Thesis
On Legibility
The concept of legibility has a contested intellectual history. Scott's Seeing Like a State (1998) used it critically: states impose legibility on complex social systems through simplification schemes (cadastral maps, standardized surnames, grid cities) that destroy the local, tacit, and practical knowledge they cannot capture. The resulting high-modernist failures (Soviet collectivization, Brasília's planned urbanism, Tanzanian villagization) demonstrated that legibility imposed from above can be a tool of authoritarian control rather than a public good.
This thesis uses the same term but examines a different problem: systems that resist external understanding through deliberate or structural opacity, where the failure mode is too little legibility rather than too much. The tension is real and the thesis must confront it directly. The question across both uses is the same: what can be made formally knowable about a system, at what cost, and for whose benefit?
Adjacent traditions have addressed fragments of this question, each contributing a structural insight that the thesis integrates. Akerlof's (1970) market for lemons demonstrated that information asymmetry degrades market quality by driving out high-quality goods; the thesis extends this to knowledge markets where disclosure quality determines whether external reasoners can distinguish sound systems from unsound ones. Jensen and Meckling's (1976) principal-agent theory identified the structural incentive for agents to misrepresent: operators benefit from opacity when their interests diverge from their stakeholders'.
Hurwicz (1960) and Myerson (1981) formalized the mechanism design problem: what institutional rules produce honest revelation when agents have private information and strategic incentives? Transaction cost economics (Williamson 1985) asks what makes organizational boundaries visible to contractual partners, and concludes that the cost of measuring what a partner is actually doing determines whether the transaction occurs at all.
Science and technology studies (Winner 1980; Jasanoff 2004) asks how technical artifacts encode and conceal political choices, and demonstrates that the design decisions embedded in infrastructure shape the political possibilities available to the communities that depend on it. Social epistemology (Longino 1990) asks what conditions must hold for knowledge claims to withstand collective scrutiny. Bovens (2007) defines accountability as requiring an actor, a forum, and an obligation to explain, and demonstrates that without structured information, the obligation remains formal rather than substantive. Each tradition addresses the same structural question from a different vantage: whether an external party can form reliable beliefs about a system it depends on but did not build.
Legibility is the name this thesis gives to that shared property. A system is legible to the extent that an external reasoner (human analyst, regulatory body, autonomous agent) can determine what the system claims to do, from publicly available structured evidence.
Two boundary conditions qualify the definition. First, Polanyi's (1966) observation that tacit knowledge resists formalization sets a ceiling on what legibility can capture: some system properties (the judgment calls embedded in governance decisions, the operational intuitions of experienced node operators, the social dynamics of core development teams) may be structurally resistant to structured disclosure. The thesis focuses on the remediable failures where legibility is achievable but absent due to design choices.
Second, Grossman and Stiglitz's (1980) information paradox applies: if legibility were costless and universal, the competitive advantage of analysis would disappear, removing the incentive to produce legibility in the first place. Legibility is a public good (non-rivalrous and partially non-excludable once produced), subject to the classic under-provision problem that affects all public goods. The question of who bears the cost of legibility production and who captures its value is an economic question the research program must address alongside the structural one.
The definition is tested against three independent bodies of work: a protocol disclosure standard (PDAS), a knowledge architecture framework for heterogeneous reasoners, and a series of infrastructure analyses. Across all three, the same five failure patterns recurred regardless of domain, system, or consumer class.
A methodological note on circularity: the definition of legibility and the taxonomy of its failures were developed iteratively rather than derived independently. The failure patterns were observed first, the unifying property was proposed to explain them, and the definition was then refined against the observed patterns. This is abductive reasoning (inference to the best explanation) rather than deduction from first principles, and the definition's value rests on its explanatory power across independent domains rather than on a priori necessity. The research program would be weakened if the failure patterns turned out to be artifacts of the analytical method rather than properties of the systems examined. The independent development of PDAS and the knowledge architecture framework, by different methods for different audiences, provides partial protection against this risk: the same five patterns appeared across work that was conducted without a shared theoretical frame. Treating them as instances of one property revealed compounding effects that domain-specific analysis obscured.
Observations
Four patterns that motivated the unifying concept
Across infrastructure analyses, systems substituted narrative for structural description (narrative substitution), restricted material facts to private channels (context privatization), published aggregate statistics instead of component-level evidence (granularity collapse), maintained documentation that described previous system states (temporal unbinding), and published claims without verification mechanisms (verification absence). When developing PDAS, the same five patterns appeared in the disclosure practices the standard was designed to address. When developing the knowledge architecture framework, the same patterns appeared in platform documentation and agent-facing metadata. The recurrence of identical failure patterns across three independent problem domains, three distinct classes of external reasoner, and three separate bodies of work suggests a shared underlying cause. DiMaggio and Powell's (1983) institutional isomorphism offers one explanation for why the same patterns appear so consistently: organizations facing similar environmental pressures converge on similar structural responses, including similar modes of opacity. The failure patterns may recur because the institutional incentives that produce them are themselves isomorphic across domains.
