arXiv CSopen access

Actionable Understanding: Action Units for Bridging the Knowledge-Action Gap in Post-FAIR Knowledge Infrastructures

Actionable Understanding: Action Units for Bridging the Knowledge-Action Gap in Post-FAIR Knowledge Infrastructures Vogt, Lars1 1

orcid.org/0000-0002-8280-0487

Leibniz Institute for the Analysis of Biodiversity Change (LIB), Museum of Nature Hamburg, Martin-8 Luther-King Platz 3, 20146 Hamburg, German

1

Abstract Despite unprecedented growth in biodiversity data and formally structured knowledge, a persistent gap remains between what is known and what is effectively acted upon. Existing frameworks such as the FAIR Principles and the CLEAR Principle have substantially improved the findability, accessibility, interoperability, and human interpretability of data and knowledge. However, these approaches remain primarily focused on representation and do not provide the components required to translate knowledge into context-sensitive action. In this paper, we argue that closing the knowledge–action gap requires a shift from concept- and data-centric representations toward statement-centred and action-oriented knowledge infrastructures. We identify a fundamental distinction between actionability as the structural capacity of a knowledge representation to support operations and applicability as the epistemic validity of using that knowledge in a specific context. This distinction, which existing frameworks conflate or ignore, is essential for ensuring that action is not only executable but contextually justified. Building on the Semantic Units Framework, we introduce Action Units as a new class of formal knowledge objects that extend plan specifications and thus information content entities specifying how processes are to be carried out, by integrating semantic content with contextual information, explicit objectives, procedural knowledge, and applicability conditions as first-class typed components. We distinguish three types of action units corresponding to three classes of operations: epistemic action units, which ground representations in real-world situations through recognition, designation, and description; transformational action units, which process and derive representations required for interpretation and decision-making; and intervention action units, which integrate causal and procedural knowledge to guide goal-directed modification of systems. Together, these define a minimal operational architecture for actionable knowledge. Action units can also be granularly composed, with higher-order action units consisting of ordered sequences of lower-order units across different operation classes, reflecting the constitutively cross-class character of real-world knowledge-driven processes. We further show that conditional action units, operationalized as executable IF–THEN structures within knowledge graphs, enable knowledge graphs to function not merely as repositories but as graph-native decision-support and workflow orchestration systems. Applicability conditions are formalized as executable queries, with directive actions becoming executable procedures. This constitutes a transition toward post-FAIR knowledge infrastructures in which knowledge is not only findable and interoperable, but operationally integrated and context-sensitive. Using biodiversity research and ecosystem management as a primary use case, we illustrate how action units support context-sensitive decision-making through worked examples across all three operation classes, reinterpret recurring patterns of intervention and epistemic failure in biodiversity science as consequences of incomplete actionable knowledge structures that support both human and machine agents. We conclude that actionable understanding emerges not from representation alone, but from the coordinated alignment of knowledge structures, operations, context, and agents—and that, in many real-world settings, such understanding is refined through action itself. On this basis, we propose the TripleA Principle—Actionability, Applicability, and Auditability—as a guiding framework for next-generation knowledge infrastructure design that extends and complements the FAIR and CLEAR Principles.

2

1. Introduction: The Knowledge—Action Gap in Biodiversity Science In the face of accelerating biodiversity loss, ecosystem degradation, and increasingly urgent demands for evidence-based decision-making, the scientific community has never had access to more data, metadata, and formally structured knowledge. Global biodiversity monitoring initiatives, museum collection databases, ecological observatories, remote sensing platforms, and citizen science programs continuously generate vast and rapidly expanding information resources. Yet despite this unprecedented abundance, translating knowledge into effective action remains a persistent challenge. The problem is no longer primarily the absence of data, but rather the lack of knowledge representations that enable scientists, practitioners, policymakers, citizens, and computational agents to interpret knowledge in context and derive knowledge that can support context-sensitive decisions and actions. In other words, we are increasingly data-rich but still action-poor, resulting in a persistent knowledge-action gap (Fig. 1). This challenge is not unique to biodiversity research but reflects a broader problem across data-intensive sciences.

Figure 1: From semantic content to action. The missing layer in FAIR and CLEAR knowledge infrastructures.

Over the past decade, major advances have been made in the representation and accessibility of knowledge. The FAIR Principles (1) have advanced the machine Findability, Accessibility, Interoperability, and Reusability of data and metadata, while recent additions such as the CLEAR Principle (2) emphasize the importance of Cognitively interoperable, semantically Linked, contextually Explorable, intuitively Accessible, and human-Readable and -interpretable data and knowledge. In parallel, ontologies and semantic knowledge graphs have enabled formal, concept-centric representation of entities and their relationships (3,4). Together, these approaches support increasingly sophisticated forms of machine-interoperability and human-interpretability. However, these frameworks remain primarily focused on representing and organizing data and knowledge, rather than enabling their direct use for context-sensitive action. They excel at answering questions such as “What is X?” or “How is X related to Y?”, but provide limited support for determining how terms and knowledge can be applied in practice and what should be done in a specific situation to achieve a specific goal. As a result, a critical gap remains between knowledge representation and knowledge-informed action.

3 This limitation reflects a deeper conceptual issue. Existing approaches do not clearly distinguish between the structural capacity of knowledge representations to support actions (what can be done with a given knowledge representation) and the contextual validity of applying such knowledge in a given situation (whether it should be done in a specific situation). In this work, we address this gap by introducing a distinction between actionability and applicability. Informally, actionability refers to the operations that a given knowledge representation enables, whereas applicability concerns whether this knowledge can be validly used in a specific context. While closely related, these dimensions capture fundamentally different aspects of how knowledge supports action. This distinction is particularly critical in ecology and other highly complex systems, where context-dependence, scale effects, and system variability frequently limit the transferability of general knowledge. In such settings, the availability of data and formalized knowledge alone is insufficient as effective decision-making requires the ability to evaluate whether knowledge is appropriate for the situation at hand. Building on this perspective, we introduce the concept of actionable understanding, defined as the capacity to interpret knowledge, assess its relevance in a given context, and use it to guide appropriate action. We argue that actionable understanding is inherently statement-centred, as scientific knowledge is primarily conveyed through statements, i.e., propositions describing relationships, observations, or processes, rather than isolated terms. Enabling action therefore requires representations that support not only the interpretation of statements, but also their contextualization and operational use. Actionable understanding emerges not only from the availability of knowledge representations, but from the processes that enable their use. These include epistemic operations that relate concepts and knowledge representations to real-world situations through recognition, designation, and description, transformational operations that process and derive relevant representations, and intervention operations that integrate causal and procedural knowledge to effect goal-directed changes in real-world systems. These operation types are interdependent. Together, they define a progression from situational understanding, to knowledge preparation, to system-level action, forming the operational backbone of actionable understanding. To address the knowledge-action gap, we propose a framework for actionable knowledge representation built on three core components. First, we adopt a statement-centred perspective grounded in the Semantic Units Framework (5–8), treating statements rather than isolated terms or triples as the primary units of meaning. Second, we introduce Action Units as a structured extension of plan specifications within the Semantic Units Framework, integrating semantic content with goals, procedures, and explicit applicability conditions as first-class typed semantic unit components, and we distinguish three corresponding types—epistemic, transformational, and intervention action units—that together define a minimal operational architecture for actionable knowledge. Third, we formalize the distinction between structural actionability—the capacity of a representation to support operations—and contextual applicability—the epistemic validity of using that knowledge in a specific situation—as a principled evaluation layer linking general knowledge to concrete contexts. A further implication of the framework is that knowledge graphs can be extended through conditional action units to function as graph-native decision-support and workflow orchestration systems rather than passive repositories. In this sense, A3ctionable Understanding rests on three interdependent pillars—Actionability, Applicability, and Action Units—whose systematic integration constitutes the central contribution of

4 this work and forms the basis for the TripleA Principle proposed in the conclusion as a guiding framework for next-generation knowledge infrastructure design. We want to explicitly note that this paper is a conceptual framework contribution. Implementation and empirical evaluation are deferred to subsequent work. The remainder of the paper develops this argument in stages. Chapter 2 analyses the representational and operational limitations of existing concept- and data-centric approaches. Chapter 3 establishes statements as the foundational units of knowledge representation within the Semantic Units Framework. Chapter 4 formalizes the role of context and applicability in linking general knowledge to specific situations. Chapter 5 introduces operations and action units as the operational and representational backbone of actionable understanding. Chapter 6 synthesizes the resulting architecture and its dual forward and reverse perspectives on knowledge use. Chapter 7 applies the framework to biodiversity science, reinterpreting recurring failure patterns in conservation and monitoring as consequences of underspecified actionable knowledge structures, constructing worked action unit examples across all three operation classes, and illustrating how individual action units compose into conditional and ecological fingerprint structures. Chapter 8 develops the implications for the design of next-generation knowledge infrastructures, including conditional action units and the transition toward post-FAIR systems. Chapter 9 concludes.

