Evidence-First Regulation: A Procedural Architecture for Governance Under Deep Uncertainty

Introduction

Federal rulemaking today reflects an impressive commitment to analytic rigor. Major regulations are accompanied by detailed Regulatory Impact Analyses (RIAs), extensive interagency review under Executive Order 12866, and formal notice-and-comment procedures designed to surface relevant evidence and arguments. Agencies routinely estimate compliance costs, model behavioral responses, and test sensitivity assumptions. In many domains, they do so with increasing methodological sophistication. 

Yet alongside this rigor lies a structural imbalance. Rulemaking remains heavily oriented toward ex ante prediction, while institutionalized mechanisms for ex post learning are comparatively underdeveloped. Agencies devote substantial effort to forecasting costs and benefits at promulgation, but far less to systematically measuring whether predicted outcomes occur,  whether assumptions hold under real-world conditions, or whether emerging evidence warrants reconsideration. Retrospective review is formally required, but not systematically embedded within the analytic architecture of rulemaking itself.

This imbalance is not the result of indifference or institutional failure. It is a predictable consequence of how the regulatory process concentrates analytic effort. Executive Order 12866 directs agencies to adopt regulations only upon “a reasoned determination that the benefits of the intended regulation justify its costs,” and to quantify impacts “to the fullest extent that these can be usefully estimated.”  Circular A-4 reinforces that estimates should not create a “false sense of precision”  and cautions that analysis must reflect the limits of available evidence. These directives emphasize reasoned judgment, not numerical perfection. At the same time, they place understandable weight on ex ante modeling because that is the portion of analysis most readily documented, reviewed, and audited at the time of promulgation.

Retrospective review, by contrast, appears elsewhere in the executive framework. Executive Order 12866 instructs agencies to “periodically review” significant regulations to determine whether they should be modified or eliminated. Executive Order 13563 reaffirmed this commitment.  In parallel, the Foundations for Evidence-Based Policymaking Act has strengthened evaluation capacity within agencies.  What remains less developed is a procedural architecture that connects ex ante analysis with ex post evidence in a systematic and transparent way.  The Evidence Act builds evaluation capacity across agencies; this framework integrates that capacity into rule-level analytic design and opens the resulting learning agenda to researchers, industry, and the public—who can see what the agency intends to measure, refine it through notice-and-comment, and contribute evidence of their own during the rulemaking or in anticipation of the review.

This architecture is not without precedent. EPA’s review of the National Ambient Air Quality Standards, refined through five decades of Clean Air Act implementation and judicial review, embodies much of this structure—grading evidence by confidence level, connecting promulgation to mandatory review, and recognizing the Administrator’s judgment as normative discretion. Yet even this mature process has gaps: its normative judgments remain embedded in the scientific record rather than surfaced as separable premises; its review cycle is time-based rather than condition-based, and it operates under unique statutory authority most agencies lack.

Each component of the framework proposed below has precedent in existing federal practice; what no current rulemaking does is implement them together as a coherent architecture. The contribution of this paper is to make the underlying architecture’s logic explicit, portable, and complete.

The need for such an architecture is most evident in domains marked by rapid change and uncertainty. In slower-moving sectors, gaps between predicted and realized outcomes tend to emerge gradually, and institutions can absorb modest forecasting error. In faster-moving areas—such as artificial intelligence, autonomous systems, digital platforms, and advanced biotechnology—uncertainty appears more quickly and more visibly. Model drift, context-dependent performance, and feedback effects can undermine static assumptions. Under these conditions, an approach focused primarily on one-time prediction risks entrenching estimates that may diverge from real-world outcomes.

The underlying issue is not that prediction is misguided. Forecasting remains a critical directional tool, and benefit–cost analysis remains the analytic engine for estimable effects. Nothing in this framework alters the welfare objective embedded in Executive Order 12866. But BCA alone does not exhaust the analytic demands of rulemaking under deep uncertainty.  When key effects are learnable only through post-implementation observation, and when some elements of a decision rest on contested normative judgments, the analytic architecture needs to extend beyond estimation—to structured learning and transparent normative reasoning. The framework provides that extension: BCA handles what can be estimated; the Learning Summary plans for what must be measured over time; and the Decision-Rule Table surfaces the value judgments that no amount of data will resolve.

The challenges addressed here have received sustained scholarly attention. Adaptive governance theorists have developed frameworks for managing regulatory problems under conditions of deep uncertainty, emphasizing institutional flexibility, iterative decision-making, and feedback as alternatives to static ex ante optimization.  Experimentalist approaches to governance have explored how regulatory commitments can be structured as provisional and subject to deliberate revision based on performance data, embedding learning mechanisms into the architecture of governance itself.  Evidence Act scholarship has examined how federal evaluation infrastructure can be strengthened to support more systematic learning across agencies. 

The contribution of this paper is to integrate existing analytic practices into a single procedural architecture that operates at the level of individual rule design—actionable structure instead of aspirational theory. Existing approaches address modeling, evaluation, or adaptive governance separately; this framework embeds all three within the administrative record at promulgation. It does so by introducing an explicit epistemic grammar—the Three Decision States—that distinguishes what can be reasonably estimated, what must be learned through observation, and what reflects normative judgment.

Building on this structure, the paper develops four procedural tools—the Four Pillars of Evidence-First Regulation. Adequacy over Precision formalizes a stop-rule for modeling grounded in marginal informational value, in line with Circular A-4’s guidance on useful estimation.  The Adaptive Statement organizes retrospective review into a learning plan. The Decision-Rule Table makes explicit the normative commitments that shape welfare calculations. The Review Spine connects ex ante assumptions to ex post evaluation through scheduled, nonbinding review points. Although the architecture is domain-general, it is especially salient for emerging technologies—artificial intelligence, autonomous systems, advanced biotechnology—where uncertainty is structural and the gap between prediction and reality materializes quickly.

The architecture extends the temporal horizon of benefit–cost analysis. By pairing ex ante modeling with structured ex post evaluation, it allows welfare analysis to be progressively refined rather than fixed at promulgation. Courts retain their traditional role under State Farm; the framework organizes the agency’s explanation more transparently.

The sections that follow proceed in six stages. Section II establishes the value-of-information logic that supports reallocating analytic effort toward observation. Section III introduces the Three Decision States. Section IV develops the Four Pillars. Section V applies the framework to a hypothetical AI auditing rule for high-risk systems, demonstrating how the architecture operates in an emerging technology domain where uncertainty is structural and feedback cycles are compressed. Sections VI and VII address legal compatibility and institutional feasibility. Section VIII examines precedent in existing federal practice—illustrating both statutory and voluntary adoption of framework-like elements, and identifying where the architecture adds value. The paper concludes by returning to the core proposition: that structured learning, embedded within existing authorities, can enhance both the rigor and legitimacy of federal regulation in a changing technological environment.

This project was made possible through the support of Grant 63641 from the John Templeton Foundation. The opinions expressed in this publication are those of the author(s) and do not necessarily reflect the views of the John Templeton Foundation. For more information, visit The Next Frontier: Rethinking Regulation in an Era of Rapid Innovation.