Across the infrastructure analyses, the quality of the analytical output was bounded by what each system made available for external examination, regardless of the analytical method applied. Sophisticated evaluation frameworks applied to systems with poor disclosure produced conclusions that were either superficial (restating what the system claimed about itself) or speculative (filling evidence gaps with assumptions). The same frameworks applied to systems with structured, verifiable disclosure produced substantive, falsifiable conclusions. The evaluator's methodology improved the analysis only up to the point where the evidence ran out. Past that point, additional analytical sophistication yielded diminishing returns. This pattern was consistent across multiple systems and across multiple evaluation dimensions (mechanism analysis, governance assessment, economic model evaluation). Gigerenzer's (2000) ecological rationality framework illuminates why: cognitive strategies perform well only when matched to the informational structure of their environment. An evaluation heuristic calibrated for structured evidence degrades when the environment supplies unstructured narrative instead, because the heuristic's performance depends on the environment's information architecture rather than on the heuristic's internal sophistication. Kahneman and Tversky's (1979) framing research identifies a complementary mechanism: when structured evidence is absent, evaluators default to whatever frames the available information supplies, and narrative-heavy disclosure systematically frames toward the system's preferred interpretation.
PDAS addressed disclosure failures for the protocol-disclosure domain specifically. The knowledge architecture framework addressed documentation failures for the heterogeneous-reasoner domain specifically. Each intervention was effective within its domain. Yet the failure taxonomy each addressed was the same taxonomy, arrived at independently. This suggested that the interventions were treating symptoms of a shared structural problem rather than addressing the problem itself. A protocol that adopted PDAS would improve its disclosure legibility without necessarily improving its documentation legibility or its governance legibility. The improvements were domain-bound because the framing was domain-bound. Star and Griesemer's (1989) concept of boundary objects explains both why domain-specific solutions work locally and why they fail to generalize: each solution functions as a local adaptation that coordinates action within one community of practice, while lacking the structural plasticity to serve communities it was designed without. The legibility concept, by contrast, operates as a higher-order boundary object that maintains coherence across all three domains precisely because it identifies the shared structural property rather than prescribing domain-specific remedies.
When the five failure patterns were analyzed as instances of legibility failure rather than domain-specific problems, a compounding structure became visible. Narrative substitution in a system's primary documentation (a disclosure failure) propagated into evaluation frameworks that relied on that documentation as input (an analytical failure), which propagated into governance decisions based on those evaluations (a governance failure), which propagated into integration decisions by systems that relied on the governance assessment (a knowledge architecture failure). Each domain-specific analysis captured one segment of this chain. The legibility lens captured the full propagation path and identified the originating failure. Callon (1998) and MacKenzie (2006) describe a related mechanism in financial markets: economic models do more than describe market behavior; they actively shape it, a process MacKenzie terms 'performativity.' Disclosure documents operate through a parallel logic. A primary document that substitutes narrative for structural evidence does more than fail to describe the system; it actively constitutes the informational environment within which all downstream reasoning occurs. The compounding structure is the primary analytical payoff of the unifying concept, and performativity theory explains why the compounding is directional: the originating document shapes the epistemic environment for every consumer that follows.
Three domains
Three domains, one structural problem
Each domain below involves a different system, a different external reasoner, and a different legibility failure. The structural pattern is identical: the external reasoner's ability to form reliable beliefs is bounded by the system's legibility.
Legibility failure
Evaluation frameworks relied on market narratives, token price, and self-reported metrics rather than structural mechanisms. The analyst lacked the evidence to determine whether a protocol's market position resulted from technical advantage, regulatory capture, network effects from early-mover timing, or token economics that created artificial switching costs. The mechanism remained opaque because the protocol's disclosure practices did not make it legible. Akerlof's (1970) market for lemons provides the economic frame: when quality signals are unreliable and buyers cannot distinguish high-quality systems from low-quality ones, the market rewards cheap signaling (narrative, brand, ecosystem size) over costly structural disclosure, driving a race to the bottom in information quality.
What we built
A corpus of structural analyses that isolate one mechanism per project and test whether the evidence supports the claim. Each article applies what Latour (1987) calls a trial of strength: the protocol's structural mechanism is subjected to empirical scrutiny against its public claims, and the discrepancy between claim and evidence constitutes the analytical finding. The method draws on Callon's (1986) four moments of translation (problematization, interessement, enrolment, mobilization) as an analytical vocabulary for tracing how protocols construct and maintain networks of support around structural positions that the evidence does not sustain.
The analytical method is itself a legibility operation. It takes a system whose mechanism is obscured by narrative and produces a structured account of how the mechanism works, what evidence supports it, and where the evidence is weak. The quality of each analysis was bounded by the protocol's disclosure practices. Bazerman's (1988) sociology of scientific texts illuminates the method's rhetorical dimension: analytical artifacts construct the objects they describe, and the choice of what to make visible (structural mechanisms) and what to subordinate (market narratives) is itself a consequential act of framing.
Legibility failure
Formal standards requiring public blockchain systems to disclose architecture, governance, operational risk, and assurance in a unified, release-bound framework remained absent. Disclosure was scattered across white papers, blog posts, GitHub repositories, and community forums. The information existed, but unstructured, unversioned, and unverifiable. The historical parallel is instructive: securities markets before the Securities Act of 1933 and the Exchange Act of 1934 faced the same structural problem. Brandeis's (1914) 'sunlight is said to be the best of disinfectants' reflected the Progressive Era recognition that voluntary disclosure regimes fail when issuers benefit from opacity. The clinical trial registration movement (ICMJE 2004) and environmental reporting under the Toxic Release Inventory (EPA 1986) confirm the pattern across domains: mandatory structured disclosure emerges when the costs of opacity become politically unsustainable.