2. Limits of Current Knowledge Infrastructures 2.1 Concept- and Data-Centric Paradigms Current knowledge infrastructures are primarily shaped by four complementary paradigms: ontologies, RDF/OWL-based knowledge graphs, the FAIR Guiding Principles, and user-oriented frameworks such as CLEAR. While each substantially improves the representation, accessibility, and interoperability of knowledge, they remain insufficient for supporting actionable understanding. Ontologies provide formal, concept-centric representations by defining classes, relations, and logical constraints. Upper-level ontologies such as the Basic Formal Ontology (BFO) (9) and DOLCE (10) provide foundational categories for entities, processes, agents, and their relationships that have informed ontology development across many scientific domains. The representational choices made throughout this work draw on BFO-aligned upper-level categories, particularly for the characterisation of processes, information content entities, material entities, and their roles. Ontologies are highly effective in addressing the question “What is X?” through ontological and intensional definitions and by specifying semantic relationships between entities. However, they generally lack the diagnostic and procedural knowledge required for recognizing instances in practice or for guiding action in real-world situations (11,12). As a result, they support conceptual clarity but not operational use. RDF/OWL-based knowledge graphs provide the primary technical infrastructure for implementing ontologies and linking data and knowledge across distributed systems (3,13). They are based on a triple-centred representation model that encodes information as binary Subject–Predicate–Object relations (14). This model offers a flexible and formally well-founded basis for representing relational knowledge and supports logical reasoning over structured data. However,

5 it is inherently optimized for binary relations. Representing more complex, n-ary statements, such as context-dependent claims or procedural specifications, requires decomposition into multiple interconnected triples, often through reification or the introduction of intermediate nodes (15). This leads to a fragmentation of semantically coherent statements across graph structures, making it difficult to treat statements as unified entities, to associate them with provenance and applicability conditions, and to support their direct use in context-sensitive reasoning and action. As a result, while knowledge graphs excel at linking data and representing binary relations, they provide only limited support for representing statements as first-class, actionable units of knowledge. The FAIR Principles have transformed data management by improving the findability, accessibility, interoperability, and reusability of data and metadata (1,16,17). However, as a fundamentally data-centric framework, FAIR focuses on discovery and integration rather than on how knowledge can be translated into context-sensitive decisions or actions. The CLEAR Principle complements FAIR by emphasizing cognitive interpretability and contextual exploration (2). It highlights the importance of semantically linked, human-interpretable knowledge and implicitly points toward the role of statements as meaning-carrying units. However, CLEAR does not provide explicit mechanisms for linking knowledge to procedures, goals, or interventions. Taken together, these paradigms significantly improve access, structure, and interpretability of knowledge, but they do not provide the representational components required for guiding action, such as procedural knowledge, explicit goals, or referential, formal, or context-dependent applicability conditions. As a consequence, they support understanding what is the case, but not what should be done.

2.2 Actionability of Terms and Statements A central limitation of existing approaches lies in the absence of a clear distinction between the actionability of terms and statements, and between human- and machine-actionability (see Box 1). This section distinguishes between different forms of actionability across representation levels (terms vs statements) and agent types (human vs machine), and introduces three classes of statements that play distinct roles in actionable knowledge. For machine agents, actionability is defined by the availability of executable operations. Machine-actionable knowledge is therefore tied to workflows, APIs, scripts, and reasoning procedures that can process formally structured representations (18). This perspective aligns with emerging visions of FAIR ecosystems (7,12) that include terminology, schema, operations, and workflow services, enabling automated reasoning and (agentic AI driven) execution. For human agents, actionability requires more than formal structure. Human-actionability depends on interpretation, contextualization, and procedural guidance. Interpretation is necessary but not sufficient for actionability. Knowledge must also provide guidance on how to recognize relevant situations and how to act within them. As such, human-actionability is strongly tied to epistemic and intervention operations. Ontologies and terminologies primarily address the meaning of individual terms and thus play an important role in epistemic operations. Regarding the human-actionability of terms, it is useful to distinguish between inferential and referential lexical competence (19) (Box 1). Inferential lexical competence refers to understanding of the intensional meaning of a term—how it is defined, interpreted, and related to other concepts. This is supported by ontological and intensional definitions, which answer questions such as “What is it?” and “How is it understood?” (12).

6 Referential lexical competence, in contrast, enables actionable use by supporting recognition (given a term, identify a corresponding instance) and designation (given an instance, identify the appropriate term). This requires diagnostic knowledge, i.e., knowledge of how instances can be identified in practice, often in the form of method-dependent recognition criteria that provide the extensional meaning of a term, such as measurement protocols, identification keys, chemical assays, medical diagnostics, decision rules, or observational procedures. It includes tasks such as identifying a species, classifying a habitat, or diagnosing a condition depend on such diagnostic procedural knowledge (11). While recent work has emphasized the importance of diagnostic knowledge, ontology-based frameworks typically provide only ontological and intensional definitions (11). As a result, ontologies typically only support inferential but not referential lexical competence, thereby limiting the operational utility of terms. Regarding the machine-actionability of terms, when modelled using Description Logics and documented in OWL-based ontologies, formal reasoning can be applied for recognition and designation tasks to automatically classify a given instance based on necessary and sufficient properties defined within class axioms. Crucially, even when terms are made actionable in this sense, they do not enable actionable understanding. Meaning in scientific discourse is primarily conveyed through statements, which express propositions about relationships, observations, hypotheses, or processes. As will be established in Chapter 3, statements rather than isolated terms are the primary units of meaning in scientific discourse. Statements can be categorized into three fundamentally different types with respect to their role in actionable knowledge: (i) instance-level knowledge in the form of assertional statements and general knowledge in the form of (ii) universal and prototypical statements and (iii) directive and conditional directive statements. Assertional statements describe the relationships between individual entities, refer to specific situations or observations, and therefore function as contextual information and empirical evidence. Their human-actionability relies on epistemic operations such as recognizing whether a statement applies to a given situation or describing a situation using an appropriate statement. Their machine-actionability, in contrast, depends on the interoperability and the availability of computational functions that can process the corresponding statement types, which typically requires standardized representations such as SHACL-constrained data models. Universal and prototypical statements express generalizable knowledge, including causal relationships and statistical correlations, that are not tied to specific situations. They provide the basis for explanation and intervention but are not directly actionable without additional structure. Procedural knowledge representing methods, algorithms, plan specifications, and practices instructions takes the form of directive statements with detailed, step-by-step instructions on what to do to achieve a specific result, frequently including if-then conditional directive statements. At the representational level, such statements specify how actions can be performed, often in the form of conditional and goal-oriented structures. Applying such general knowledge typically requires contextual information describing the situation, contextual applicability conditions specifying when the general knowledge holds or can be applied successfully, and procedural knowledge, including objective specifications that define how to act. Actionable understanding emerges from the interaction between these three forms of propositional content: assertional statements provide the situational grounding, while universal and prototypical statements provide the knowledge structures that inform action, which are specified using (conditional) directive statements.

7 This taxonomy reflects a fundamental distinction in the philosophy of language. Assertional, universal, and prototypical statements share a word-to-world direction of fit in Searle's sense (20,21): they aim to describe how the world is, either at the level of specific situations or at the level of general regularities and dispositions. Directive and conditional directive statements, by contrast, share a world-to-word direction of fit: they specify how the world should be changed to match a desired state. The distinction between these two classes thus captures not merely a representational difference but a fundamental difference in the epistemic function of statements, i.e., between representing reality and prescribing action. Within the present taxonomy, the finer-grained distinctions between assertional, universal, and prototypical statements on one side, and between directive and conditional directive statements on the other, follow Jansen & Schulz (22) as further developed within the Semantic Units Framework (6), but the two-class structure is grounded in this more fundamental philosophical distinction. Applicability assessment links these layers (assertional, universal and prototypical, and directive and conditional directive statements) by evaluating whether general knowledge can be validly applied in a given situation. Different classes of operations thereby impose fundamentally different requirements on applicability, making actionability also operation-dependent. Some operations are intrinsically applicable once representational compatibility in the form of compliance with a specific data schema and format is satisfied, including most transformational operations such as unit conversion that take data as input and produce data as output. In these cases, applicability is deterministic and does not require contextual information. In contrast, epistemic operations of identifying real-world situations to which general concepts (e.g., ecological niche) or statements (e.g., a causal hypothesis) apply or intervention operations that apply methods and practices (e.g., a restoration method) to a given real-world system are contextually applicable. Their valid execution depends on evaluating contextual applicability conditions against contextual information and is often associated with uncertainty. This distinction already shows that support for operations on knowledge does not, by itself, guarantee valid use in a specific situation and highlights a critical limitation of current knowledge infrastructures, which usually only support the execution of operations on data and knowledge but do not support the evaluation of whether these operations are valid in a given context. As a result, they enable action without ensuring its appropriateness. Actionable knowledge thus spans three interdependent classes of operations. Epistemic operations establish the relationship between knowledge representations and thus between information content entities and material systems, transformational operations manipulate and derive representations within the space of information content entities, and intervention operations effect goal-directed changes in material systems. Intervention operations represent the goal-oriented endpoint, but depend on prior epistemic grounding, particularly applicability assessment, and on transformation-enabled preparation and integration of knowledge representations. Box 1 Readability, Interpretability, and Actionability of Terms and Statements and Their Human and Machine Dimensions Human-Centred Actionability Human-readable term (12)​ A term is human-readable if it consists of a sequence of characters that can be mapped to sounds. Communication requires sender and receiver to use the same set of characters and agree on a reading direction. Human-interpretable term (12)​ A term is human-interpretable if it is readable and can be assigned a cognitive representation of its intensional meaning. This requires a shared inferential lexical competence (19), i.e., knowledge of how the term is defined and related to

