The Architecture of Not-Knowing: Strategic Ignorance, Fragmented Oversight, and the Requester Game in Regulated Industries
Joseph Phelan
The Global Centre for Advanced Studies, Dublin; F'nAround Media
ORCID: 0009-0008-3135-3083
October 26, 2026.
DOI: 10.13140/RG.2.2.16528.98566
Abstract
This article asks how multi-agency regulatory systems sustain strategic ignorance as a stable outcome in public records practice under freedom of information (FOIA) law. Drawing on a multi-year Illinois public records file on cannabis regulation, with a Kentucky comparison, it argues that fragmented oversight produces partial knowledge as an equilibrium, and that this equilibrium is a problem for regulatory governance, not only for transparency. The article develops a five-mechanism typology of not-knowing: sequestering, decompiling, delay, referral loops, and exemption with partial visibility. It then models, in prose, the requester game that makes the typology stable: given the costs each side faces, partial compliance by the system and persistence by the requester are the strategies that survive, while full release and full denial are both costly. The Illinois evidence, drawn from a dated chronology of more than 425 entries between June 2023 and September 2026, illustrates each mechanism with documented episodes, and the suit record that surrounds parts of the file is used only as dated evidence. A Kentucky medical cannabis FOIA chronology from December 2025 to October 2026 provides a second site, coded with the same rules. The article states five falsifiable propositions and a coding protocol so that other researchers can test the typology against other records files. Its claim is structural, not personal: not-knowing in this account is a property of fragmented systems, not a charge against any official, and the article claims no motive that the records cannot show.
Keywords: strategic ignorance; public records; FOIA; regulatory governance; fragmented oversight; institutional theory; administrative burden
Introduction
Consider a requester who asks a straightforward question of a regulatory system: who approved what, on what record, and when? The requester files with one agency and is told that another unit holds the relevant fragment. The requester files with that unit and receives a partial production. A third unit extends its deadline, then corrects its own extension letter on the same day. A fourth unit answers that it holds no records of a referral, while an earlier answer from the same department, concerning a transfer, said it held no responsive records. A credential listing shows two licenses active and in good standing, with recent issue dates, while the approval materials behind those licenses are withheld under exemption. A large production arrives from another office in tranches, and the requester narrows a sub-request to a slice of files in order to see anything at all. No single response in this sequence is scandalous on its own. Together, the responses add up to a system in which no one holds the whole picture, and no one is positioned to assemble it.
This article asks how that outcome is sustained. Its question is: how does a multi-agency regulatory system sustain not-knowing as a stable outcome, rather than as an accident or a failure? The question matters for two reasons. First, the standard account of transparency failure emphasizes either bad faith or backlog. Bad faith explains secrecy as concealment by officials who know what they are hiding. Backlog explains delay as a resource problem, in which too many requests chase too few staff. Both accounts have force in particular cases, but neither fits a pattern in which every unit complies in part, every deadline is eventually met in some form, and the whole picture still never arrives. Second, the stakes of that pattern are regulatory, not only democratic. Licensing oversight depends on the knowledge the system holds. A regulator that cannot assemble its own file cannot evaluate compliance, whatever its rules say. A public that cannot assemble the file cannot evaluate the regulator. Not-knowing is therefore a capacity problem as much as a transparency problem, and it belongs at the center of regulation research rather than at its edge.
The article's answer has three parts. The first is a typology. Drawing on ignorance studies, institutional theory, and public administration, it identifies five mechanisms by which fragmented systems produce not-knowing: sequestering, decompiling, delay, referral loops, and exemption with partial visibility. The second is a game-theoretic account, stated in prose without formal mathematics, of why the typology is stable. The requester and the agency system face asymmetric costs, and the strategies that survive their interaction are partial compliance on one side and persistence on the other. The third is evidence. The article illustrates each mechanism with documented episodes from a multi-year public records file on cannabis regulation in Illinois, uses the suit record that surrounds parts of that file only as dated evidence, and tests portability against a second site in Kentucky medical cannabis records, coded with the same rules. It then states five falsifiable propositions and a coding protocol, so that the argument can be confirmed or refuted against other files.
Three scope decisions shape the article, and each is stated here so that reviewers can judge it. First, the article studies structure, not motive. It claims that fragmented systems produce partial knowledge through regular operations, and it does not claim that any official intended that result. Where the records cannot show intent, the article says so, and its strictest coding rule enforces the point: a non-response is coded as indeterminate unless a later release proves that the fragment existed. Second, the article's evidence is a single requester's file, supplemented by a second site. That design can establish mechanisms and their observable implications, but it cannot estimate how common each mechanism is across requesters or systems. Prevalence is the next study, and the article names it as such rather than claiming it. Third, the article is written from inside the file it studies. The author is the requester whose chronology and response file supply the Illinois evidence, and was a party to the civil action whose record appears here as dated evidence. That position gives the article access no outside researcher currently holds, and it also creates a bias risk that the article manages through conservative coding, disconfirming-evidence procedures, and the propositions themselves, which are stated so that a second file can refute them. Positionality is addressed in full in the method section.
Three further clarifications about what this article is not will help reviewers place it. It is not a prevalence study. It does not estimate how often each mechanism occurs across requesters, agencies, or states, and no such estimate should be read into its evidence section, which illustrates mechanisms with documented episodes. It is not a legal brief. It makes no argument about whether any exemption, extension, or referral was lawful, and it draws no conclusion about any party's rights or liability. It is not a case against any official. It names no individual, attributes no motive, and treats every episode as an operation of a fragmented system rather than as conduct by a person. Those restraints are not evasions. They are what make the article testable, because an argument that claims only structure can be refuted by structure, while an argument that claims motive can always retreat into unobservable intention.
The article proceeds as follows. The next section places the argument in the literatures it joins: transparency and secrecy research, information strategy, organizational institutionalism, mobilization and burden research, and regulation theory. The theory section then defines the five mechanisms and develops the requester game. The evidence section presents the Illinois file organized by mechanism, followed by the suit record as dated evidence. The Kentucky comparison follows, based on a Kentucky FOIA chronology coded with the same rules as the Illinois file. The method section states the tracing design, the coding rules, and the five propositions. The discussion answers the strongest objections in advance, states the limits plainly, and specifies what would falsify the argument. The conclusion draws out the implications for regulation research and for the design of oversight.
