pragmatics using curated expert judgment items delivered at the point of statistical reasoning. The three conditions draw from the same 354 pages of Census Bureau documentation, differing only in how that knowledge is represented and delivered.
The results demonstrate that 36 curated pragmatic items produce very large improvements in consultation quality relative to no support (Cohen's d = 1.440) and large improvements relative to RAG (d = 0.922), with the strongest effects on uncertainty communication — precisely the dimension where fitness-for-use judgment matters most. Pragmatic context achieves 91.2% fidelity to authoritative data sources compared to 74.6% for RAG, at a marginal cost of nine cents per query.
The contribution is not a better retrieval system. It is the identification and operationalization of a missing layer in the federal statistical data ecosystem — a layer that has been conceptually present in quality frameworks for decades but has never been delivered computationally. Making data AI-ready requires three things: refactoring how data is exposed to AI systems, accelerating metadata curation, and encoding the expert judgment needed to evaluate fitness for use. The first two are underway. The third is the subject of this paper.