8

other concepts. Ontological and intensional definitions provide this knowledge. Ontologies provide this knowledge for structured data. Human-actionable term (12) A term is human-actionable if it is interpretable and supports the operations of recognition (given term, identify a corresponding instance) and designation (given instance, identify the appropriate term). This requires a shared referential lexical competence (19), i.e., diagnostic knowledge of how instances can be identified in practice, typically via method-dependent recognition criteria such as measurement protocols or decision procedures. Logical reasoners support designation tasks by automatically classifying instances based on their properties and on necessary and sufficient properties specified in an ontology. Human-readable statement (12)​ A statement is human-readable if it consists of human-readable terms organized according to rules that delimit terms and define statement boundaries. A data structure is human-readable if it can be represented in this form. Human-interpretable statement (12)​ A statement is human-interpretable if (i) it is human-readable, (ii) its constituent terms are interpretable, and (iii) it can be mapped onto a syntactic structure (i.e., a syntax tree) with syntactic positions that are associated with specific semantic roles. This requires sender and receiver to share a terminology and a set of common statement structures. Approaches such as Rosetta Statements (7,8,23) provide formalized templates that support such interpretation and act as semantic anchors in knowledge infrastructures. Human-actionable statement (12)​ A statement is human-actionable if it is interpretable and supports the operations of recognition (given a statement, identify the situation to which it applies) and description (given a situation, identify a statement that accurately describes it). Actionable statements can be combined into coherent narratives, provided that entities can be tracked consistently across statements. For universal and prototypical as well as directive and conditional directive statements, human-actionability further requires the capacity to assess whether the statement’s applicability conditions are satisfied in a given context—that is, to determine whether the general knowledge the statement encodes is validly applicable to the specific situation at hand. Machine-Centred Actionability Machine-readable term (12)​ “[E]lements in bit-sequences that are clearly defined by structural specifications”[(p. 3 in (18)). In the Resource Description Framework (RDF), these are resources in the form of IRIs and typed literals. Machine-interpretable term (12)​ “[E]lements [in bit-sequences] that are machine-readable and can be related with semantic artefacts in a given context and therefore have a defined purpose” (p. 3 in (18), emphasis added). In RDF/OWL, this is achieved through the use of ontology terms with formal semantics. Machine-actionable term (12)​ “[E]lements in bit-sequences that are machine-interpretable and belong to a type [of element] for which operations have been specified in symbolic grammar” (p. 3 in (18), emphasis added). In OWL, such operations are provided by logical reasoning based on Description Logics. Machine-readable statement (12)​ A statement composed of machine-readable terms. In RDF, this corresponds to triples following the Subject–Predicate–Object structure. Since RDF triples represent binary relationships, there is no one-to-one correspondence between human-readable statements and RDF triples. Within the Semantic Units Framework, independent of the underlying n-arity, all statements are represented as statement units. Machine-interpretable statement (12)​ A machine-readable statement encoded in a formal language (e.g., OWL) that supports semantic interpretation. Machine-actionable statement (12)​ A machine-interpretable statement for which operations are available that take the statement as input. These

9

operations may include logical reasoning, data transformation, statistical analysis, or other computational procedures. Interoperability is a key requirement; formal constraints such as SHACL shapes support the development of reusable operations for specific statement types. For conditional directive statements, machine-actionability is further realised through conditional action units—executable IF-THEN structures in which the IF clause is operationalized as a query evaluating whether applicability conditions are satisfied in the current context, and the THEN clause triggers a corresponding computational action or workflow (see Section 8.5). This enables knowledge graphs to function as graph-native decision-support and workflow orchestration systems rather than passive repositories.

2.3 Context, Contextual Information, and Applicability A further limitation of current knowledge infrastructures lies in the absence of explicit mechanisms for linking knowledge to specific situations and evaluating its validity in context. Addressing this limitation requires distinguishing three related but conceptually distinct notions, i.e., context units, contextual information, and applicability conditions. Within the Semantic Units Framework (5) (introduced in Section 3.2), context units provide structural frames of reference that organize a knowledge graph into coherent subgraphs by separating for instance information about the structure of a scientific publication into different sections and paragraphs from information about the research activity it reports (the materials and methods applied), and from information about the real-world system it investigated (the results). This is a representational notion of context that organizes knowledge but does not determine its applicability. Its role within the Semantic Units Framework is discussed in Section 3.2. Contextual information, by contrast, refers to instance-level descriptions of a specific situation, including system states, environmental conditions, and observations, represented as assertional statements. It provides the situational input required for interpreting and applying knowledge. Contextual information may be retrieved from the knowledge graph or provided by a user or system at decision time. Referential, formal, and contextual applicability conditions specify the constraints under which a statement can be validly applied. Referential applicability conditions specify method-dependent recognition criteria, observability conditions, and method validity for epistemic operations. Formal applicability conditions specify the data schema of input data for transformational operations. Contextual applicability conditions define the relationship between general knowledge and specific situations and can be operationalized as executable question units (6) that evaluate whether a given context satisfies the required conditions. Applicability conditions function as the operational bridge between general procedural knowledge and situational decision-making. They enable not only the identification of applicable actions but also the detection of unmet conditions, thereby supporting decision-making under uncertainty and guiding data acquisition or system modification. Applicability should not be understood as a type of knowledge, but as a relation between knowledge and context. It characterizes whether, and to what extent, a given knowledge representation can be validly used in a specific situation. Importantly, applicability is not a Boolean property but graded, ranging from unknown to plausible, supported, or empirically validated, depending on the availability of evidence and contextual information. This reflects the inherent uncertainty of real-world systems, particularly in ecology.

2.4 Core Gap

10 The preceding analysis reveals a fundamental limitation of current knowledge infrastructures. Even when terms are made actionable through diagnostic knowledge, this remains insufficient for actionable understanding, because actionable understanding fundamentally depends on contextualized general statements, applicability assessment, and procedural knowledge. Statements become actionable only when they are linked to specific situations and when their applicability conditions are explicitly defined. General knowledge alone cannot reliably guide action without mechanisms for evaluating whether it applies in a given context. For example, a causal statement about an ecological process is only actionable if its validity can be assessed in a given environmental context and if it can be linked to potential interventions. Without this linkage, general knowledge cannot reliably guide action. Existing infrastructures provide formal ontological and intensional definitions (ontologies), accessible and interoperable data and metadata for transformation operations (FAIR RDF/OWL-based knowledge graphs), and human-interpretable representations (CLEAR), but they do not provide integrated representations that combine universal and prototypical statements documenting causal and statistical knowledge, directive and conditional directive statements documenting procedural knowledge, contextual applicability conditions, and assertional statements documenting contextual information for knowledge-driven action. As a result, current knowledge infrastructures make knowledge interpretable and executable, but not reliably actionable. Bridging this gap requires a shift from concept- and data/triple-centric representations toward statement-centric and action-oriented knowledge structures that explicitly integrate applicability as a core component. In the following chapter, we address this challenge by establishing statements as the primary units of knowledge representation and by introducing a framework that enables their contextualization and operationalization.

3. Statement-Centred Knowledge Representation 3.1 Statements as Meaning-Carrying Units The limitations identified in the previous section point to a fundamental shift in how knowledge should be represented. While traditional approaches emphasize terms and concepts as primary units, scientific knowledge is not conveyed by isolated terms, but by statements, i.e., propositions describing relationships, observations, hypotheses, or processes (12). Statements are the primary carriers of meaning because they express how entities are related, what properties they have, and under which conditions specific claims hold. A term acquires operational meaning only when embedded in a statement that specifies its semantic role within a proposition. Consequently, actionable understanding depends not only on the correct interpretation of individual terms, but on the ability to interpret and apply statements in relation to specific situations. The importance of treating statements rather than isolated terms or triples as primary units of knowledge representation has been recognized across several communities. In the nanopublications literature, Groth, Gibson, and Velterop (24) proposed the nanopublication model precisely to make core scientific statements, rather than raw data or isolated triples, findable, connected, citable, and associated with their provenance. In the Linked Data community, Heath and Bizer's foundational

11 work (25) established best practices for publishing structured data on the web, emphasizing that meaningful, reusable data requires coherent, self-contained statements with clear semantic roles rather than decontextualized triples. These developments reflect a shared recognition that scientific propositions are often inherently n-ary, involving multiple participants, conditions, and roles, and therefore cannot be adequately captured within a single Subject-Predicate-Object triple. Even technical approaches, such as RDF-star (26), that enable metadata to be attached to individual triples do not resolve this more fundamental representational challenge, i.e., a statement that requires multiple triples to be adequately modelled cannot be referenced or reasoned about as a unified representation without an explicit representational commitment to the statement level. This motivates the statement-centred approach adopted in this work. As established in Chapter 2, it is essential to distinguish between different types of statements. These statement types can be organized along a fundamental distinction between instance-level and general knowledge. Assertional statements represent instance-level knowledge by describing specific situations or observations. General knowledge, in contrast, comprises two complementary forms: (i) universal and prototypical statements, which express causal or statistical regularities, and (ii) directive and conditional directive statements, which represent procedural knowledge in the form of methods, algorithms, or instructions of practices. These three types of statements play complementary roles across the knowledge-action spectrum: assertional statements provide descriptions of particular real-world situations, while universal and prototypical statements provide knowledge that may extend across situations, and (conditional) directive statements provide instructions for goal-oriented actions, specifying how to change a given real-world system. A defining characteristic of statements is their inherent context-dependence. Their meaning and relevance depend on factors such as environmental conditions, spatial and temporal scale, methodological assumptions, and background knowledge. For assertional statements, this context is typically inherent in the situation being described. For general knowledge, i.e., universal and prototypical as well as (conditional) directive statements, however, context-dependence is more subtle: although formulated in general terms, their validity is often contingent on specific conditions. This observation has an important consequence. General knowledge cannot be assumed to hold universally in practice; its validity is contingent on context. Its meaningful use requires relating it to specific situations and evaluating the conditions under which they apply. This need motivates a shift from term-centric to statement-centric representation as the foundation for actionable knowledge.