Literature: The Conversation This Paper Joins
This article does not merely cite the literatures it draws on. Each body of work below supplies one part of the argument, and each part is tied to a mechanism or a proposition, so that the literature section positions the article rather than decorating it.
Transparency, secrecy, and the limits of the right to know
Roberts (2006) asked, in Blacked Out, whether the old presumption of secrecy had really been overthrown in favor of a new presumption of openness, and answered that it had not. Access legislation spread widely, from a handful of countries in the early 1980s to dozens within two decades, yet governments continued to ignore or frustrate release requirements, and structural changes in how government is organized, including opaque networks and the contracting out of public services, complicated every campaign for openness. Roberts gives this article its field-level premise: law does not equal visibility. The article's contribution is to specify, at the level of a single file, how that premise operates. Delay, referral, splitting, and partial visibility are the micro-operations of Roberts's structural point. Where Roberts shows that secrecy persists after FOIA laws, this article shows the operations by which it persists, request by request.
Pozen (2017) pressed the critique further by arguing that FOIA embodies a distinctively reactionary form of transparency. Disclosure responds to ad hoc demands. That feature, Pozen argued, saps regulatory capacity, distributes government goods in an inegalitarian fashion, feeds a culture of adversarialism around the domestic policy bureaucracy, and insulates the most secretive agencies and corporations from similar scrutiny. This article accepts the force of that critique and changes its level of analysis. Where Pozen asks whether FOIA is the right instrument, this article asks how the instrument behaves when it is used: what architecture turns ad hoc demands into partial knowledge? Two of Pozen's points tie directly into the article's machinery. His inegalitarian point, that FOIA's benefits flow disproportionately to the resourced, is the distributive logic behind Proposition 2, which predicts that persistence, and therefore success, tracks resources. His adversarialism point is not only a culture around FOIA. In this article's account, adversarialism is the structure of the requester game itself, in which each side's strategies are defined against the other's.
Pozen and Schudson (2018), in Troubling Transparency, and Schudson's (2015) history of the right to know, place FOIA's mixed legacy in view: a landmark achievement that is also beset by backlogs, strategic use by the powerful, and doubts about whether it serves the democratic purposes it was meant to advance. Worthy (2017) adds that FOIA law carries symbolic value, in that governments gain from the promise of openness even as bold policies are weakened before they reach the statute book, and that FOIA's disruptive and democratic effects continue to cause controversy once the law is in operation. Meijer (2013) adds that transparency is constructed in interaction between governments and stakeholders, from strategic, cognitive, and institutional perspectives at once, rather than delivered by a single decision to disclose. This article sits inside that mixed and constructed legacy. It neither romanticizes FOIA as sunlight that disinfects nor dismisses FOIA as failure. Its mechanisms describe the middle that both romantic and dismissive accounts miss: a system that releases a great deal and still withholds the whole.
Bok (1982), in Secrets, treated concealment and revelation as a moral problem cutting across public and private life, with chapters on secrecy, power, and accountability, on secrets of state, and on whistleblowing and leaking. Bok supplies the normative stakes without requiring this article to moralize. Secrecy is power because it controls what others can know and therefore what they can do. The article's contribution is to show the organizational operations behind that power: not grand secrets, but fragments, tranches, and exemptions. Bok's attention to the position of the one kept out also supports the article's method. The requester's file is a legitimate vantage point on the system, not a biased sample to be discounted, provided its selection effects are named, which the method section does.
Berliner and coauthors (2021) provide the closest empirical complement to this article's argument. Studying several hundred thousand information requests to Mexican federal agencies, they found that requests from locales with higher governing-party vote shares received more favorable responses across multiple indicators of the nature and timing of responses, consistent with officials managing political risks in disclosure. Their finding matters here for a precise reason. It shows that disclosure responsiveness is patterned rather than neutral, even in a routine mode of citizen and government interaction with decentralized decisions by large numbers of officials. This article's claim is parallel but organizational rather than partisan: in a fragmented system, responsiveness is patterned by unit, fragment location, and request structure, and that patterning is itself a finding about how oversight works. Where Berliner and coauthors show who gets better answers, this article shows how answers are structured so that no one gets the whole answer.
Information as strategy
Crawford and Sobel (1982) developed the classic account of strategic information transmission. A better-informed sender sends a possibly noisy signal to a receiver who acts on it. In equilibrium, the sender partitions its private information and reports only which partition the observation lies in, and revelation is more informative when the sender's and receiver's preferences are more similar. Stated verbally, without the formalism and without importing the model, that logic maps directly onto the requester game. The agency system is a sender that partitions its file, by unit, by tranche, and by exemption, and the requester is a receiver who must act on the partition rather than the file. Preference divergence between sender and receiver predicts coarser partitions, which is to say less revelation. That logic is the game-theoretic ancestor of this article's decompiling mechanism and of Proposition 3, which predicts partial release as the equilibrium.
The article's advance on Crawford and Sobel is organizational, and it is the reason the article refuses to formalize. The sender in this article is not one mind but many units, so the partition is produced by structure rather than by a single sender's choice. A multi-unit sender does not fit the single-sender model without distortion. The article therefore uses the logic qualitatively, as the reasoning behind its mechanisms, and states its expectations in prose to be tested against a coded file. No equilibrium is computed and no payoff is estimated, in keeping with the article's rule against mathematics. The gain from this restraint is fit: the article describes the system it actually studied, rather than a simplified system that can be solved.
Organizations that decouple, legitimize, and mediate law
Institutional theory supplies the article's account of why formal openness and substantive visibility come apart. Meyer and Rowan (1977) argued that organizations adopt formal structures to signal conformity with institutional myths, while decoupling those structures from actual work. Bromley and Powell (2012) extended the argument to a world of proliferating formal commitments, in which decoupling takes new forms, including the separation of policy from practice and of means from ends. The article's mechanisms are decoupling operations at the inter-unit level. A unit can be formally open, with a records office, a published procedure, and a released tranche, while the system as a whole remains substantively closed, because the whole picture is never assembled in one place. Decoupling is therefore not a deviation in this account. It is the normal operation of a fragmented system, and the article's contribution is to name its records-specific forms.