What we built
PDAS: a disclosure and assurance standard that organizes around material facts rather than market narratives. Primitive-first (mechanism over category), release-bound (a specific configuration over a brand), and designed to produce machine-readable manifests alongside human-readable guides. The design draws on Timmermans and Epstein's (2010) analysis of how standards achieve authority: through precision of requirements (which constrains gaming), transparency of process (which builds legitimacy), and graduated conformance (which enables adoption across capability levels).
The standard is a legibility infrastructure. It specifies what a protocol must publish such that external reasoners can evaluate it from structured evidence rather than marketing materials. The design decisions (release-bound conformance, machine-readable manifests, explicit disclosure categories) were each responses to specific legibility failures observed across the infrastructure analyses. Bovens's (2007) accountability framework identifies the structural function: effective accountability requires a forum, an actor, and an obligation to explain. PDAS provides the structured information without which the obligation to explain remains formal rather than substantive.
Legibility failure
Knowledge systems were designed for a single consumer class. Documentation assumed a human reader. API specifications assumed a developer. Regulatory filings assumed a compliance officer. When a different class of reasoner arrived (autonomous agents, AI verification systems, AI-augmented human workflows), the knowledge required rebuilding from the new consumer's requirements. Each rebuild compounded technical debt and created periods where the knowledge layer failed the consumers it was designed to exclude. Hutchins's (1995) distributed cognition framework explains why single-consumer design fails at scale: cognitive work distributes across human and material artifacts in configurations that the system's designers did not anticipate, and knowledge architecture that assumes a fixed distribution breaks when the configuration shifts.
What we built
Eight consumer-agnostic design primitives for knowledge systems, each extending an established information architecture concept for the case where the consumer is unknown at design time. The DNCE framework renders these primitives for autonomous agents specifically. Context efficiency principles treat every reasoner's context constraint as a first-class architectural concern. The design approach extends Dewey's (1927) argument that effective publics form only when citizens have access to the material facts about conditions that affect them: the consumer-agnostic commitment ensures that the knowledge infrastructure serves any public that depends on the system, including publics that have yet to form.
Knowledge architecture is a legibility discipline. The primitives that make a system legible to an autonomous agent are the same primitives that make it legible to a human analyst, a regulatory body, and reasoner forms that have yet to emerge. The design question is universal; the rendering is consumer-specific. This convergence emerged independently from the same convergence observed between PDAS and the design principles. Luhmann's (1979) distinction between personal trust (based on familiarity with specific actors) and system trust (based on confidence in abstract mechanisms) maps onto the architecture: consumer-agnostic legibility enables system trust by providing the structural evidence that personal trust cannot scale to supply.
Argument
External operations are bounded by system legibility
A consistent empirical finding emerged across the research program: the quality of every external operation on a system (evaluation, trust assessment, governance participation, integration) correlated strongly with what the system made legible. Improving the sophistication of the external operation produced diminishing returns once the legibility ceiling was reached. The causal claim requires qualification: the finding is observational, drawn from comparative analysis across multiple systems and domains, and alternative explanations exist (systems with better disclosure may also have better engineering, better governance, or more mature operational practices, and the quality improvement may stem from those correlated properties rather than from disclosure alone). The structural argument for causation rests on specific cases where the same analytical framework, applied with equal rigor, produced substantively different outputs depending solely on what the system disclosed. Williamson's (1985) analysis of bounded rationality under information asymmetry, Ostrom's (1990) identification of information access as a precondition for commons governance, and Gigerenzer's (2000) demonstration that decision quality depends on the match between cognitive strategy and environmental information structure each provide independent theoretical grounds for expecting a causal relationship. This section presents the ceiling finding across four domains.
Observation 2 documented the empirical finding. The structural argument is stronger: every evaluation framework assumes access to the facts its criteria require. When those facts are inaccessible, the framework operates on substitute inputs (self-reported metrics, market data, community sentiment, analogies). The quality of the substitute input, rather than the quality of the framework, determines the output. This means the return on investment in evaluation methodology is bounded by the system's disclosure quality. Past the legibility ceiling, investment in better evaluators yields diminishing returns; investment in better disclosure does not. Toulmin's (1958) argument model makes the structural dependency explicit: an evaluation conclusion (claim) requires both data (the evidence examined) and a warrant (the justificatory principle connecting evidence to conclusion). When the system's disclosure is poor, the evaluator lacks adequate data, and the warrant connecting available evidence to evaluative conclusions becomes speculative rather than grounded. Improving the warrant's sophistication without improving the data's quality produces more elaborate speculation.
Trust requires evidence. Evidence requires structured access to material facts. Without legibility, trust depends on reputation, narrative, or social proof, all of which degrade under adversarial conditions. In infrastructure systems, this manifested as platforms with strong community trust but weak structural foundations: the trust was based on narrative (founding team reputation, growth metrics, ecosystem size) rather than on verified structural evidence. The structural analyses revealed cases where community trust ratings diverged substantially from structural assessment because the community's trust was based on information the system made legible (roadmap promises, partnership announcements) while the structural assessment required information the system did not make legible (operational concentration, governance mechanism specifics, economic model dependencies). Granovetter's (1985) embeddedness thesis explains the mechanism: trust in economic transactions flows through social networks, and when structural evidence is absent, network-embedded trust (who vouches for the protocol, which investors participated, which conferences featured the founders) substitutes for evidence-based trust. The substitution is stable under cooperative conditions but fragile under adversarial ones, because network-embedded trust lacks the verification infrastructure that evidence-based trust provides. Luhmann's (1979) distinction between personal trust and system trust sharpens the diagnosis: personal trust depends on familiarity with specific actors, while system trust depends on confidence in abstract mechanisms whose operation can be monitored. Complex infrastructure systems require system trust (users depend on mechanisms they cannot personally verify), but the opacity of those mechanisms forces participants back onto personal trust (the founding team's reputation, the investor roster, social proof from ecosystem size). Legibility is the structural precondition for the transition from personal to system trust. Without it, trust remains anchored to persons, and the trust ceiling corresponds to the limits of personal familiarity. Legibility is necessary for evidence-based trust, though alone insufficient: a system can be legible and still untrustworthy if its disclosed behavior is harmful.