3.2 Semantic Units Framework Building on this perspective, the Semantic Units Framework provides a formal representation model in which statements are treated as first-class, identifiable units of meaning. Semantic units are modular, self-contained, and referenceable units of semantic content that can be interpreted independently within a knowledge space (5–8,23). Within this framework, statement units represent individual propositions, while compound units represent semantically coherent collections of statements (and potentially other compound units), corresponding to structured knowledge objects such as causal hypotheses, workflows, or system descriptions. This modular and potentially nested organization enables knowledge to be decomposed into meaningful units that retain semantic integrity while remaining linkable and reusable across contexts. Assertional, universal, prototypical, directive, and conditional directive

12 statements can be represented within the Semantic Units Framework as different types of statement units, and specific combinations of them as different types of compound units (6). A key structural feature of the framework is the explicit representation of context units, which define frames of reference within a knowledge graph. Context units are compound units that organize semantic units into coherent subgraphs corresponding to different frames of reference, such as the organizational structure of a document, the research activity it reports, and the real-world system the activity investigated. For example, within a knowledge graph representing a scientific publication, a context unit might organize information about the structure of the publication itself as a distinct frame of reference, separating it from a frame containing the materials and methods used in the research activity the document reports on and another frame containing the results of the study. Relations such as iao:isAbout connect different frames of reference within the graph and enable cross-references between them, for example, linking a methods description to the real-world system it was applied to, or linking a result to the observation context in which it was obtained. It is essential to distinguish these structural context units from the notion of context as situational information. As established in Section 2.3, context units organize and structure representations within the knowledge graph, whereas contextual information refers to instance-level descriptions of specific real-world situations. Context units do not determine whether knowledge is applicable in a given context but establish representational boundaries and relationships between frames of reference. This distinction becomes operationally critical in Chapters 4 and 5, where linking general knowledge to particular situations requires explicit contextual information and applicability conditions rather than structural framing alone. The framework further supports the representation of multiple granularity perspectives within the same knowledge graph via granularity tree units (27–29), enabling users to navigate between different types of abstractions and generalizations and their respective levels. This enables heterogeneous knowledge to be integrated and explored while preserving relationships between different levels of abstraction and alternative hierarchical representations of the same entity; for example, functional versus spatio-structural partonomies within a single coherent graph (5). Importantly, the Semantic Units Framework is technology-agnostic. It can be implemented across different infrastructures, including RDF/OWL-based knowledge graphs, property graphs, or relational systems. Its central contribution lies in the explicit representation of meaning through modular, referenceable units, rather than in any specific implementation technology (6). In this way, the Semantic Units Framework provides the formal backbone for statement-centred knowledge representation. In the following section, we consider what this foundation requires in terms of operational extension to support the transition from representation to action.

3.3 Toward Operational Knowledge Representation While semantic units provide a necessary foundation for representing structured and meaningful knowledge, they do not by themselves ensure that knowledge can be used in practice. As established in Chapter 2, enabling the use of knowledge requires more than representing statements. It requires linking statements to specific situations and supporting their contextual interpretation. In particular, the use of general knowledge (universal, prototypical, directive, and conditional directive statements) depends on relating general statements to descriptions of concrete situations and on assessing whether the conditions under which they hold are satisfied.

13 The Semantic Units Framework provides the representational basis for this integration, representing each of the required components, i.e., situational descriptions, causal and statistical knowledge, procedural knowledge, and constraints, as distinct semantic unit types, as established in Section 3.2. However, the framework does not yet prescribe how these components are systematically combined to support the transition from knowledge representation to its practical use. In particular, it does not define integrated structures that systematically link situational descriptions, causal and statistical knowledge, and potential actions through procedural knowledge. Addressing this limitation requires extending semantic units beyond representation toward operational structures that explicitly support both interpretation and action. In the following chapter, we therefore deepen the role of context and develop the conceptual basis for relating general knowledge to specific situations through explicit contextualization and applicability assessment.

4. Context and Applicability of Knowledge 4.1 Contextualization as Epistemic Process Building on the distinctions introduced in Chapter 2, actionable knowledge requires the coordinated interaction of four core components: 1.​ universal and prototypical knowledge representations specifying causal and statistical correlation relationships, 2.​ contextual information describing a specific situation, 3.​ procedural knowledge (i.e., methods, practice instructions) associated with explicit objectives, and 4.​ contextual applicability conditions specifying when knowledge can be validly used. While these components are conceptually distinct, their practical relevance emerges only through their integration. This integration is captured by the process of contextualization. Contextual information provides an instance-level description of a real-world situation, including system states, environmental conditions, and observations. It provides the empirical grounding required for interpreting and applying general knowledge. Contextual applicability conditions define the constraints under which statements or procedures can be validly applied. They specify the relationship between general knowledge and specific situations and thus enable the evaluation of whether a given piece of knowledge is relevant in a particular context. Procedural knowledge, together with explicitly defined objectives, specifies how actions can be performed to achieve desired outcomes. Contextualization is the process that relates knowledge to a specific situation by combining contextual information with applicability conditions. It enables the interpretation of knowledge in context and supports the evaluation of whether knowledge representations are relevant in a given situation. For assertional statements, contextualization is largely implicit, as they already refer to specific situations and therefore provide contextual evidence. However, additional contextual information, sometimes retrieved from intervention operations, may still be required to interpret them correctly or to support subsequent operations. For universal or prototypical statements, as well as

14 (conditional) directive statements, contextualization is essential and must be performed explicitly. The validity of these types of statements cannot be assumed but depends on whether the conditions under which they hold are satisfied in the situation at hand.

4.2 Causal Hypotheses as Context-Dependent Knowledge In biodiversity science and many other domains, much of the knowledge relevant for decision-making is expressed in the form of causal hypotheses. These describe how changes in one part of a system influence others, such as environmental drivers of biodiversity change, ecosystem responses to interventions, or interactions between species and their environment. The representation of causal hypotheses proposed here is consistent with interventionist accounts of causation (30), in which causal relationships are characterised by their invariance under interventions and the conditions under which they hold. Causal hypotheses provide the basis for anticipating the effects of interventions and therefore frequently provide the scientific justification for the effectivity of intervention operations documented as methods and practice instructions and thus procedural knowledge. However, their usefulness depends critically on whether they are applicable in a specific context. To support such use, causal hypotheses must be represented in a structured and operationalizable form that allows their components to be related to observable aspects of real-world systems. Recent work has proposed modelling them within the Semantic Units Framework as structured networks of binary relationships between causal variables, where each variable corresponds to a process affecting a quality or realizing a disposition of a specific type of entity. To be able to link causal variables to observations, measurements, or system states within the graph, these qualities and dispositions must be associated with identifiable material entities as their bearers (31). This enables a crucial integration: assertional statement units provide context-specific evidence, while universal or prototypical statement units provide generalizable causal or statistical knowledge and (conditional) directive statement units the relevant procedural knowledge. Actionable understanding depends on the interaction between these layers, mediated by applicability assessment.

4.3 Applicability as Epistemic Operationalization A central challenge for actionable knowledge is that general knowledge (universal, prototypical, directive, and conditional directive statements) is typically context-dependent. Its validity depends on factors such as environmental conditions, spatial and temporal scale, system composition, methodological assumptions, and boundary conditions. Contextual applicability conditions make this dependency explicit. They define the constraints under which a causal, statistical, or procedural statement can be validly applied and thereby establish the relationship between general knowledge and a specific situation. For universal, prototypical, directive, and conditional directive statements, this step is indispensable. Without explicit contextualization and applicability assessment, such statements often remain abstract and cannot reliably support decision-making or intervention. Applicability requirements are also operation-dependent. Different classes of operations impose different demands on contextual evaluation:

15 ●​ Transformation operations, such as unit conversions, are usually intrinsically applicable operations that only require that their input, typically assertional statements, satisfy specific data schema and format constraints. In these cases, applicability is determined by schema compatibility and thus formal applicability conditions. Consequently, transformation operations are effectively deterministic. ●​ Epistemic and intervention operations, such as applying the concept of an ecological niche or a restoration method, are contextually applicable operations and require evaluation of referential and context-specific applicability conditions and often involve uncertainty. This distinction highlights that suggesting operations based on or involving general knowledge is not sufficient for actionable understanding. A knowledge system may contain FAIR general knowledge, but without mechanisms to evaluate their applicability, it cannot determine whether their execution is valid in a given situation. Applicability conditions define formal constraints under which a statement may be used. However, satisfying these conditions does not guarantee that the statement is empirically valid in a given context. Applicability assessment therefore remains an epistemic process that may require evidence, inference, or judgment. A useful distinction can be made between applicability and actionability as two complementary forms of operationalization of general knowledge (formally defined in Box 3, Section 5.2.4). This distinction forms the conceptual foundation of the framework developed in the following chapters. Applicability refers to the epistemic operationalization of universal, prototypical, directive, and conditional directive statements, enabling the assessment of whether such knowledge is valid in a specific context (e.g., “Is causal hypothesis X applicable to context Y?”). Actionability, in contrast, encompasses transformational, epistemic, and interventional operationalization. In its interventional form, it enables the selection and execution of procedures that can modify a system in a given context to achieve a desired goal (e.g., “What can be done in context Y, based on knowledge X, to achieve goal Z?”). In this sense, applicability constitutes an epistemic precondition for interventional actionability: knowledge must first be assessed as applicable before it can be reliably used to guide knowledge-driven action, and general knowledge becomes usable only when its applicability has been established in context. This distinction is particularly important for causal hypotheses whose epistemic validity cannot be assumed but must be evaluated relative to specific contexts. Applicability alone, however, does not yet define what should be done. It determines whether knowledge can be validly used, but not how it can be operationalized to achieve a goal. As will be elaborated in the following chapter, actionability is realized through the coordinated interaction of epistemic, transformational, and intervention operations.

16

5. From Operations towards Action Units 5.1 Operations as the Missing Link between Knowledge and Action The preceding chapters established that improving the readability and interpretability of knowledge representations, as achieved by frameworks such as FAIR and CLEAR, is a necessary but insufficient condition for enabling action. The transition from interpretability to actionability requires an additional layer: the explicit representation of operations through which knowledge is used. Actionability is therefore not a property of representations alone, but emerges from the interaction between representations and the processes that operate on them. We define operations as processes that use knowledge representations to produce outputs, transform representations, or effect changes in real-world systems. In this sense, operations constitute the functional interface between knowledge and action. They specify how representations can be used, what kinds of inputs they require, and what kinds of outputs they produce. Without explicit representation of such operations, knowledge remains structurally accessible but operationally inert. The distinction between representing and intervening has deep roots in the philosophy of science. Hacking's foundational work Representing and Intervening established a principled distinction between the epistemic activities of representing the world and the experimental activities of intervening upon it, arguing that these two modes of scientific engagement are irreducible to one another and that intervention has a life independent of representation (32). The three operation classes introduced in this work can be understood as a systematic operationalization of this distinction within a knowledge representation framework, with epistemic operations corresponding to Hacking's representing dimension, establishing what is the case, while intervention operations correspond to his intervening dimension, effecting goal-directed changes in real-world systems. Transformational operations constitute a third, intermediate class that operates entirely within the representational domain, manipulating, deriving, and preparing information content entities, i.e. a class of operation that is well established in computational and data processing practice. A related distinction appears in artificial intelligence (AI) planning and robotics, where the sense–plan–act architecture (33,34) has long distinguished perception (sense), reasoning and planning (plan), and action execution (act) as three functionally distinct layers of an intelligent agent's operation. This architecture anticipates the tripartite structure proposed here in important ways, as sensing corresponds broadly to epistemic operations, planning to transformational operations over representational states, and acting to intervention operations on the world. However, the sense–plan–act paradigm was designed for autonomous agent architectures operating in real-time environments, and its primary concern is the sequencing and timing of agent behaviour rather than the representational structure of knowledge and its operational integration. The framework proposed here differs in that it addresses the prior challenge of how heterogeneous, partially formalized, and context-dependent knowledge can be represented so that all three classes of operations can be reliably performed, which is a challenge that the sense–plan–act paradigm presupposes rather than addresses.