Edelman (1992) showed how organizations mediate ambiguous law through symbolic structures: formal arrangements that signal compliance while leaving substantive practice largely untouched. FOIA compliance can function in exactly that way. Acknowledgment letters, extension notices, and partial productions are forms of openness that signal responsiveness while the complete record remains unassembled. Edelman gives the article a way to describe that pattern without alleging bad faith, since symbolic structures are normal organizational responses to ambiguous law rather than conspiracies. Suchman (1995) showed how organizations manage legitimacy strategically and institutionally, by conforming to expectations in ways audiences can see. Partial visibility is legitimacy management in that sense. Releasing a large tranche while withholding approval materials demonstrates responsiveness and protects the sensitive fragment at the same time. The mechanism is therefore not merely obstruction. It is an organization maintaining its standing with two audiences at once, the requester who demands openness and the interests that benefit from continued opacity.
Scott (2014) provides the broader institutional frame, in which institutions rest on regulative, normative, and cultural-cognitive elements together. The article's mechanisms operate across all three: rules that permit exemption and extension, professional norms of caution inside units, and a taken-for-granted fragmentation in which no unit expects itself to hold the whole file. Naming all three keeps the article from reducing not-knowing to a single cause, whether legal, cultural, or structural. Abbott, Levi-Faur, and Snidal (2017) add the regulatory intermediary frame, in which third parties stand between regulators and targets and shape what regulators can know and do. The requester in this article is an unacknowledged intermediary of a particular kind: not hired by the regulator and not employed by the target, but positioned by the FOIA system itself to assemble fragments the system does not assemble. That positioning is the article's distinctive claim about who does the work of oversight when the system does not.
Who can play, and at what cost
Galanter (1974) distinguished repeat players from one-shotters, showing that repeat players enjoy expertise, relationships, and the ability to play for rules rather than outcomes. The requester game is a repeat-player game. The author, as a repeat requester with a chronology, a publication platform, and appeal experience, can persist where a one-shot requester drops. Proposition 2 follows from that distinction, and it carries a warning about the evidence: the file overrepresents what persistence can extract, and it must be read as a record of that extraction rather than as a typical requester's experience. Zemans (1983) made the parallel point about legal mobilization, which depends on resources, organization, and the perceived efficacy of turning to law. FOIA mobilization follows the same logic. Requesters file and persist when they have resources and a reason to believe the record matters. The game is therefore selective twice over: first by who files, and then by who persists. That double selection is a limit on generalizability, and it is also a finding, since it explains why the system's not-knowing is rarely challenged by those it affects most.
The burden and oversight literatures complete this part of the argument. Herd and Moynihan (2018) showed that administrative burden functions as policy by other means, since the costs imposed on claimants determine who receives benefits. The same logic applies to requesters: delay, re-filing, and appeal are costs that determine who receives records. Lipsky (1980) located those costs where they are actually imposed, in frontline discretion, which is where delay and splitting decisions are made in practice. McCubbins and Schwartz (1984) explained why the system waits for the requester in the first place: fire-alarm oversight depends on affected parties pulling the alarm, rather than on continuous monitoring. The requester is the alarm, and this article studies what happens when the alarm must also assemble the evidence. Cohen, March, and Olsen (1972) add the garbage-can insight that, in organized anarchies, problems, solutions, and participants arrive independently and are coupled by timing rather than by design. A multi-unit records system behaves in that way: a request arrives at one unit, a referral arrives at another, an extension arrives at a third, and the coupling of those streams into a single answer is the exception rather than the rule.
Regulation, capture, and why knowledge control matters
Baldwin, Cave, and Lodge (2012) treat regulation as a matter of strategy, rules, enforcement, and evaluation together. This article's question is a regulation question for that reason. Licensing oversight depends on the knowledge the system holds. A regulator that cannot assemble its own file cannot evaluate compliance, whatever its rules say. Not-knowing is therefore not only a transparency problem. It is a regulatory capacity problem, and it belongs in a regulation journal for that reason. Levi-Faur (2011) and the regulatory governance literature extend the point: contemporary regulation operates through networks of public and private actors, and knowledge in such networks is distributed by design. The question is not whether knowledge is distributed, but whether any actor in the network is positioned to reassemble it, and what happens when no one is.
Stigler (1971) and the capture literature, as synthesized by Carpenter and Moss (2014), ask who influences regulators. This article asks a prior question: who controls what regulators can know? Knowledge control precedes influence. A target that benefits from fragmentation need not capture anyone. It need only benefit from a system that never assembles the file. That claim is structural and comparative, and the article makes no claim that any target did so in the cases studied. The point is that fragmentation, whatever its cause, has distributive consequences, and a theory of regulation that ignores it will mistake a thin file for a clean record. Shavell (1984) and the enforcement literature add the complementary point that enforcement design assumes an informed enforcer. Where the enforcer's information is fragmented, enforcement will be selective in ways the statute does not predict, and the pattern of selectivity will track the pattern of fragmentation. That prediction connects this article's typology to the study of enforcement outcomes, which is where future tests should look.
Taken together, these literatures call for exactly the study this article proposes. Roberts shows secrecy persisting after FOIA laws. Pozen shows FOIA's limits as a strategy. Bok shows secrecy as power. Crawford and Sobel show partial revelation as equilibrium. The organizational literature shows forms decoupling from substance. The mobilization literature shows who can play. The regulation literature shows why the resulting ignorance matters beyond transparency. The article's five mechanisms are where those arguments meet a real file, and its propositions are the test of whether they meet there in the way the literatures predict.
Theory: Five Mechanisms and the Requester Game
Ignorance studies gives this article its starting distinction. McGoey (2012) argued that ignorance in organizations is often strategic, in the sense that not-knowing is produced and maintained by regular arrangements rather than by individual failure to learn, and Gross and McGoey (2015) collected the field's core claim that ignorance is a social product with its own organization. Proctor and Schiebinger (2008) named the study of that production agnotology, and Alvesson, Einola, and Schaefer (2022) develop the dynamics of wilful ignorance in organizations, in which sequestering, by which knowledge is kept apart, and decompiling, by which knowledge is broken into pieces that no one reassembles, operate as key dimensions. This article takes those two operations, adds three more that the records file forced into view, and defines all five so that each can be coded in a file. The definitions are deliberately behavioral. Each mechanism is identified by what happens to a request, not by why anyone acted, so that the typology can be applied without attributing motive.