Governance mechanisms assume that participants can observe the state of the system being governed, assess proposed changes against that state, and verify that approved changes were implemented correctly. In the infrastructure analyses, several systems maintained formal governance mechanisms that allowed stakeholders to vote on system changes, while the coordination layer that implemented those changes had no formal specification. The governance system could authorize a change, but no participant could verify whether the change was implemented as authorized, because the implementation layer was opaque. Ostrom's (1990) institutional analysis of commons governance identifies monitoring as one of the essential design principles for long-enduring common pool resource institutions: participants must be able to observe resource conditions and the behavior of other participants. When the system being governed is opaque, monitoring fails regardless of the governance structure's formal properties. Mechanism design theory (Hurwicz 1960; Myerson 1981) formalizes the problem: the designer seeks institutional rules that produce desirable outcomes even when participants have private information and strategic incentives. The fundamental insight is that information structure determines which outcomes are achievable: no mechanism can elicit honest revelation from agents whose private information is structurally inaccessible. Black's (2002) analysis of responsive regulation extends the point: regulators calibrate their enforcement intensity based on observed compliance, but when the system's compliance state is illegible, the regulatory response defaults to either blanket permissiveness or blanket restriction, both of which are suboptimal. Habermas's (1984) conditions for rational discourse (inclusion of all affected parties, equal voice, absence of coercion, willingness to revise beliefs based on evidence) provide a normative framework: governance that meets these conditions requires that all participants have access to the same material facts about the system being governed. The governance quality was bounded by what the governance infrastructure made legible to its participants. Governance also requires aligned incentives, participant capability, and enforcement mechanisms, but all of these operate downstream of legibility: participants with perfect incentives and full capability still fail if they cannot observe the system state.
Integration requires that the integrating system can determine what the target system does, under what constraints, with what guarantees, and at what cost. When the target system's capabilities are documented in prose rather than structured metadata, integration depends on the integrator's ability to extract structured requirements from unstructured descriptions. This was observable across infrastructure ecosystems (systems integrating with other systems based on white paper descriptions rather than structured interface specifications) and in the agent-platform ecosystem (agents selecting tools based on natural language descriptions without constraint metadata). Latour's (1987) sociology of translation describes the process by which actors enroll other actors into networks: successful enrollment requires that the translated entity (the target system's capabilities) be rendered in terms the enrolling actor can act upon. When the translation is lossy, because the target system's self-description lacks the structural precision the integrator needs, the resulting network is fragile. As the integrating reasoner shifts from human developer to autonomous agent, the extraction step becomes the primary failure point, because the agent lacks the contextual understanding that allows human developers to compensate for unstructured documentation. Lessig's (1999) thesis that code is law applies with particular force here: the integration layer's architecture determines which systems can participate and on what terms, making the legibility of integration interfaces a governance question as much as a technical one.
A single source of structured, verified material facts can serve regulatory requirements, agent integration, developer experience, and governance participation simultaneously. The economic advantage comes from reducing maintenance cost (one source rather than many), reducing inconsistency (one truth rather than divergent versions), and enabling new audiences to self-serve. This compounding effect was the design insight behind both PDAS (which structures disclosure so that a single set of facts serves regulators, integrators, and counterparties) and the knowledge architecture design primitives (which structure knowledge so that consumer-specific renderings derive from a shared foundation). Star and Griesemer's (1989) boundary object theory provides the conceptual architecture: the shared fact base functions as a boundary object that maintains structural coherence while enabling different communities to extract the renderings they need. Callon's (1998) performativity framework adds a dynamic dimension: the shared fact base does more than represent the system; it actively constitutes the informational environment within which regulatory, evaluative, and integrative decisions are made, meaning that improvements to the fact base improve the quality of all downstream decisions simultaneously. The benefits are real but require investment in consumer-specific rendering: the same facts, structured by the same principles, presented differently for different reasoners.
Taxonomy
Five remediable failure modes
The following taxonomy emerged from the research program and recurred across infrastructure analysis, PDAS development, and knowledge architecture research. Each failure pattern was observed independently in multiple systems, confirmed across multiple evaluation dimensions, and found to produce a distinct information deficit in external reasoners. The taxonomy focuses on remediable failures: cases where legibility degrades because of design choices rather than inherent domain complexity. Several of these patterns have parallels in established research traditions. Burke's (1969) dramatistic analysis of motives, Bowker and Star's (1999) work on the politics of classification, Williamson's (1985) analysis of opportunism under information asymmetry, and Popper's (1963) falsificationist epistemology each illuminate different failure modes from different disciplinary vantage points. The convergence of these traditions on the same structural patterns strengthens the claim that the taxonomy identifies general properties of legibility failure rather than domain-specific artifacts.