17 The three operation classes proposed here are thus not without precedent, but, to our knowledge, they have not previously been systematically integrated into a unified knowledge representation framework that explicitly links operations to statement types, applicability conditions, and formal knowledge structures. The contribution of this section is precisely this integration. From an ontological perspective, operations can be characterized as processes involving: ●​ participants, including agents, information content entities, and material entities or processes, some of which take on specific roles, such as input and output, within the process; ●​ associated knowledge, represented as information content entities, which may either play the role of plan specifications (e.g., algorithms, methods, practice documentations) that are executed by the process, or diagnostic, causal, and statistical knowledge that informs, constrains, or justifies the process without being executed; ●​ applicability conditions that specify the constraints under which the operation can be validly performed and is expected to achieve its intended outcome, potentially under conditions of uncertainty; and a ●​ goal or objective specification that characterizes the intended outcome of the process. This characterization provides a formal basis for linking knowledge representations to their use in both human and machine contexts and applies uniformly across epistemic, transformational, and intervention actions, while allowing for systematic variation in the types of participants, applicability conditions, and goals involved. It therefore enables a systematic distinction between different classes of operations based on the types of entities they relate and transform. For the purpose of actionable knowledge systems, we distinguish three primary classes of operations that together form a minimal operational basis: epistemic operations, transformational operations, and intervention operations. These classes differ in their input–output structure, the types of entities they relate, and the forms of knowledge and capabilities they execute and realize (Fig. 2). Taken together, they define a progression from establishing knowledge about a situation, through preparing and transforming representations, to acting upon real-world systems.

5.1.1 Epistemic Operations: Relating Concepts and Knowledge to Reality Epistemic operations establish the relationship between knowledge representations and real-world entities, processes, and situations. They employ the diagnostic capabilities required to determine what is the case in a given context and whether specific knowledge applies. In this sense, epistemic operations constitute the grounding layer of actionable knowledge, linking information content entities to material entities, their states, and the processes they are involved, and vice versa (Fig. 2a). Figure 2: Three classes of operations underlying actionable knowledge. Epistemic, transformational, and intervention operations differ in their input–output structure and in the types of knowledge they execute. a) Epistemic operations relate information content

18 entities (ICE) to material entities or processes (here not shown), using ‘rdf:type’ for terms and ‘iao:is about’ for statements, and execute diagnostic knowledge, which is the basis for establishing referential lexical competence. b) Transformational operations operate on representations, taking ICE as both input and output, and execute algorithms that realize computational functions. c) Intervention operations act on material entities, transforming them from an initial to a target state by executing procedural knowledge and realizing goal-directed capabilities. Together, these operation types form the operational backbone of actionable understanding, linking knowledge representation, data transformation, and real-world action. The properties used in this figure are in line with upper-level ontologies such as the Basic Formal Ontology (BFO) with its OBO Relations Ontology (RO) and extensions such as the Statistical Methods Ontology (STATO), the Ontology for Biomedical Investigations (OBI), and the Information Artifact Ontology (AIO) (see refs (9,35)). For instance the relation ‘stato:executes’ is understood as a derived relation (a chained property combining ro:realizes and ro:concretizes) linking a process to a plan specification via the intermediate realization of some plan that concretizes the specification, whereas ‘iao:is about’ is a representation relation between an ICE and the entity this information is about.

From an ontological perspective, epistemic operations are processes that relate representations and reality in a bidirectional manner. They may take an information content entity as input and identify a corresponding instance in the world, or take a material entity or situation as input and produce a representation that describes it. Through these processes, epistemic operations establish type and aboutness relations between representations and the entities or situations they refer to, thereby enabling the evaluation of applicability (Fig. 2a). Epistemic operations can be systematically characterized along two dimensions: the level of representation (terms vs. statements) and the direction of the operation (from representation to world, or from world to representation). At the term level, epistemic operations support referential use of concepts and include: ●​ Recognition: given a term, identify the corresponding instance or occurrence in the world. ●​ Designation: given an observed instance, identify the appropriate term or concept. These operations realize referential lexical competence, enabling agents to connect conceptual representations to concrete instances. They depend on diagnostic knowledge, such as method-dependent recognition criteria, measurement protocols, or classification rules, which specify how instances can be identified in practice. At the statement level, epistemic operations extend this grounding to propositional knowledge. They include: ●​ Recognition: given a statement, identify the situation or context in which it holds. ●​ Description: given a situation, identify or construct a statement that accurately describes it. These operations apply across different types of statements. For assertional statements, which describe specific observations or situations, epistemic operations enable the alignment between representations and empirical reality, providing contextual information and evidence. For general statements, epistemic operations play a crucial role in evaluating applicability. This includes: ●​ Universal and prototypical statements, expressing causal or statistical relationships, whose validity must be assessed relative to specific conditions; and ●​ Directive and conditional directive statements, representing procedural knowledge, whose applicability depends on whether the required preconditions are satisfied in a given context. In the case of such general statements, epistemic operations establish whether assertional statements can be interpreted as instances or realizations of them. This relation is analogous to the

19 instance–class relation at the term level, but differs in that it is mediated by applicability conditions. Determining whether a situation represented by an assertional statement instantiates a general statement therefore requires evaluating whether the relevant contextual and methodological conditions are satisfied. In all cases, epistemic operations enable the assessment of whether a statement can be validly related to a particular situation. This process often relies on structured representations that make the components and constraints of statements explicit. Formalized statement schemata, such as constraint-based models (e.g., SHACL shapes), support consistent interpretation across contexts and facilitate both human reasoning and machine-based matching and querying. Epistemic operations can be characterized as processes involving the following components (Fig. 2a): ●​ participants, including information content entities and material entities or processes, which take on the roles of input and output in either direction (representation → recognition operation → real-world part; real-world part → description and designation operation → representation); ●​ diagnostic procedural knowledge, represented as information content entities, which plays the role of plan specifications that are executed by the process; ●​ referential applicability conditions, which determine whether a representation can be validly related to a material entity or system state and vice versa using diagnostic criteria, measurement protocols, or classification rules and are often method-dependent; and an ●​ epistemic goal or objective specification, namely the establishment of a correct and adequate relationship between representations and real-world entities or states. A defining characteristic of epistemic operations is that they realize epistemic or cognitive dispositions, in particular referential lexical competence and related inferential capabilities. While these operations are supported by diagnostic knowledge, they are usually not fully determined by explicit plan specifications in the way that algorithms or procedures determine transformational or intervention processes. Instead, they involve context-sensitive interpretation, evaluation, and judgment. For human agents, epistemic operations rely on interpretation, experience, and domain expertise. For machine agents, they are implemented through classification, pattern matching, constraint checking, and querying over formally structured representations. Despite these differences, their functional role is the same, i.e., to establish a reliable relationship between knowledge representations and the world. Epistemic operations thus answer the fundamental question: “What is the case in this situation, and does this knowledge apply here?” They provide the necessary foundation for both transformational operations, which prepare and integrate representations, and intervention operations, which act upon real-world systems. Without epistemic grounding, knowledge cannot be reliably applied and therefore cannot support actionable understanding. Epistemic operations may, however, depend on prior intervention operations when access to diagnostically relevant properties of a material entity requires active manipulation. In many scientific and practical settings, the identification or characterization of an entity presupposes preparatory procedures, such as dissection, staining, chemical treatment, or measurement protocols. These preparatory steps are themselves intervention operations that modify the material entity to render

20 specific features observable. In such cases, epistemic operations are embedded in composite processes in which intervention operations provide the necessary conditions for subsequent recognition or description. This highlights an important dependency: epistemic access to the world is often mediated by intervention. A further important class of epistemic operations concerns the application of general knowledge to specific situations. Evaluating whether a universal, prototypical, directive, or conditional directive statement holds in a given context constitutes a form of epistemic operation that operates on representations grounded in reality. These processes take general knowledge and contextual information, typically represented by assertional statements, as input and produce judgments about applicability. In this sense, applicability assessment is itself an epistemic operation, enabling the evaluation of whether general knowledge is valid in a specific situation.