Sequestering
Sequestering is the holding of a fragment in one unit while the requester is told, by that unit or by another, that the fragment does not exist, cannot be found, or belongs elsewhere. Its observable implication is a non-response that is later contradicted, qualified, or narrowed by a subsequent release, or a pair of non-responses from the same department that answer different questions and are mistaken for one answer. The mechanism is coded conservatively. A non-response on its own is indeterminate. It becomes sequestering only when a later release proves that the fragment existed at the time of the non-response, or when two answers from the same department can be shown to address different questions. Without that proof, the article reports absence, not concealment, and the distinction is maintained throughout the evidence section.
Decompiling
Decompiling is the splitting of a record into tranches, slices, and sub-requests so that no single production contains the whole. Its observable implication is a large production that arrives in parts, with the requester required to narrow, re-file, or appeal in order to see successive slices. Decompiling differs from ordinary volume management in its effect, not its intent: whether the splitting is driven by capacity, by caution, or by design, the requester receives partitions rather than a file, and must supply the labor of reassembly. The mechanism is identified by the partition structure itself: a production that can only be understood by combining tranches that arrive at different times and under different cover letters. Crawford and Sobel's partition logic, stated verbally, is the game-theoretic reason decompiling is stable: a sender that partitions its information gives the receiver less than the whole, and the receiver must pay in time and persistence to narrow the partitions.
Delay
Delay is the extension of time beyond the statutory or expected window, including incorrect extensions that must themselves be corrected. Its observable implication is elapsed time coded from the file: the gap between filing and response, the number and length of extensions, and the correction of extension letters. Delay matters in this typology not as rudeness but as cost. Every extension raises the price of persistence, and the requester game turns on whether the requester pays it. Delay is therefore coded by duration and by correction, not by characterization. An extension letter that is issued and then corrected on the same day is evidence of two things at once: a system that moves slowly, and a system that monitors its own letters closely enough to correct them. Both facts belong in the record, and the article reports both.
Referral loops
A referral loop occurs when a request is sent from one unit to another, and then back again, or onward to a unit that refers it further, without any unit accepting responsibility for assembling the answer. Its observable implication is a chain of referrals in the chronology in which the requester's question returns to a unit that has already answered it, or in which an oversight body refers a complaint back to the subject of the complaint. Referral loops are the mechanism that most directly expresses fragmentation, because they show that the system's map of itself does not include a unit charged with the whole. The loop is identified by the chain, not by any single referral, since a single referral to the correct holder is ordinary administration. It is the return, or the indefinite onward passage, that marks the mechanism.
Exemption and partial visibility
Exemption with partial visibility occurs when a system releases a conclusion while withholding the record behind it, under a lawful exemption. Its observable implication is a pair: a public output, such as a credential listing showing a license in good standing, alongside a withheld input, such as the approval materials that would show how that standing was determined. The mechanism is defined to include lawful acts. Exemptions are part of FOIA design, and the article makes no claim that invoking one is improper. The theoretical point is about effect: a system that routinely pairs visible conclusions with withheld premises produces a public that can see outcomes but cannot check them. Partial visibility is therefore the mechanism that most directly connects this article to legitimacy research, since it describes how a system can appear open and remain uncheckable at the same time.
The requester game
The five mechanisms do not operate in a vacuum. They operate in a game between a requester who wants a complete picture and an agency system that is not a single player but a set of units, each of which wants to close requests at acceptable cost. Behind them sits the regulated target, whose interests are affected by what becomes visible, but the core game is between requester and system, and the target enters mainly through the political cost that visibility carries. The game has three stable features that the evidence section will illustrate. First, full release is rare, because it is costly to the system and because no unit can release what another unit holds. Second, full denial is also costly, because it invites appeal and publicity. Third, the stable middle is a large partial production, a set of non-responses from units that hold fragments, and a requester who keeps going because the file is now too valuable to abandon.
Stated without any matrix, the expectations are these. Narrowing plus persistence yields a small complete picture at low cost to both sides when the system releases, and slow accumulation with requester fatigue when it resists. Appeal plus publicity yields faster release on appealed items and sets a precedent when the system releases, and higher political cost with a longer timeline when it resists. Dropping closes the file at no further cost either way, and leaves any hidden fragment hidden with no appeal record. These are qualitative expectations to be tested against the coded file, not payoffs measured from data. No formal mathematics is used anywhere in this article, and no payoff table is presented. The game is presented as analytic reasoning in prose, in the tradition of game-theoretic reasoning about institutions, and its value lies in the predictions it generates rather than in any computation it performs.
The game also predicts its own selection effects, and the article states them plainly. Because persistence is costly, the requesters who obtain the most complete records will disproportionately be repeat players with resources, platforms, and appeal experience. Because dropping is cheap, the files that close early will disproportionately belong to one-shot requesters. Any study built on a persistent requester's file, including this one, therefore overrepresents what persistence can extract. The propositions in the method section are designed with that bias in view: they predict patterns that should hold across requesters, not merely outcomes for this requester, and they can be tested on files the author has never touched.
Evidence: The Illinois File, Coded by Mechanism
The Illinois evidence comes from the author's dated chronology and response file: more than 425 entries from June 2023 to September 2026, covering requests, responses, referrals, extensions, productions, denials, and appeals across multiple state units concerned with cannabis regulation. The evidence is presented below organized by mechanism, in prose chronology rather than in tables, with dates given where the chronology fixes them. Two cautions govern the whole section. First, a non-response is coded as indeterminate unless a later release proves that the fragment existed. The article never upgrades a non-response into a finding of concealment. Second, the article reports what the records show and labels inference as inference. Where the chronology records a letter, a listing, or a production, the article says so. Where it offers an interpretation, it says that instead.