Present across all infrastructure analyses. Primary documentation and ecosystem reports consistently described what the system aimed to achieve rather than what it structurally was. Evaluation frameworks that consumed this narrative input produced narrative output, because the input lacked the structured evidence needed to enforce analytical rigor. Perelman and Olbrechts-Tyteca's (1969) new rhetoric identifies this pattern as audience adaptation taken to an extreme: the discourse is optimized for persuasion rather than for enabling independent judgment by the audience.
Where it appears
White papers, pitch decks, ecosystem reports, token launch documentation
Consequence
External reasoners form beliefs about what the system claims to be rather than what it structurally is. The belief deficit propagates: evaluations built on narrative input produce narrative conclusions, which inform governance decisions based on narrative evidence. In Toulmin's (1958) terms, the warrant connecting evidence to claim is absent, but the rhetorical structure disguises the gap by substituting narrative coherence for evidential support.
Documented across the majority of infrastructure analyses. The most informative technical details about system mechanisms were consistently available through private channels, informal conversations, or institutional relationships with core teams rather than through published documentation. Analysts with institutional access produced better analyses than analysts relying solely on public materials.
Where it appears
Early-stage protocols, venture-backed platforms, enterprise systems with restricted documentation
Consequence
Legibility becomes a function of social access rather than public infrastructure. Reasoners with institutional relationships form better beliefs than reasoners without them, creating information asymmetry that market mechanisms alone cannot correct. Zelizer's (2005) work on the social meaning of money demonstrates an analogous dynamic: access to financial circuits depends on social relationships, and the relationship structure determines who can participate on what terms. Context privatization creates an equivalent circuit for knowledge, where the terms of access are social rather than technical.
Present across the majority of infrastructure analyses. Systems published ecosystem dashboards showing aggregate activity metrics while omitting the component-level data (operational distribution, governance participation rates, economic model parameters) that external analysts needed to form structural conclusions. Winner's (1980) thesis that artifacts have politics applies directly: the dashboard's aggregation level embodies a decision about who can reason about the system and at what depth.
Where it appears
Dashboards, high-level architecture documents, summary reports
Consequence
The information is technically public but practically opaque. The reasoner can see that the system exists and observe its aggregate behavior but cannot form beliefs about the mechanisms producing that behavior. This creates what Floridi (2011) terms an 'informational environment' that is structurally impoverished: the environment contains data but lacks the semantic structure that would allow an epistemic agent to convert that data into actionable knowledge.
Present across all infrastructure analyses. Every system's documentation described a state that was between 6 months and 3 years behind the current implementation. The most common pattern was documentation that described the system's architecture at its last major version while the implementation had undergone significant changes through governance decisions that were themselves documented only in snapshot records. No system maintained a mechanism to signal the divergence between documentation and system state.
Where it appears
Documentation that is unversioned, changelogs that summarize without specifying, specifications that lack effective dates
Consequence
Reasoners form beliefs about the current system based on documentation that describes a previous state. The gap between documentation and reality widens silently, and no mechanism signals the divergence. March and Olsen's (1989) logic of appropriateness suggests that organizations maintain outdated documentation because institutional norms reward the existence of documentation over its accuracy: having a white paper satisfies the institutional expectation, regardless of whether the white paper describes the current system.
Present across nearly all infrastructure analyses. Systems published detailed claims about reliability, security, and scalability while providing insufficient evidence for external verification. The most common pattern was primary documentation that made specific technical claims alongside marketing materials that amplified those claims, with no structured mechanism for an external party to independently check any of them. In Toulmin's (1958) argument model, these claims provided data and made assertions but omitted the warrant (the justificatory principle connecting evidence to conclusion) and the backing (the grounds supporting the warrant's authority).
Where it appears
Marketing materials presented as technical documentation, claims without corresponding evidence surfaces
Consequence
Legibility is mimicked without substance. The system appears legible because it publishes detailed descriptions, but the descriptions are unfalsifiable assertions rather than verifiable disclosures. The external reasoner cannot distinguish a legible system from one that performs legibility without substance. Lamport's (2002) work on formal specification through TLA+ demonstrates what structured verifiability looks like in practice: a specification language where every system property is expressed as a temporal logic formula that can be mechanically checked against an execution trace. The gap between this standard and current disclosure practice measures the verification deficit.
Discipline
Legibility as engineering discipline
The ceiling argument and failure taxonomy together imply a design discipline: if legibility bounds all external operations, and if the five failure patterns are remediable through design choices, then designing for legibility is an engineering concern with concrete components. Feenberg's (2002) critical theory of technology provides the philosophical grounding: technology embodies values through design decisions, and those decisions can be made deliberately rather than defaulted to incumbent incentive structures. Jasanoff's (2004) co-production framework adds that the relationship between knowledge and social order is mutually constitutive, meaning that legibility infrastructure shapes the social and institutional arrangements that form around it.
Designing systems that are structurally legible to external reasoners by default. In practice, this means publishing machine-readable manifests alongside human-readable documentation (as PDAS requires), encoding material facts at the point of origin rather than reconstructing them later, and tracking legibility degradation with the same urgency as security debt. Meyer's (1992) design by contract methodology provides an operational model: every system interface specifies preconditions (what the consumer must provide), postconditions (what the system guarantees), and invariants (what holds regardless of state). Lamport's (2002) TLA+ demonstrates that temporal properties of systems can be specified formally and checked mechanically, offering a standard for what structured verifiability can look like in practice. Concrete starting points: a release-bound disclosure surface, a canonical artifact registry, and a machine-readable privilege register.