5.1.2 Transformational Operations: Processing Representations While epistemic operations establish the relationship between knowledge representations and real-world situations, a second class of operations operates entirely within the domain of representations. These transformational operations are processes that take information content entities as input and produce modified or derived information content entities as output (Fig. 2b). Their primary function is to prepare, manipulate, and derive representations that can subsequently support epistemic interpretation and intervention. Transformational operations are thus information-processing processes that establish relationships between representations by transforming them from one form into another. Typical examples include data transformation, aggregation, filtering, feature extraction, format conversion, graph-based transformations, statistical analyses, and visualizations. In each case, the operation produces a new representation that is related to the input through a well-defined transformation. From an ontological perspective, transformational operations are characterized by the following components (Fig. 2b): ●​ participants, with information content entities serving as both input and output, where input representations are transformed into output representations; ●​ methods and algorithms, represented as information content entities, which function as plan specifications executed by the process; ●​ formal applicability conditions, defined by schema constraints, which determine whether a given transformation operation can be validly applied to information content entities in the input role; and a ●​ transformational goal or objective specification, namely the production of representations in a form that satisfies specific structural, semantic, or functional requirements. In contrast to epistemic operations, transformational operations do not directly relate representations to material entities or processes. Instead, they operate exclusively within the representational domain, achieving the transformational goal of producing representations in a form that satisfies specified structural, semantic, or functional requirements. Transformational goals are representational in nature, but are realized through transformation processes. Their applicability is therefore typically determined by formal constraints on representations, such as schema compatibility, data types, or structural requirements. These constraints can be expressed as formal applicability conditions of the form: if an input representation conforms to schema X, then operation

21 Y can be validly executed. In contrast to epistemic and intervention operations, whose applicability depends on context-specific properties of real-world systems and is often uncertain, the applicability of transformational operations is largely deterministic and can be evaluated through formal validation. In many cases, this makes their execution largely deterministic, provided that the required input conditions are satisfied. Transformational operations play a critical enabling role in actionable knowledge systems. They provide the means to integrate heterogeneous data sources, derive new representations from existing ones, and prepare information in forms that are suitable for interpretation or action. For example, raw observational data may be aggregated into summary statistics, converted into standardized formats, or transformed into features that can be used in classification or decision processes. Transformational operations can be applied to different types of statements, including general knowledge. In practice, they frequently operate on assertional statements that represent instance-level data, for example when aggregating observations, filtering datasets, or deriving new representations from empirical records. However, transformational operations may also process general knowledge, such as formalizing causal relationships, translating statistical models, or compiling procedural knowledge into executable workflows. Importantly, in contrast to epistemic operations, the role of statement types in transformational operations is not constitutive but contextual. Transformational processes do not evaluate whether a statement is true, applicable, or relevant in a given situation; rather, they manipulate representations according to formally specified rules. Their applicability is therefore determined by structural compatibility and formal constraints, rather than by contextual or epistemic considerations. In this sense, transformational operations are agnostic to the epistemic status of the statements they process, even though the resulting representations may subsequently be used in epistemic or intervention contexts. This distinction reflects a fundamental separation between the processing of representations and their interpretation and application. This way, transformational operations do not themselves establish whether a representation correctly describes a real-world situation, nor do they directly effect changes in such systems. Instead, they support these tasks indirectly by producing representations that can be used in epistemic and intervention operations. In this sense, they form an intermediate layer between the grounding of knowledge in reality and its application to achieve specific goals. Transformational operations thus answer the question: “How can available representations be processed or transformed so that they can support interpretation or action?” They are an indispensable component of the operational backbone of actionable knowledge, enabling the preparation and derivation of representations required for both epistemic grounding and intervention.

5.1.3 Intervention Operations: Acting on Systems Intervention operations are processes that deliberately modify the state of a material system in order to achieve a specified goal. They represent the culmination of actionable knowledge, translating epistemically grounded and transformation-enabled representations into concrete changes in the world. From an ontological perspective, intervention operations are material processes that take a material entity in a given state as input and produce a modified material entity as their output (Fig.

22 2c). These processes are goal-directed and causally efficacious, and they are guided by information content entities that specify how the transformation of the system is to be achieved. More specifically, intervention operations are characterized by the following components (Fig. 2c): ●​ participants, with material entities serving as both input and output, where a system in an initial state is transformed into a system in a modified state; ●​ procedural knowledge, represented as information content entities, which function as plan specifications executed by the process and guiding the transformation of the system; ●​ contextual applicability conditions, which determine whether a procedure can be validly applied in a given situation based on relevant constraints and on ●​ contextual information, typically represented by assertional statements describing relevant aspects of the situation; and an ●​ intervention goal or objective specification, namely the realization of a desired change in the state of a material system. Within the typology of statement types introduced earlier, intervention operations rely on the coordinated interaction of assertional statements (instance-level) providing situational grounding and describing the current state of the system, universal or prototypical statements providing causal or statistical relationships that justify or inform potential interventions, and directive and conditional directive statements (procedural knowledge) specifying how actions are to be performed to achieve desired outcomes. A critical principle follows from this integration: general knowledge is not directly actionable. Universal and prototypical statements, such as causal hypotheses or statistical regularities, as well as directive and conditional directive statements, such as method specification and practices instructions, become operational only when they are embedded in structures that specify how they can be used to achieve a goal under defined conditions. This requires linking them to contextual information. This linkage is mediated by contextual applicability conditions, which define the constraints under which a given piece of knowledge can be validly used in a specific situation. Operationalizing applicability therefore requires connecting general knowledge to observable properties of real-world systems. Because the variables of causal hypotheses correspond to changes of qualities of material entities, they can be related either directly or indirectly to concrete observations, measurements, or derived indicators represented as assertional statements (31). This enables the formulation of queries that match applicability conditions against contextual information, allowing systems to identify relevant empirical evidence, evaluate whether the conditions for applying a given procedure are satisfied, and detect missing information required for decision-making. Moreover, explicitly defined contextual applicability conditions enable the formulation of counterfactual scenarios, which are essential for testing context dependence and refining hypotheses. By varying contextual parameters and evaluating their effects on the applicability of a statement, it becomes possible to systematically investigate the limits of universal and prototypical knowledge and specify corresponding applicability conditions. Failures to account for context-dependent applicability are not merely theoretical concerns but have been observed across biodiversity science and conservation practice, and they are a major source of ineffective or misguided interventions, particularly in complex systems such as ecological

23 environments. General ecological knowledge is often transferred across systems and contexts without sufficient consideration and evaluation of environmental conditions, scale dependencies, or system-specific constraints, leading to ineffective or even counterproductive interventions. These cases highlight the practical importance of explicitly representing contextual applicability conditions. Examples and their reinterpretation through the proposed framework are provided in Chapter 7. Through applicability conditions based applicability assessment, a systematic connection is established between instance-level data and general knowledge, enabling the context-sensitive selection and execution of interventions. Intervention operations therefore answer the question “What can be done in this situation to achieve a desired outcome?” However, this question can only be addressed reliably if a prior epistemic step has been completed. Applicability assessment constitutes an epistemic precondition for intervention operations. Before a procedure can be selected or executed, it must be established that the underlying knowledge, particularly causal and procedural knowledge, is applicable in the given context. This dependency highlights the sequential and interdependent nature of the three operation types: ●​ epistemic operations establish the relationship between knowledge and the current situation, ●​ transformational operations prepare and derive the representations required for analysis, ●​ intervention operations use this integrated knowledge to effect change in the system. Explicit representation of contextual applicability conditions, and their integration into intervention operations, is essential for ensuring that actions are not only executable but also contextually valid.

5.1.4 A Minimal Operational Framework for Actionable Knowledge The three classes of operations introduced above, i.e., epistemic, transformational, and intervention operations, together define a minimal operational framework for actionable knowledge systems. Each class fulfils a distinct but complementary role in enabling the transition from knowledge representation to effective knowledge-driven action. All three classes of operations are goal-directed but differ in the nature of their objectives and their functional role in the progression from knowledge to action. Epistemic operations pursue epistemic goals, i.e., establishing correct relationships between representations and real-world entities or states, and thereby ground knowledge in specific situations. Transformational operations pursue transformational goals, i.e., producing representations that satisfy specified structural, semantic, or functional requirements, and thereby prepare and derive representations required for subsequent operations. Intervention operations pursue intervention goals, i.e., achieving desired changes in the state of a material system, and thereby translate epistemically grounded and representationally prepared knowledge into concrete action. Together, these three operation types define a functional progression from situational understanding, through representational transformation, to goal-directed system change. However, this progression should not be understood as strictly linear, as the three operation classes are not interdependent at the level of sequences of distinct operations but can also be constitutively composed within a single higher-order operation.

24 An operation of one class may require, as a constitutive step, an operation of a different class whose completion is a necessary precondition for the higher-order operation to proceed. A recognition operation, for example, may require as its first step an intervention operation that prepares and transforms the material entity to be recognized, as in histological tissue preparation and staining, where the material must first be dissected and then chemically modified before its composition can be correctly identified by microscopy (Fig. 3). In such cases, the intervention is not a separate preceding operation but a constitutive component of the recognition process itself. Similarly, a description operation may require prior transformational steps that derive or aggregate representations into a form suitable for the intended description, and an intervention operation may require a prior epistemic step that explicitly assesses whether its applicability conditions are satisfied before the intervention can be validly executed.

Figure 3: A granular composite operation spanning multiple operation classes. A coarser-grained epistemic recognition operation, i.e., Histological Tissue Composition Analysis, decomposes into two constitutive ordered steps of different operation classes, connected by a precedes relation. Step 1 is an intervention operation (Tissue Preparation and Staining) that takes the untreated tissue sample as input and produces a stained tissue section as output, governed by contextual applicability conditions and a preparation and staining method. Step 2 is an epistemic recognition operation (Compositional Identification) that takes the stained tissue section as material input together with diagnostic knowledge and referential applicability conditions, and produces a compositional description as output. The coarser-grained operation has both component operations as parts, takes the untreated tissue sample as its overall input, and produces the compositional description as its overall output. The intervention operation in step 1 constitutes a necessary precondition for the applicability of the epistemic operation in step 2. Without prior staining, the referential applicability conditions for compositional recognition cannot be satisfied. For clarity, the executes relations, agent participants, and referential applicability conditions of the coarser-grained parent operation are not shown; these follow the same structural pattern as the component epistemic operation and are inherited from the component operations through the has part relation (see Figure 2). This cross-class compositional structure illustrates that a higher-order operation of one class may constitutively require a step of a different class as a necessary precondition, and that the three operation classes are not mutually exclusive at the compositional level. The compositional structure is carried through to the representational level as composite action units (Section 5.2.1).