Sequestering in the file
The clearest sequestering pattern in the file concerns the agriculture department, which gave non-responses on two distinct requests that must not be collapsed into one. On a referral question, the department answered on September 21, 2026 that it held no records of a referral. On a transfer question, an earlier answer stated that it held no responsive records concerning a transfer. Those are two different questions with two different answers, and treating them as a single refusal would misstate the record. Under the article's conservative coding rule, each non-response stands as indeterminate on its own, since no later release in the file, as of the dates covered, proves that the fragments existed at the time of the answers. The mechanism is therefore illustrated here in its weak form: fragments addressed by different questions to the same department, answered separately, with no unit assembling the two answers into a single account of what the department held. That weak form is theoretically important, because it shows that sequestering does not require contradiction. Separation alone, across questions and across time, is enough to leave the requester without the whole.
Decompiling in the file
The largest production in the file came from the attorney general's office: a multi-thousand-file production released in tranches, with the requester narrowing a sub-request to a slice of files in order to obtain a usable set. The production is the article's central decompiling instance, because its structure is exactly the partition structure the theory predicts. No single tranche contained the file. Each tranche arrived under its own cover, and the requester's narrowing was the labor that made any part of the production usable. The article draws no conclusion about why the production was structured in tranches, since the records do not show motive, and capacity, caution, and design are all consistent with the observed structure. What the records do show is effect: the requester received partitions, paid in time and narrowing to combine them, and still held only a slice of the whole. That effect is the mechanism, and it is sufficient for the article's purpose, which is to name the operation and its observable implication rather than to explain its intention.
Delay in the file
Delay appears throughout the chronology as elapsed time, and one episode illustrates its coding with particular clarity. On September 24, 2026, the licensing agency issued extension letters that proved incorrect and were corrected on the same day. The episode is coded by duration and by correction, not by characterization. The incorrect letters extended time that the corrected letters then withdrew, and the requester was required to track both versions in order to know where the request stood. The article reports both facts, because both are theoretically relevant. A system that issues incorrect extensions and corrects them on the same day is a system that moves slowly and monitors itself closely at the same time, and any account that reports only the slowness or only the correction would misstate the record. Delay elsewhere in the file follows the same pattern: extensions that lengthen the timeline, followed by responses that arrive in part, with the requester carrying the cost of waiting and of tracking.
Referral loops in the file
On September 18, 2026, the inspector general's office referred a matter back to the subject agency. That referral is the article's referral loop instance, because it returns the question to the unit whose conduct or record is at issue, rather than to a unit positioned to assemble an independent answer. The article makes no claim about the propriety of that referral, which may reflect jurisdiction, procedure, or practice that the records do not fully show. The theoretical point is structural: when an oversight body refers a matter back to its subject, the requester's path loops rather than advances, and the requester must begin again with the unit that has already answered. The chronology shows that loop as one link in a longer chain of referrals across units, in which each referral was individually reasonable and the chain as a whole produced no assembled answer. It is the chain, not any single referral, that marks the mechanism, and the article codes it accordingly.
Exemption and partial visibility in the file
On September 24, 2026, the licensing agency's public credential listing showed two licenses active and in good standing, with August 31, 2026 issue dates, while the approval materials behind those licenses were withheld under exemption. That pair, a visible conclusion alongside a withheld premise, is the article's partial visibility instance, and it is presented without any claim that the exemption was improperly invoked. Exemptions are part of FOIA design, and the records do not show that this one was misapplied. The theoretical point is about effect, not propriety. A public that can see that licenses are in good standing but cannot see the approvals on which that standing rests can observe outcomes but cannot check them. The same pattern appears in miniature wherever a production releases conclusions, summaries, or listings while withholding the underlying record under exemption. The mechanism is therefore the most common in the file and the most consequential for legitimacy, since it describes a system that is open in its outputs and closed in its premises.
How the mechanisms compound
The mechanisms do not operate independently, and the file shows them compounding in ways that matter for the theory. Delay amplifies decompiling, because each tranche arrives after its own wait, and the requester must hold earlier tranches in mind, or in files, while waiting for later ones. A production split into parts and delivered slowly costs more than the same production delivered at once, and the extra cost falls entirely on the requester. Referral loops produce delay as a byproduct, because each referral restarts a clock in a new unit, and the requester must track not one timeline but several. Partial visibility caps what persistence can win, because even a requester who persists through every delay and combines every tranche still faces withheld premises behind visible conclusions. The game is therefore not five separate games but one game with five moves, and the requester's cost is the sum of all of them. That compounding is why the article's central prediction concerns the whole file rather than any single response. A system that delayed once, or split once, or exempted once would still be describable as an ordinary bureaucracy. A system in which delay, splitting, referral, and partial visibility recur across units and across years produces a different outcome: a requester who holds a great deal of paper and still cannot write the sentence that says who approved what, on what record, and when. The file's length is part of the finding. More than 425 entries is not only evidence of persistence. It is evidence of how much labor the system required before it yielded even a partial picture.
The suit record as dated evidence
A civil action in the state trial court surrounds parts of the file, and its record enters this article only as dated evidence. The action's filing prompted a round of requests, its discovery produced records that also entered the public file, its testimony forms part of the dated record, and it reached a jury verdict on a breach of contract claim in July 2026. The article makes no claim about the merits of the action and draws no legal conclusion from it. That restraint is deliberate, and it follows from the article's scope: the suit is evidence of what records were generated, prompted, or made public, and of when, not evidence for any proposition about liability, fault, or rights. Two features of the suit record matter for the theory. First, the suit generated records that the FOIA file alone would not contain, which means that the requester game and the litigation game interact: each proceeding produces fragments that the other proceeding must then assemble. Second, the suit's verdict provides a dated endpoint against which the records chronology can be read, showing which questions remained unanswered in the public file even after a trial had taken place. Both features support the article's central claim without requiring any judgment about the suit itself. A system can try a case and still leave the public file fragmented, because a trial assembles a record for a courtroom, not a record for the public.