The design primitives (structural decomposability, semantic self-description, verification surface, temporal binding, modality independence, progressive disclosure, composability, adversarial resilience) hold regardless of consumer. From these invariant foundations, reasoner-specific renderings are derived: structured evidence for human analysts, machine-readable capability metadata for agents, release-bound disclosure for regulators, queryable interfaces for AI verification systems. Star and Griesemer's (1989) boundary object framework provides the conceptual precedent: objects that inhabit multiple communities of practice simultaneously, maintaining enough shared structure to enable coordination while permitting local adaptation. The design primitives formalize this insight for knowledge infrastructure: the shared structure is the consumer-agnostic fact base; the local adaptation is the reasoner-specific rendering. This architecture was derived from the observation that rebuilding knowledge for each new consumer class is economically unsustainable and produces inconsistency.
A system's legibility can be assessed along concrete dimensions: disclosure coverage (what percentage of material facts are published and structured), temporal coherence (what percentage of documentation matches current system state), and evidence completeness (what percentage of claims are independently verifiable). The knowledge architecture framework proposes five specific metrics. Lakatos's (1970) methodology of scientific research programmes provides the evaluative standard: a research programme is progressive if it predicts novel facts and some of those predictions are corroborated. The legibility metrics serve as the programme's predictive instrument: they should predict downstream outcomes (faster integration, more reliable evaluation, better governance), and the correlation between metric values and observed outcomes provides the empirical test of the programme's progressiveness. These metrics are operational starting points for teams that want to track legibility as a quality dimension. Their correlation with downstream outcomes requires empirical validation, which is part of the ongoing research program.
In adversarial environments, systems have incentives to appear legible while remaining opaque. The five failure patterns documented in the taxonomy can be deployed deliberately: narrative substitution to deflect scrutiny, context privatization to control information access, granularity collapse to provide the appearance of transparency without its substance. Zuboff's (2019) analysis of surveillance capitalism documents a systemic version of this pattern: platforms that extract behavioral data while rendering their own extraction mechanisms opaque, creating a structural asymmetry between the platform's legibility of its users and the users' legibility of the platform. Noble's (2018) work on algorithmic discrimination demonstrates that opacity can perpetuate structural harm, as classification systems that resist external scrutiny embed biases that compound over time. A legibility discipline must account for adversarial conditions. The PDAS standard addresses this by making disclosure categories explicit, so that a missing category is a visible signal rather than a silent gap. The knowledge architecture framework addresses it through the adversarial resilience principle. Braithwaite's (2002) responsive regulation pyramid offers a governance model: enforcement intensity escalates in proportion to observed non-compliance, but the model's effectiveness depends on the regulator's ability to observe compliance state, which depends on legibility.
Limits & tensions
What the thesis does and does not claim
A legibility thesis that does not confront its own limits risks the high-modernist overconfidence Scott (1998) documented. This section identifies five structural tensions the research program must navigate rather than resolve prematurely.
Scott's central insight is that legibility has historically been an instrument of power. Cadastral maps made land taxable. Standardized surnames made populations conscriptable. Grid cities made neighborhoods policeable. In each case, the state's legibility project destroyed local knowledge (mētis) that sustained the communities being made legible. Emerging infrastructure ecosystems exhibit a version of this tension: regulatory demands for structured disclosure serve the state's need to govern a domain it cannot currently see, and systems that resist legibility may be protecting properties (censorship resistance, pseudonymity, permissionless participation) that structured disclosure would compromise. The thesis argues that remediable opacity in systems that claim to be transparent represents a failure, and that the five failure patterns are design defects rather than features. Where opacity serves a legitimate function (privacy, security, competitive necessity), that function should be disclosed explicitly rather than exercised silently. The distinction between deliberate privacy and accidental opacity is the thesis's response to the Scott critique.
Polanyi (1966) demonstrated that skilled practitioners know more than they can articulate: a cyclist cannot specify the physics of balance, a master diagnostician cannot fully externalize their pattern recognition. Nonaka and Takeuchi (1995) built an entire theory of organizational knowledge creation around the conversion between tacit and explicit forms (the SECI model: socialization, externalization, combination, internalization). Some system properties resist formalization. The operational judgment of a protocol's core developers, the social dynamics that determine whether governance proposals succeed, the contextual knowledge that experienced node operators use to diagnose infrastructure failures: these are partly tacit and will resist structured disclosure regardless of how well the disclosure standard is designed. The thesis's scope is the remediable zone: system properties that can be made legible through design choices but are not, due to institutional incentives, design neglect, or strategic opacity. The boundary between the remediable and the structurally tacit is empirical rather than theoretical, and mapping that boundary is an open research question.
Legibility is a public good: once produced, its benefits are non-rivalrous (one reasoner's use of structured disclosure does not diminish another's) and partially non-excludable (published disclosure is available to any consumer). Public goods are systematically under-provided by markets because producers cannot capture the full social value of their investment. Grossman and Stiglitz (1980) identified a sharpened version of this problem for information goods: if information were freely available, no one would bear the cost of producing it, because the returns on costly analysis would be competed away by free riders. This paradox applies directly to legibility: the infrastructure analysis corpus demonstrates that producing structured legibility for a system requires substantial analytical investment, and if that analysis were freely available, the incentive to produce it diminishes. Mandatory disclosure standards (PDAS, EU AI Act requirements) represent one institutional response: shifting the cost of legibility production to the system operator, where it becomes a cost of doing business rather than a public good requiring external subsidy. Voluntary disclosure regimes (open-source documentation, developer experience investment) represent another: the system operator bears the cost because the competitive benefits (faster integration, better developer experience, regulatory goodwill) exceed it. Both regimes have failure modes, and the research program's treatment of legibility economics remains preliminary.