This compositional structure reflects a fundamental feature of knowledge-driven action in complex systems. The boundary between representing and intervening, in Hacking's sense (32), is not a clean sequential boundary but is itself subject to operational composition. Operations are therefore not only classifiable by their primary type but may also be granular, with a coarse-grained operation of one class decomposing into an ordered sequence of finer-grained operations, potentially of mixed types, whose collective execution realizes the coarser-grained operation's goal. The granularity of operations is governed by the same part-whole logic that applies to compound semantic units more broadly, and is carried through to the representational level as composite action units (Section 5.2.1). This interdependence reveals a central principle: knowledge becomes actionable not through representation alone, but through its integration into operations that connect representations to

25 real-world situations, transform them into usable forms, and guide context-sensitive system change. Actionable understanding thus emerges from the coordinated interplay of epistemic grounding, representational transformation, and goal-directed intervention (see Box 3 in Section 5.2.4 for definitions and characterizations of all key concepts relating to actionable knowledge). The operational architecture defined by these three classes of operations is inherently iterative. Outcomes of intervention operations generate new observations and system states, which are reintroduced as assertional statements. These, in turn, feed into epistemic operations, enabling refinement of contextual understanding, reassessment of applicability conditions, and improvement of procedural and transformational knowledge. This feedback structure supports continuous learning and adaptation of knowledge systems in response to changing environments and incomplete information, which is particularly critical in complex domains such as biodiversity science and ecological systems. The relationship between actionability and applicability within this architecture is developed as an operational continuum in Sections 5.2.5 and 6.2. In the following section, we build on this operational perspective by introducing Action Units as integrated knowledge structures that explicitly combine semantic content, contextual information, goals, procedural knowledge, and applicability conditions. Action Units provide the representational counterpart to the operational framework developed here and enable the systematic linkage between knowledge and action.

5.2 Action Units: Definition and Scope Building on the Semantic Units Framework and the operational distinction introduced in Section 5.1, we introduce action units as the representational structures that link knowledge representations to their potential use in operations. Formally, action units can be understood as structured extensions of plan specifications—in the broad OBO/BFO sense of information content entities that specify how a process is to be carried out—that make applicability conditions and contextual grounding explicit as first-class, typed, and evaluable semantic unit components. While semantic units provide modular and context-aware representations of meaning, they do not by themselves specify how these representations can participate in processes. Action units address this gap within the Semantic Units Framework by introducing this missing operational layer—explicitly representing how information content entities and, where relevant, material entities can serve as inputs and outputs of processes, and how these processes are guided or enabled by associated knowledge.

5.2.1 General Definition of Action Unit While operations are processes unfolding in time, action units provide structured representations of how such processes can, in principle, be carried out. As extended plan specifications, they do not introduce a new type of operation, but rather provide a level of representation that integrates and coordinates epistemic, transformational, and intervention operations within a unified structure. Formally, an action unit is a compound unit that specifies the structural components of an operation as a specific set of semantic units (Fig. 4): ●​ input and output units, which specify the participants of the process with their respective roles and may comprise information content entities and/or material entities;

26 ●​ plan specification units, comprising information content entities that represent diagnostic and procedural knowledge, including methods and algorithms, to be executed by the process; ●​ an applicability conditions unit, which specifies the constraints under which the operation can be validly performed, including formal (schema-based), referential, or contextual conditions; and ●​ an objective unit, which characterizes the intended outcome of the operation. Box 2 provides a consolidated reference for the core representational components introduced across Chapters 2–5 that together constitute the building blocks of actionable knowledge representation. Note that the plan specification unit listed above refers specifically to the semantic unit component of an action unit that carries the procedural or diagnostic knowledge executed by the process, such as methods, algorithms, or recognition criteria. This is distinct from the broader sense in which action units themselves constitute extended plan specifications, as action units at the compound unit level specify the full operational structure of a process including its applicability conditions, contextual grounding, and objectives, whereas plan specification units sensu stricto at the component level carry only the procedural knowledge that the process executes. In contrast to other types of semantic units, which represent meaning in isolation, action units explicitly encode the operational roles that entities assume within processes and how knowledge participates in them. They thereby establish a structured link between knowledge representations and the operations through which they can be applied. Importantly, action units do not merely represent isolated operations. They provide integrative structures that can coordinate multiple classes of operations. In particular, they may embed epistemic operations that establish relationships between representations and specific situations, as well as transformational operations that prepare and derive the representations required for interpretation and execution, in order to enable intervention operations that effect changes in material systems. Action units can also be granular and hierarchically composed. A higher-order action unit may consist of an ordered sequence of lower-order action units, potentially of different types, whose outputs serve as inputs to subsequent steps and whose collective execution realizes the higher-order unit's objective. As illustrated at the operation level in Section 5.1.4 and Figure 3, the boundary between operation classes is not a barrier to composition. A higher-order epistemic action unit may, for instance, require an intervention action unit as its first constitutive step when access to diagnostically relevant properties of a material entity depends on prior physical manipulation. In such cases, the applicability assessment that grounds the higher-order operation is itself an active representational object, i.e., an epistemic action unit with its own inputs, outputs, plan specification, and goal, rather than merely an implicit constraint. The granularity of action units follows the same part-whole logic that governs operations and compound semantic units more broadly, with explicit ordering relations between component units encoded in the plan specification unit of the higher-order action unit. The plan specification unit here encodes both the procedural knowledge guiding execution and the temporal ordering constraints on component action units, reflecting the compound nature of higher-order action unit specifications.

27

Figure 4: Action unit. A compound unit consisting of several distinct types of semantic units. Arrows indicate 'has part' relations directed from the compound unit to its component semantic units parts. Action units can also be granular and hierarchically composed, with a coarser-grained action unit having finer-grained action units as its ordered component parts (here not shown). This composite structure is illustrated in Figure 3 and discussed in Section 5.2.1.

In this sense, action units encode structural actionability, i.e., they define what operations can, in principle, be performed given a particular representation, independent of a specific situation. Whether such operations can be validly carried out in a given context depends on the evaluation of their applicability conditions. Conceptually, action units can thus be understood as a structured extension of plan specifications in the broad sense, that is, of information content entities that specify how processes are to be carried out. Within this broader structure, the plan specification unit serves as one specific component, carrying the procedural or diagnostic knowledge to be executed. The novelty of action units lies in making the remaining components, i.e. applicability conditions, contextual information, and objectives, equally explicit, modular, and referenceable as first-class semantic units. While plan specifications in the OBO sense may already include objectives, conditions, and input–output specifications, they do not distinguish structural actionability from contextual applicability, nor do they represent applicability conditions and contextual information as typed, evaluable semantic unit components. Action units address precisely this gap by integrating plan specifications with assertional and general statements, contextual information, and applicability conditions, thereby supporting their interpretation, evaluation, and coordinated use across different types of operations. Based on the three classes of operations distinguished in Section 5.1, action units can be specialized into three corresponding types, reflecting their primary operational role. Box 2 Core Components of Actionable Knowledge Representation (new semantic unit types introduced in this work are explicitly indicated) Semantic Units and Action-Oriented Extensions Semantic Unit A modular, self-contained, and referenceable representation of meaning, typically corresponding to a single statement (statement unit) or a structured collection of statements (compound unit). Semantic units constitute the foundational units of knowledge representation. Context Unit A semantic unit that defines a frame of reference within a knowledge graph. It is a compound unit that organizes other semantic units into a coherent subgraph (e.g., document, experiment, real-world system) and specifies representational boundaries, but does not determine applicability.

28

Contextual Information Semantic units representing instance-level information about a concrete situation, including system states, environmental conditions, and observational data. Contextual information provides the situational input required to interpret and apply knowledge. It may be retrieved from the knowledge graph (e.g., via query-based question units) or provided by an agent in a specific decision context. Applicability Condition Unit (introduced in this work) A semantic unit that specifies the conditions under which a statement can be validly applied. These compound units relate general knowledge to specific situations and can be operationalized as executable question units that evaluate contextual information. They provide the bridge between general knowledge and situational applicability. Action Unit (introduced in this work) A semantic unit that integrates semantic content (i.e., knowledge) with contextual information, explicit goals, procedural knowledge, and applicability conditions. Action units extend plan specifications within the Semantic Units Framework by making applicability conditions and contextual information explicit as first-class, typed, and evaluable semantic unit components. Action units represent knowledge in a structurally actionable form that supports both epistemic and intervention actions. They define how an action can, in principle, be performed, independent of whether it is applicable in a specific situation. Action units can also be granular and hierarchically composed, with a higher-order action unit consisting of ordered lower-order action units of potentially different types (see Section 5.2.1). Applicable Action Unit (introduced in this work) A semantic unit that links an action unit to contextual information for which its applicability conditions are satisfied or inferred or assumed to be satisfied. Applicable action units represent contextually applicable knowledge, indicating that an action can, in principle, be performed in a given situation based on available information and reasoning. Applicable action units may exist without empirical validation; they represent hypothesized or inferred applicability, whereas validated applicable action units represent empirically supported applicability. Validated Applicable Action Unit (introduced in this work) A semantic unit representing an applicable action unit whose applicability is empirically supported, for example through prior successful applications, observational data, or experimental validation. These units document contexts in which an action has been shown to work and provide a basis for assessing reliability and transferability.

5.2.2 Epistemic Action Units Epistemic action units specify the structural components of operations that establish relationships between representations and real-world entities or system states. They provide the representational counterpart to epistemic operations such as recognition, designation, and description by defining how information content entities and material entities participate in such processes. Following the general definition of action units, epistemic action units comprise the following specialized components: ●​ Input and output units, which include both information content entities and material entities or processes, taking on the roles of input and output in either direction. In recognition processes, representations (terms or statements) serve as input and material entities or situations as output. In designation and description processes, material entities or situations serve as input and representations as output. ●​ Diagnostic plan specification units, representing diagnostic procedural knowledge in the form of recognition criteria, classification rules, or observational methods. These information content entities define how representations are to be related to material entities and are executed by the process. ●​ A referential applicability conditions unit, specifying the conditions under which a representation can be validly related to a material entity or system state. These conditions constrain the correct application of diagnostic criteria and may depend on methodological, observational, or contextual factors.