The Kentucky Comparison
Kentucky provides the article's second site, based on a FOIA chronology from December 2025 to October 2026. Requests were directed to the medical cannabis office within the health and family services cabinet unless stated otherwise, with additional requests to the finance cabinet's general counsel and to the governor's office. The requester is the same in both settings, namely the author, so the comparison tests portability across settings, not across requesters. The Kentucky episodes below are coded with the same rules as the Illinois file, under the same conservative rule for non-responses. The section establishes no conduct by any agency or licensee, claims no motive, and draws no legal conclusion. It describes dated requests and responses, and what those dated sequences show about how a second fragmented system produces partial knowledge.
The sequence begins in December 2025 with portal registration as a media requester and a first request in January 2026. On January 15, 2026, an email from the medical cannabis office's general counsel stated that no records existed in response to the inquiry and invited direct contact. On January 23, 2026, an email from the deputy executive director provided records responsive to the same inquiry. Coded under the article's strict rule, that pair is sequestering in a stronger form than the Illinois instance, because a later release proves that records existed at the time of the earlier non-response. The article draws no conclusion about why the first answer was given. It records the dated contradiction, which is sufficient for the mechanism: a fragment held in one place, denied in one message, and provided in another, with the requester carrying the cost of reconciling the two.
Decompiling appears in Kentucky in two forms. First, the file splits across three units, so no single office supplies a complete account: the medical cannabis office holds program and ownership records, the finance cabinet's general counsel holds review records concerning an inspector general report, and the governor's office holds a reconciliation request and its response. A June 2026 sequence illustrates the split: a request to the finance cabinet's general counsel asked whether an inspector general report had taken into account records later provided by the medical cannabis office, and the June 22 response stated that the finance cabinet's review had been limited to materials it held. Each unit thus answers from its own holdings, and no unit assembles the holdings of the others. Second, responses shift channel across the chronology, from portal filings to email responses, so the requester must assemble a record across channels as well as across units. That channel shift is coded here as a decompiling variant, described as a shift in channel, not as an intention to conceal.
Delay appears as extensions and as elapsed time around an under-review reply. An April 2026 request concerning ownership information drew an extension at the end of April from the director of enforcement and compliance. A July 31 request for internal reporting records drew an extension on August 7. On August 26, a sequence of emails concerned a response deadline, including a message recall and a reply from general counsel stating that the inquiry was under review and that counsel would circle back as soon as possible. By October 10, no substantive response from the medical cannabis office and public affairs office had been received following that under-review reply. Delay is coded here by duration and by the under-review reply itself, not by characterization, and the requester carries the cost of waiting and of tracking each timeline.
Partial visibility appears in the May 2026 ownership sequence. A May 14 response provided ownership information with redactions. A May 22 response stated that the office held management agreements, withheld them on privacy grounds, and stated that an internal review had found management agreements and therefore no violation. A May 29 response provided transfer records bearing on the same ownership question. The article reports those three responses as dated response content, attributed to the responses, and draws no conclusion about the underlying ownership question or about whether any response was correct. The theoretical point is the same as in Illinois: a visible conclusion alongside withheld premises leaves the requester able to observe outcomes but unable to check them from the record provided.
Two further patterns are handled as boundary cases rather than forced into the typology. On August 21, a response stated that the office had not referred, investigated, or sent the matter to the inspector general before the report at issue, but had reviewed and discussed potential compliance issues concerning ownership, which it stated it had authorized. That statement is coded here as a non-referral observation, not as a referral loop, because no chain returning the question to a unit that had already answered it is documented in the chronology as presented. Separately, two no-records responses are coded as indeterminate under the conservative rule: an August 24 response from the finance cabinet's general counsel, after a diligent search, finding no responsive records concerning correspondence about correcting the inspector general report, and a September 15 response from the governor's office showing no records in response to a request for a reconciliation of stated values with department actions. No later release in the chronology proves that those fragments existed, so the article reports absence, not concealment. No referral loop chain appears in the Kentucky chronology as presented, and the article does not claim one.
Taken together, the Kentucky chronology supports Proposition 5 in a stronger form than correspondence alone could: four mechanisms appear in coded form, with the same directional pattern as Illinois, under the same rules. The comparison remains limited in one decisive way: the requester is the same in both settings, so it demonstrates portability across settings, not across requesters, and a multi-requester study remains the next test. The Kentucky episodes are drawn from the author's chronology and response file, coded under the conservative rules stated in the method section, and the coding protocol is written so that a second researcher can replicate the classification.
Method and Propositions
The article uses comparative process tracing, following Bennett and Checkel (2015) and Yin (2018), to connect mechanisms to outcomes through dated sequences rather than through correlation alone. Process tracing is appropriate here because the argument is about sequences: a request arrives, a referral follows, an extension intervenes, a tranche arrives, and the requester narrows or appeals. Each link in the sequence is dated in the chronology, and the article tests the theory by asking whether the links appear in the order and form the mechanisms predict. The Illinois file supplies the primary sequence, the Kentucky FOIA chronology supplies a comparative sequence, and the suit record supplies dated anchors against which both sequences can be read. The design is comparative in the limited sense that two settings are examined with the same coding rules, not in the sense that variables are controlled across a sample. That limit is stated here so that the article is not mistaken for a controlled comparison.
Coding followed rules stated in advance. Each episode in the chronology was classified by mechanism using the behavioral definitions in the theory section, with no reference to motive. A non-response was coded as indeterminate unless a later release proved that the fragment existed. A production was coded as decompiling only when its partition structure required the requester to combine tranches arriving at different times. A referral was coded as a loop only when the chain returned the question to a unit that had already answered it, or passed it onward without any unit accepting responsibility for assembly. An exemption was coded as partial visibility only when a visible conclusion and a withheld premise could be paired in the record. Episodes that did not meet a rule were left uncoded, and the article reports that restraint rather than filling gaps with interpretation. A disconfirming-evidence procedure was applied throughout: for each mechanism, the article sought episodes that would contradict the mechanism's observable implication, such as full releases, single-unit assemblies, or referrals that ended in a complete answer, and reports the pattern those episodes form rather than suppressing them. The author's position as requester is disclosed as a limitation, and the coding protocol is specified in full so that a second researcher can replicate the classification on the same file and so that other researchers can apply it to theirs.
