Notice-and-comment rulemaking is the core mechanism through which the public can shape regulation. As such, courts ask whether agencies have responded adequately to significant comments and scholars often examine whether final rules differ from proposals that precede them—but those approaches typically measure change at the level of entire regulations or, more broadly, documents.

That is the wrong unit of analysis.

The impactful element of a rule is not the document itself, but the discrete instructions to regulated actors contained within; that is, the clauses or sentences that say what a regulated person or entity must, must not, or may do. Those individual commands are what ultimately give rise to compliance obligations in the real world. A given regulation could contain one, one hundred, or one thousand discrete obligations.

When a proposed rule becomes final, what happens to the rule at the obligation level? Do most make it through the process unchanged? Are they tweaked around the edges in response to comments and stakeholder pressure? Or does the rulemaking process reshape them more fundamentally?

New data examining regulatory action at the obligation level suggests that the rulemaking process may operate chiefly in an all-or-nothing fashion. Across hundreds of proposed-to-final rule pairings, only about one-third of proposed regulatory obligations survived unchanged into the final rule—most are removed entirely.

Notice-and-comment, in other words, seems to function in practice less like a line-by-line editing process and more like a filter.

Methods

To measure how regulatory obligations evolve during rulemaking, a dataset was built that extracts sentence-level obligations from Federal Register rules spanning 2016-2026. Each obligation is defined as a discrete deontic instruction directed at an actor—for example, “the operator MUST inspect the equipment” or “no person MAY enter the facility.” Using a deterministic and reproducible filtering mechanism, these sentences are separated and structured into four components: the actor, the modal verb, the action, and the object.

The underlying corpus includes more than 520,000 Federal Register documents, from which hundreds of thousands of individual obligations were extracted and processed. Proposed rules were then paired with their corresponding final rules using metadata like docket identifiers, allowing for obligation sets in each version to be compared directly—like running redlines at scale.

This comparison produces novel metrics, including an overall obligation survival rate, a net obligation flow, and a deontic drift rate. Together, these measures make it possible to observe rulemaking change at the level of the regulatory commands themselves—and compare those measures across both agencies and time.

Results

The results suggest that proposed rules are far less stable than one might expect. Across 954 paired proposed-to-final rule comparisons, less than a third (32%) of proposed obligations survive into the final rule in syntactically recognizable form. The remaining obligations are either removed wholesale or rewritten beyond the similarity threshold that was used to match them.

Among the matched rule pairs, roughly 13,700 proposed obligations disappear from the final rule, while only about 2,050 appear as modified versions of their counterparts. The final rules also introduce about 6,700 obligations that have no counterpart in the proposal at all—indicating that agencies frequently add new commands in addition to eliminating proposed ones.

The asymmetry is noteworthy—proposed obligations are more than six times as likely to be removed as they are to be modified. If notice-and-comment served as an incremental editing function, like a “Track Changes” version of a Microsoft Word document, we would expect most obligations to appear in the modification category. Instead, this portion of the rulemaking process appears to function less like collaborative editing and more like a selection process in which a large universe of proposed obligations is winnowed, reorganized, and replaced before the rule becomes final.

At the same time, when obligations do survive the transition from proposal to final rule, they rarely change their fundamental character. One might expect notice-and-comment to soften or strengthen regulatory commands, for example converting “must” requirements into “may” permissions, or vice versa. But the data suggests those shifts are rare. Across the paired rules the deontic drift rate is roughly 0.6%, meaning that fewer than one percent of surviving obligations change their modal “force.” When an obligation survives, its mandatory or permissive character almost always remains unchanged.

Agency Variance

The stability of proposed obligations is not consistent across agencies—rather, it varies dramatically. Some agencies finalize rules that very closely resemble their initial proposals, while others substantially restructure their obligations during the rulemaking process.

For example, the analysis finds that Treasury regulations exhibit a survival rate of about 57%, meaning that well more than half of proposed obligations appear in the final rule in syntactically recognizable form. Interior rules show a survival rate of about 19%.

The data does not by itself identify the cause of these differences, but several possibilities may suggest themselves. Regulatory domains differ in technical complexity, stakeholder involvement, and political visibility. Some agencies may also engage in more extensive internal drafting and revision before publishing a proposal, leaving less revision to occur during notice-and-comment.

Whatever the explanation, the variation suggests that the dynamics of the comment process in rulemaking may differ substantially across regulatory domains.

Caveats and Limitations

One important caveat turns on a given syntactic “non-survival” not necessarily meaning that the underlying obligation fully disappears in substance. Agencies can rewrite regulatory instructions during rulemaking, and a proposed obligation might be preserved conceptually even if its wording changes enough that it no longer matches the original structure. The survival concept used therefore examines and measures structural persistence rather than semantic continuity or persistence in underlying meaning—the latter being a significantly more difficult endeavor to undertake deterministically.

Even with that limitation, however, the distribution of transformation types provides insights into how rulemaking operates. If notice-and-comment primarily functioned via incremental editing of proposed requirements, we would expect most obligations to appear in the modification category. Instead, removals dominate the data.

The matching method also tolerates significant and substantial textual change before classifying an obligation as non-matching—many edited obligations already appear in the modification category. And even allowing for some rewritten obligations escaping matching, assuming arguendo that the 32% figure is only directionally correct, the overall pattern suggests that rulemaking more often filters and restructures obligations rather than granularly editing them.

It is worth caveating this caveat, however, as even if an agency intends to preserve the underlying meaning of a proposed obligation in its final form, substantial rewriting (of the sort that would register as significant syntactic change) may still have functional and legal consequences. When an obligation changes enough that it no longer appears as a recognizable continuation of the proposed version, it raises the possibility that regulated parties—or a court—may not treat the two provisions as wholly synonymous.

Administrative law frequently turns on close readings and careful interpretation of regulatory texts, and small linguistic differences can affect how a requirement is ultimately interpreted or applied. In that sense, syntactic non-survival may still matter even where an agency itself views the final obligation as conceptually consistent with the proposal.

Conclusions

These findings strongly suggest that proposed rules are less reliable proxies for final regulatory obligations than one might otherwise assume. Empirical studies, impact analyses, and public commentary may treat proposed rules as rough previews of the requirements that will ultimately regulate covered actors—but if only one-third of proposed obligations survive, that focus may be incomplete.

The data suggests a different way of thinking about the role of notice-and-comment in the rulemaking process. Rather than functioning as a disaggregated editing mechanism, in which proposed obligations are incrementally refined and tweaked, the process may in fact operate more like a selection mechanism. Agencies appear to propose a larger universe of potential regulatory commands and then narrow, winnow, restructure, or replace many of them before the rule is made final.

If two-thirds of proposed obligations will not survive to final rule publication, then regulated parties, trade associations, and public interest groups may be devoting resources to commenting on provisions that were never likely to become law. Conversely, final rules routinely introduce obligations with no antecedent—raising questions about whether notice-and-comment adequately provides the public with notice of those obligations that will ultimately bind them. In each case, the dynamic suggests the notice portion of notice-and-comment may deserve closer empirical scrutiny.

Understanding rulemaking in this way highlights, first, how much regulatory change occurs on a lower level of abstraction than the document or regulation itself. The real evolution of a rule lies in how its individual obligations survive, split, are deleted, or are replaced during the process. Measuring those transformations provides an empirical window into how agencies write rules—and how the notice-and-comment process shapes the regulatory obligations that ultimately govern the public.

Andrew Leahey is an Assistant Teaching Professor of Law at Drexel Kline School of Law. Data for this study was derived from his research project FRTracker.app.