Foucault's (1980) analysis of the power-knowledge nexus raises an unavoidable question: legibility serves the interests of those who can act on the information it provides. A protocol made legible to regulators enables regulatory action. The same protocol made legible to competitors enables competitive response. Made legible to users, it enables informed participation. Made legible to adversaries, it enables targeted attack. The thesis argues for consumer-agnostic legibility: structured facts from which any reasoner can extract what it needs. Habermas's (1984) discourse ethics provides one normative framework: the conditions for rational discourse (inclusion, equal voice, absence of coercion, willingness to revise beliefs based on evidence) require that all participants have access to the same material facts. Anti-gatekeeping, in this framing, is a discourse ethics requirement rather than merely a documentation best practice. The thesis acknowledges that this position is normative rather than empirically derived: the claim that legibility should be consumer-agnostic is a value commitment that can be challenged by those who argue that some information should be structurally restricted. The crypto-anarchist tradition and the cypherpunk movement represent one such challenge; securities regulation represents another. The tension between these positions is productive rather than resolvable.
Simon (1962) demonstrated that complex systems exhibit emergent properties that cannot be predicted from component-level specifications. Holland's (1995) complex adaptive systems framework extends the point: systems with heterogeneous agents, feedback loops, and adaptive behavior produce macro-level patterns that resist micro-level description. Some complex infrastructure properties (emergent behavior under stress, governance dynamics across competing stakeholder groups, economic equilibria in composable systems) may be structurally resistant to the kind of specification the legibility framework requires. The thesis addresses system properties that are specifiable but unspecified. Whether the five failure patterns account for the full space of remediable legibility failures, and where the boundary between specifiable and emergent lies, remain open questions. Lamport's (2002) formal specification work demonstrates that temporal system properties can be specified with remarkable precision, suggesting the boundary may be further toward the specifiable end than current practice assumes. The research program's task is to push the boundary of what can be made legible while acknowledging that the boundary exists.
Convergence
A research program in system legibility
The three domains described above were developed as independent responses to independent problems. The unifying claim is that all three are applications of a single discipline: making systems structurally legible to external reasoners. The analytical method (Chain of Evidence) produces legibility for a specific system. The disclosure standard (PDAS) produces legibility infrastructure for a class of systems. The design primitives produce legibility architecture for any system, regardless of consumer. Each operates at a different scale; all address the same property. In Lakatos's (1970) terms, the legibility concept functions as the hard core of a research programme: the non-negotiable theoretical commitment around which a protective belt of auxiliary hypotheses (the failure taxonomy, the ceiling argument, the design primitives) can be tested, revised, and extended without abandoning the core insight.
The convergence is empirical rather than imposed. PDAS and the knowledge architecture framework were developed for different audiences through different design processes. They arrived at the same failure taxonomy, the same structural requirements (release-bound conformance maps to temporal binding, machine-readable manifests map to verification surfaces, explicit disclosure categories map to adversarial resilience), and the same underlying principle (structured, verifiable evidence as the foundation for external reasoning). Kuhn's (1962) account of paradigm formation through the accumulation of anomalies that a new framework resolves provides one model for this convergence: the same structural anomalies (the five failure patterns) appeared across domains that lacked a shared explanatory framework, and the legibility concept resolves them by identifying the common cause. Longino's (1990) social epistemology adds a normative dimension: the convergence gains credibility to the extent that it emerges from independent inquiry subject to mutual criticism, and the independent development of PDAS and the knowledge architecture framework satisfies this condition. The convergence strengthens the case that legibility is a property of knowledge systems themselves rather than an artifact of a particular domain or consumer class.
Trajectories
Forward trajectories
The following trajectories are conditional on three observable trends continuing: autonomous systems becoming widespread consumers of platform metadata (current MCP adoption, enterprise agent deployment), market reward for structured knowledge investment (emerging in developer experience metrics), and regulatory mandates for structured disclosure (EU AI Act effective August 2026, MiCA white paper requirements). North's (1990) institutional economics provides the analytical framework: institutions (formal rules and informal norms) shape the incentive structures that determine organizational behavior, and regulatory mandates, market rewards, and technological capabilities each alter the institutional landscape in ways that make legibility investment more or less attractive. If these trends stall, the trajectories slow accordingly.
Just as every networked system today maintains an API surface, public systems will increasingly maintain a legibility surface: a structured, versioned, consumer-agnostic account of what the system is, what it does, how it is governed, and how its claims can be verified. Lessig's (1999) four modalities of regulation (law, norms, markets, architecture) suggest that legibility surfaces will emerge through multiple channels simultaneously: regulatory mandates (EU AI Act), market pressure (developer experience metrics), normative expectations (open-source disclosure culture), and architectural necessity (agent integration requirements). If autonomous systems become widespread consumers of platform metadata, and if regulation mandates structured disclosure, these surfaces will serve human analysts, agents, AI verification systems, regulatory bodies, and future reasoner forms from a single structured source. Early indicators (EU AI Act documentation requirements, MCP adoption trajectories, enterprise knowledge graph investment) suggest these conditions are emerging, though the pace and form of adoption remain uncertain.