The article's propositions are stated so that they can fail. Proposition 1: as the number of units holding fragments of a record increases, the completeness of any single response decreases. Proposition 2: requesters who persist through appeal and narrowing obtain more complete records than requesters who file once and drop, controlling for request scope. Proposition 3: the modal outcome of a contested multi-unit request is a large partial production, not full release and not full denial. Proposition 4: delay, referral, and splitting cluster on requests that span multiple units, rather than distributing randomly across requests. Proposition 5: the mechanisms appear in a second regulatory setting with the same directional pattern, coded with the same rules. Each proposition names its falsifier. If single-unit requests show the same incompleteness as multi-unit requests, Proposition 1 fails. If persistent and one-shot requesters obtain equally complete records, Proposition 2 fails. If full releases or full denials dominate contested multi-unit requests, Proposition 3 fails. If delay, referral, and splitting distribute randomly, Proposition 4 fails. If no mechanism appears in a second setting, Proposition 5 fails. The article's contribution does not depend on all five propositions surviving. It depends on the typology being stated clearly enough that its failure can be seen.
Three limits bound the article. First, the evidence comes from a single requester's file, and that requester is the author. Repeat-player effects, publication incentives, and appeal experience all distinguish this requester from a typical one, and the article's findings must be read with that selection in view. Second, the Kentucky comparison rests on a FOIA chronology from a single requester, and it can support portability claims across settings but not across requesters. Third, the article studies a single policy domain, cannabis licensing, in two states, and its mechanisms may operate differently in domains with different stakes, different unit structures, or different exemption regimes. Those limits do not defeat the typology, but they do confine it, and the article's claims are confined accordingly.
Discussion: Objections, Limits, and What Would Falsify the Argument
A strictly academic article must survive its reviewers, so this section states the strongest objections and answers each, condensed from the fuller defense that shaped the article's design. The first objection is that this is one requester's anecdote, not a generalizable finding. The reply is that the article's contribution is a typology and a mechanism, not a prevalence estimate. Typologies are judged by portability: can another researcher code another records file with these five mechanisms and get a determinate answer? The Kentucky comparison is the first portability test, and the propositions are stated so that a second file can confirm or refute them. Generalizability to all requesters is the next study, and the article names it as such rather than claiming it.
The second objection is that strategic language smuggles in intent. The reply is the literature's own distinction: ignorance studies treats not-knowing as regular and systemic, not necessarily intentional. The article claims structure, never motive. Its strictest coding rule enforces that: a non-response is coded as indeterminate unless a later release proves that the fragment existed. Where intent cannot be shown, the article says so. The third objection is that agencies followed the law, since exemptions, extensions, and referrals are lawful. The reply concedes the point entirely. The mechanisms include lawful acts. The article studies effects, not legality. A lawful withholding that leaves the public with a conclusion and no record behind it is still a visibility gap, and naming it as such is not an accusation.
The fourth objection is that the suit contaminates a records study, or that litigant authorship biases it. The reply is structural: positionality stated up front, conservative coding, the suit used only as dated evidence with no merits claim, and unknowns reported as unknowns even where a favorable inference is available. If those safeguards fail in review, the article fails, and the article states the safeguards now so they cannot be added retroactively. The fifth objection is that game theory without mathematics is not game theory. The reply concedes the label point. The article uses game-theoretic reasoning, meaning strategies, costs, and equilibria stated in prose, and it does not present a formal model. Its contribution is the typology and the mechanism, not the model. Formalization is named as future work, possibly with a co-author, and no mathematics is faked to fill the gap.
The sixth objection is that the article duplicates the author's existing preprints on multiple voices and on the state that knows and does not know. The reply is that those papers are the base, and this article synthesizes them into something none of them contains: a five-mechanism typology, an equilibrium account, and a comparison. Self-citation is kept to what the synthesis requires, and the new contribution is stated explicitly so reviewers can judge whether it is enough. The seventh objection is that not-knowing is simply backlog and incompetence. The reply is a testable difference. Backlog predicts delay distributed roughly randomly across requests. This article predicts patterned delay, referral, and splitting correlated with fragment location and unit count. If the coded file shows random patterns, the theory fails, and the article says that in advance.
The article's implications follow from its mechanism account. For regulation research, the implication is that oversight capacity should be measured not only by rules, budgets, and staffing, but by assemblability: whether any actor in the system can combine the fragments into a checkable record. A licensing regime with strong rules and an unassemblable file is weaker, in a specific and measurable way, than its rules suggest. For institutional research, the implication is that decoupling has a records-specific form that existing accounts underdescribe. Formal openness and substantive visibility are distinct, and the gap between them can be studied through request files with the same rigor applied to budgets and decisions. For transparency reform, the implication is modest and practical. Reforms that add disclosure duties without designating an assembling unit will reproduce the pattern this article describes, because each new duty will be met in fragments. The reform question is therefore not only what must be disclosed, but who must assemble it, and what happens when no one does.
A further implication concerns who the requester game leaves out. Galanter's repeat-player distinction and Pozen's inegalitarian point converge here. The requesters who can afford the game's costs, in time, expertise, appeal experience, and publication capacity, are the requesters who obtain the most complete records. The requesters who cannot afford those costs drop early, and their files close with the fragments still hidden and no appeal record. The system therefore learns most about its own fragmentation from the requesters least representative of the public, and learns least from the requesters most affected by it. That selection effect should temper any reading of this article's file as typical, and it should also shape reform. Lowering the cost of persistence, through assembling units, through consolidated responses, or through appeal procedures that do not require the requester to reassemble the file, would change the game's equilibrium more reliably than adding disclosure duties that each unit meets in fragments. The article's reform implication is therefore institutional rather than hortatory: change who must assemble, and the mechanisms lose their stability.
Conclusion
This article began with a simple question: how does a multi-agency regulatory system sustain not-knowing as a stable outcome? Its answer is that fragmentation produces partial knowledge through five regular mechanisms, that the requester game stabilizes those mechanisms by making partial compliance cheap for the system and persistence expensive for the requester, and that the resulting equilibrium is visible in a dated Illinois records file, in a Kentucky FOIA chronology coded with the same rules, and in the suit record that surrounds parts of the file. The argument is structural throughout. It claims no motive, alleges no misconduct, and draws no legal conclusion. It claims instead that a system organized in fragments will produce fragmentary knowledge, and that the fragments will be patterned rather than random.
The article's contribution is therefore threefold. It gives regulation research a typology of not-knowing that can be coded in any records file. It gives institutional research a records-specific account of decoupling, legitimacy management, and intermediary work. And it gives future empirical research five propositions with stated falsifiers, a coding protocol, and a second-site design that can be extended to a full file. If those propositions fail against other files, the typology should be revised or abandoned, and the article will have served its purpose by making that failure visible. If they survive, then not-knowing deserves a place in the study of regulation alongside capture, burden, and enforcement, as one of the regular ways in which oversight is weakened without anyone deciding to weaken it. Either outcome advances the conversation this article joins, which is the proper test of a theory-building article.
References
Abbott, K. W., Levi-Faur, D., and Snidal, D. (2017). Theorizing regulatory intermediaries: The RIT model. Annals of the American Academy of Political and Social Science, 670(1), 14-35.
Alvesson, M., Einola, K., and Schaefer, S. M. (2022). Dynamics of wilful ignorance in organizations. The British Journal of Sociology, 73(4), 839-858. https://doi.org/10.1111/1468-4446.12963
Baldwin, R., Cave, M., and Lodge, M. (2012). Understanding Regulation: Theory, Strategy, and Practice (2nd ed.). Oxford University Press.
Bennett, A., and Checkel, J. T. (Eds.). (2015). Process Tracing: From Metaphor to Analytic Tool. Cambridge University Press.
Berliner, D., Bagozzi, B. E., Palmer-Rubin, B., and Erlich, A. (2021). The political logic of government disclosure: Evidence from information requests in Mexico. Journal of Politics.
Bok, S. (1982). Secrets: On the Ethics of Concealment and Revelation. Pantheon Books.
Bromley, P., and Powell, W. W. (2012). From smoke and mirrors to walking the talk: Decoupling in the contemporary world. Academy of Management Annals, 6(1), 483-530.
Carpenter, D., and Moss, D. A. (Eds.). (2014). Preventing Regulatory Capture: Special Interest Influence and How to Limit It. Cambridge University Press.
Cohen, M. D., March, J. G., and Olsen, J. P. (1972). A garbage can model of organizational choice. Administrative Science Quarterly, 17(1), 1-25.
Crawford, V. P., and Sobel, J. (1982). Strategic information transmission. Econometrica, 50(6), 1431-1451. https://doi.org/10.2307/1913390
Edelman, L. B. (1992). Legal ambiguity and symbolic structures: Organizational mediation of civil rights law. American Journal of Sociology, 97(6), 1531-1576.
Galanter, M. (1974). Why the "haves" come out ahead: Speculations on the limits of legal change. Law and Society Review, 9(1), 95-160.
Gross, M., and McGoey, L. (Eds.). (2015). Routledge International Handbook of Ignorance Studies. Routledge.
Herd, P., and Moynihan, D. P. (2018). Administrative Burden: Policymaking by Other Means. Russell Sage Foundation.
Illinois Freedom of Information Act, 5 ILCS 140.
Levi-Faur, D. (Ed.). (2011). Handbook on the Politics of Regulation. Edward Elgar.
Lipsky, M. (1980). Street-Level Bureaucracy: Dilemmas of the Individual in Public Services. Russell Sage Foundation.
McCubbins, M. D., and Schwartz, T. (1984). Congressional oversight overlooked: Police patrols versus fire alarms. American Journal of Political Science, 28(1), 165-179.
McGoey, L. (2012). Strategic unknowns: Towards a sociology of ignorance. Economy and Society, 41(1), 1-16.
McGoey, L. (2012). The logic of strategic ignorance. British Journal of Sociology, 63(3), 533-576. https://doi.org/10.1111/j.1468-4446.2012.01424.x
Meijer, A. (2013). Understanding the complex dynamics of transparency. Public Administration Review, 73(3), 429-439. https://doi.org/10.1111/puar.12032
Meyer, J. W., and Rowan, B. (1977). Institutionalized organizations: Formal structure as myth and ceremony. American Journal of Sociology, 83(2), 340-363.
Phelan, J. (2026). Reconciliation testing: A public-records methodology for auditing institutional policy coherence. ResearchGate preprint. https://doi.org/10.13140/RG.2.2.18066.62403
Phelan, J. (2026). When the state knows but the state does not know. ResearchGate preprint. https://doi.org/10.13140/RG.2.2.21422.06720
Phelan, J. (2026). When the state speaks with multiple voices. ResearchGate preprint. https://doi.org/10.13140/RG.2.2.16303.70565
Pozen, D. (2017). Freedom of information beyond the Freedom of Information Act. University of Pennsylvania Law Review, 165, 1097.
Pozen, D., and Schudson, M. (Eds.). (2018). Troubling Transparency: The History and Future of Freedom of Information. Columbia University Press.
Proctor, R. N., and Schiebinger, L. (Eds.). (2008). Agnotology: The Making and Unmaking of Ignorance. Stanford University Press.
Roberts, A. (2006). Blacked Out: Government Secrecy in the Information Age. Cambridge University Press.
Schudson, M. (2015). The Rise of the Right to Know: Politics and the Culture of Transparency, 1945-1975.
Scott, W. R. (2014). Institutions and Organizations: Ideas, Interests, and Identities (4th ed.). Sage.
Shavell, S. (1984). A model of the optimal use of liability and safety regulation. RAND Journal of Economics, 15(2), 271-280.
Stigler, G. J. (1971). The theory of economic regulation. Bell Journal of Economics and Management Science, 2(1), 3-21.
Suchman, M. C. (1995). Managing legitimacy: Strategic and institutional approaches. Academy of Management Review, 20(3), 571-610.
Worthy, B. (2017). The Politics of Freedom of Information: How and Why Governments Pass Laws That Threaten Their Power. Manchester University Press.
Yin, R. K. (2018). Case Study Research and Applications: Design and Methods (6th ed.). Sage.
Zemans, F. K. (1983). Legal mobilization: The neglected role of the law in the political system. American Political Science Review, 77, 690-703.