When AI systems evaluate, compare, and select platforms on behalf of human decision-makers, the quality of a platform's legibility surface influences its selection probability. Two platforms with identical capabilities are distinguished by which one any reasoner can understand more reliably. The compounding effect operates across consumer classes: platforms that agents select more often accumulate integration data that improves their legibility, which also improves human developer experience, regulatory assessability, and verification coverage. Early investment in consumer-agnostic legibility creates a structural advantage that widens across multiple dimensions simultaneously.
PDAS-style disclosure, DNCE-style agent-facing knowledge architecture, and human-facing technical writing are converging toward a single source artifact type: structured, versioned, verifiable knowledge from which consumer-specific renderings are derived. The design primitives proposed in the knowledge architecture work make this convergence architecturally possible. The distinction between 'disclosure for regulators,' 'metadata for agents,' and 'documentation for developers' dissolves as all three require the same underlying material facts, structured by the same principles, rendered for different consumption contexts.
As legibility surfaces become critical infrastructure, questions of legibility governance become unavoidable. Who controls what a system discloses about itself? What happens when AI-generated documentation diverges from system behavior? How is provenance tracked when AI co-authors knowledge artifacts? What are the consequences of deliberately degrading legibility to avoid scrutiny? Jasanoff's (2004) co-production framework predicts that the answers to these questions will shape, and be shaped by, the social and institutional arrangements that form around legibility infrastructure. The governance of legibility is co-produced with the legibility itself: the disclosure categories a standard mandates determine what becomes visible, and what becomes visible determines what governance mechanisms are possible. Ostrom's (1990) design principles for commons institutions (clear boundaries, proportional equivalence between benefits and costs, collective-choice arrangements, monitoring, graduated sanctions) provide a structural template for legibility governance. These questions will require formal governance mechanisms with the same rigor as financial disclosure governance. The EU AI Act's documentation requirements (effective August 2026) represent the opening of this regulatory trajectory.
Contributions
What this thesis contributes
To transparency research
Legibility as a unifying property across transparency, discoverability, verifiability, and understandability, situated within the broader traditions of accountability (Hood 2006; Bovens 2007), information asymmetry (Akerlof 1970; Stiglitz 2000), and co-production of knowledge and social order (Jasanoff 2004). The five-pattern failure taxonomy enables cross-domain failure propagation analysis: tracing how a legibility failure at a source system compounds through every downstream operation that depends on it, a mechanism that parallels MacKenzie's (2006) performativity analysis of how models shape the markets they describe.
To evaluation methodology
The ceiling argument reframes where organizations should invest. When evaluations, trust assessments, or governance outcomes are unsatisfying, the productive investment is in the system's disclosure quality, where returns are unbounded, rather than in evaluator sophistication, where returns diminish past the legibility ceiling. This finding converges with Gigerenzer's (2000) ecological rationality: decision strategies are effective only when matched to their informational environment, and improving the environment's information structure improves all strategies simultaneously.
To system design
Four components of a legibility engineering practice: structural legibility by default (informed by Meyer's (1992) design by contract), consumer-agnostic foundations with reasoner-specific rendering (extending Star and Griesemer's (1989) boundary objects), operational metrics for legibility quality (evaluated through Lakatos's (1970) progressiveness criteria), and adversarial resilience through explicit disclosure categories (motivated by Zuboff's (2019) and Noble's (2018) analyses of structural opacity).
To institutional theory
The failure taxonomy and ceiling argument contribute to institutional analysis by documenting how information asymmetry is produced and maintained through specific, remediable design patterns. The five failure modes extend Williamson's (1985) transaction cost analysis by identifying the disclosure-level mechanisms through which opacity is created, and complement Ostrom's (1990) commons governance principles by specifying the information infrastructure that monitoring and collective choice require.
To political economy
The limits section confronts the power question Scott (1998) raised and Foucault (1980) formalized: legibility serves the interests of those who can act on the information it provides, and the choice of what to make legible is a political decision. The thesis responds with a discourse ethics position (Habermas 1984): consumer-agnostic legibility creates the informational preconditions for rational governance, and the anti-gatekeeping principle ensures those preconditions are met. The economics of legibility production (Grossman & Stiglitz 1980), the principal-agent dynamics of disclosure (Jensen & Meckling 1976), and the mechanism design question of what institutional structures produce honest revelation (Hurwicz 1960; Myerson 1981) remain underdeveloped areas that the research program must address.
To the research program
Empirical convergence across three independent bodies of work (structural analysis, protocol disclosure, knowledge architecture) strengthens the case that legibility is a property of knowledge systems themselves, accessible to investigation from any domain. In Lakatos's (1970) terms, the programme is progressive: it generates novel predictions (the ceiling effect, the compounding structure, the taxonomy's cross-domain recurrence) that are corroborated across independent domains. The open questions include: whether the five failure patterns are exhaustive or merely the most common; whether the ceiling effect admits exceptions under conditions the current corpus has yet to encounter; where the boundary between remediable opacity and structurally tacit knowledge lies (Polanyi 1966; Nonaka & Takeuchi 1995); how legibility production should be financed as a public good (Grossman & Stiglitz 1980); and whether the emergent properties of complex adaptive systems (Simon 1962; Holland 1995) impose fundamental limits on what can be made formally specifiable.
This thesis paper is a living document. The legibility framework, failure taxonomy, and discipline definition will be refined as the underlying work in structural analysis, protocol disclosure, and knowledge architecture develops.