29 ●​ An epistemic objective unit, specifying the epistemic goal of the operation, namely the establishment of a correct and adequate relationship between representations and real-world entities or system states. Epistemic action units thus provide a structured representation of how representations can be grounded in reality. They define the conditions under which terms designate entities and statements describe situations, and they support the evaluation of whether instance-level and general statements can be validly related to a given context.

5.2.3 Transformational Action Units Transformational action units specify the structural components of operations that transform information content entities into other information content entities. They provide the representational counterpart to transformational operations by defining how representations are processed, manipulated, and derived within the space of information content entities. Following the general definition of action units, transformational action units comprise the following specialized components: ●​ Input and output units, which consist of information content entities serving as both input and output, where input representations are transformed into output representations. ●​ Method and algorithm plan specification units, representing procedural knowledge in the form of computational functions, methods, algorithms, and workflows. These information content entities define how input representations are to be transformed and are executed by the process. ●​ Formal applicability conditions unit, specifying the schema-based constraints under which a transformation can be validly applied. These conditions determine whether input representations satisfy the structural, syntactic, or semantic requirements necessary for executing the transformation. ●​ Transformational objective unit, specifying the goal of the operation, namely the production of representations that satisfy specific structural, semantic, or functional requirements. Transformational action units capture operations such as aggregation, filtering, inference, format conversion, statistical analysis, and visualization. They are representation-agnostic with respect to statement type and provide the technical means for preparing and deriving data required for epistemic evaluation and intervention. Transformational action units thus provide a structured representation of how information content entities can be systematically processed and transformed, defining the conditions under which representations can be converted, combined, or derived. They support the preparation and integration of both instance-level and general statements for subsequent epistemic or intervention operations. In this sense, transformational action units answer the question of how input representations can be technically transformed into forms that are usable for epistemic evaluation or intervention.

5.2.4 Intervention Action Units Intervention action units specify the structural components of operations that deliberately change the state of material systems in order to achieve explicit goals. They provide the representational

30 counterpart to intervention operations by defining how material entities and relevant knowledge participate in processes that effect system change. Following the general definition of action units, intervention action units comprise the following specialized components: ●​ Input and output units, which consist of material entities serving as both input and output, where a system in an initial state is transformed into a system in a modified state. ●​ Procedural plan specification units, representing procedural knowledge in the form of methods, practices, or workflows, typically expressed as directive and conditional directive statements. These information content entities define how the system is to be transformed and are executed by the process. ●​ Contextual applicability conditions unit, specifying the conditions under which a procedure can be validly applied in a given situation. These conditions define constraints related to the system, its environment, and the intended intervention, and they determine whether the execution of the procedure is appropriate in context. ●​ Contextual information units, representing instance-level information about the current state of the system, typically expressed as assertional statements. These units provide the empirical basis for evaluating applicability conditions and selecting appropriate procedures. ●​ Intervention objective unit, specifying the goal of the operation, namely the realization of a desired change in the state of a material system. Intervention action units thus provide a structured representation of how material systems can be changed in a goal-directed manner. They define how contextual information, general knowledge, and procedural specifications are integrated to enable the selection and execution of interventions under context-specific constraints. In this sense, intervention action units are the representational structures through which world-to-word directive knowledge becomes operationally integrated with contextual information and applicability conditions. In contrast to epistemic and transformational action units, intervention action units are inherently integrative, as they require the coordinated alignment of contextual information, applicability conditions, and procedural knowledge to achieve a specified goal. Together, these three types of action units reflect the fundamental ways in which knowledge can participate in operations: by grounding representations in reality (epistemic), transforming representations (transformational), and guiding goal-directed changes in material systems (intervention). In this sense, action units operationalize the transition from epistemic evaluation to interventional action by integrating transformation processes and applicability assessment within a unified representational structure. The key concepts introduced in this work (actionability, applicability, the three forms of actionable knowledge, and actionable understanding) are defined and characterized in Box 3. Box 3 Key Concepts Introduced in This Work Actionability​ The extent to which a given representation can be used by an agent to perform epistemic, transformational, or intervention actions. Actionability depends on interpretability and the availability of operations that can process the representation as input. Actionability is not a Boolean property, but a multidimensional spectrum determined by:​ (i) agent type (human or machine), reflecting differences in interpretative and operational capabilities;​

31

(ii) meaning-bearing representation level, distinguishing between non-propositional representations (terms) and propositional, meaning-bearing representations (statements); and ​ (iii) statement generality (for propositional representations), distinguishing between assertional (instance-level) statements and universal or prototypical as well as directive and conditional directive (type-level) statements.​ These dimensions jointly determine the degree and type of actionability. Notably, statement generality is only defined for propositional representations. Applicability​ The extent to which a given knowledge representation can be validly applied in a specific context. Applicability depends on the relationship between contextual information and the applicability conditions associated with the representation. Like actionability, applicability is not a Boolean property but graded and often uncertain. In many cases, it can only be assessed with varying degrees of confidence, depending on the completeness of contextual information, the validity of underlying assumptions, and the availability of empirical evidence, reflecting ecological uncertainty and the challenges of real-world decision-making. Applicability can be understood as the epistemic characterization of actionability in context. While actionability describes the structural potential for action, applicability determines whether such action can be validly realized in a given context. Applicability requirements depend on the type of operation: some operations are intrinsically applicable based on representational compatibility, whereas others require context-dependent evaluation of applicability conditions. Actionable knowledge​ Knowledge that enables one or more classes of operations—epistemic, transformational, or intervention—depending on the agent, representation type, and context. It is restricted to propositional representations (statements), as isolated terms do not carry propositional knowledge independently. Actionable knowledge includes both human- and machine-actionable statements, and takes three forms depending on the class of operation it supports: epistemically actionable knowledge, transformationally actionable knowledge, and interventionally actionable knowledge, each defined below. Epistemically actionable knowledge​ A form of actionable knowledge that supports epistemic operations such as recognition, designation, description, and the evaluation of whether a statement applies to a given situation. For general statements (universal and prototypical as well as directive and conditional directive statements), epistemic actionability is realized through applicability assessment, i.e., determining whether the referential applicability conditions under which the statement holds are satisfied in a specific context. Epistemically actionable knowledge enables the question: “Does this knowledge apply in this situation?” Transformationally actionable knowledge​ A form of actionable knowledge that supports transformational operations on representations, such as data processing, aggregation, filtering, inference, or reformatting. It enables the transformation of representations into forms that support subsequent epistemic interpretation or intervention operations. Transformational actionability depends on the compatibility between input and output representations (formal applicability conditions), as well as the availability of operations (e.g., workflows, scripts, or graph-based transformations) that can be applied to them.​ Unlike epistemically actionable knowledge, which relates representations to real-world situations, and interventionally actionable knowledge, which supports changes in real-world systems, transformationally actionable knowledge operates within the space of representations (information content entities) themselves. It enables the question: “How can this representation be transformed into a form that supports interpretation or action?” Interventionally actionable knowledge​ A form of actionable knowledge that supports intervention operations aimed at modifying a system to achieve a specified goal. It requires the integration of semantic content, contextual information, procedural knowledge, explicit goals, and contextual applicability conditions. Interventional actionability depends on prior applicability assessment. Interventionally actionable knowledge enables the question: “What can be done in this situation to achieve a desired outcome?” Actionable understanding​ A form of human-centred actionable knowledge that constitutes the capacity to interpret knowledge, assess its applicability in a given context, and select and justify appropriate actions accordingly. It rests on two constitutive components: epistemic operations, which establish the relationship between a knowledge representation and the situation at hand, and intervention-oriented reasoning, which determines appropriate courses of action given that relationship. These two components are mutually dependent: epistemic grounding without intervention-oriented reasoning yields interpretation without action, while intervention-oriented reasoning without epistemic grounding yields action without justification. Actionable understanding may also depend on transformational operations that prepare or derive the representations

32

required for epistemic interpretation and decision-making. These are enabling rather than constitutive: they support the conditions under which actionable understanding can arise, but do not themselves constitute it.​ For general statements (universal, prototypical, directive, and conditional directive), actionable understanding requires explicit applicability assessment, determining whether the referential and contextual applicability conditions under which a statement holds are satisfied in the specific situation at hand. In this sense, actionable understanding is not only a prerequisite for action but is itself refined and often achieved through the application of knowledge in concrete contexts.

5.2.5 From Structural Actionability to Contextual Validity A key implication of the proposed framework is that actionability and applicability must be clearly distinguished. Action units define what can be done structurally, whereas the validity of performing an action in a specific situation depends on whether applicability conditions are satisfied. This distinction is critical. A knowledge representation may be fully actionable in a structural sense, i.e., providing well-defined inputs, procedures, and goals, while still requiring epistemic evaluation to determine whether it is applicable in a given context. To capture this progression from structural potential to context-specific validity, we distinguish three levels of operational grounding (which are formally characterised in Box 2, Section 5.2.1): ●​ Action units, which encode structurally actionable knowledge and define what can be done in principle. ●​ Applicable action units, which link action units to contextual information for which applicability conditions are satisfied or inferred to hold, defining what can be done in a given situation. ●​ Validated applicable action units, which represent applicable action units whose validity is supported by empirical evidence, defining what has been shown to work in comparable contexts. These levels reflect successive stages from structural specification to contextual validity and empirical support. They highlight that actionability is not an inherent property of knowledge alone, but emerges from the interaction between representations, operations, and context. In this structure, objectives define what should be achieved, procedures define how it can be achieved, applicability conditions define when it can be achieved, contextual information defines the current situation. When these components align, applicable action units emerge. More generally, actionability should be understood as a graded and multi-dimensional spectrum rather than a Boolean property (12). Knowledge representations may support different types of operations to varying degrees, depending on their level of formalization, the availability of associated plan specifications, and the explicitness of applicability conditions. At one end of the spectrum, representations may be interpretable but not operational, supporting epistemic understanding without enabling further use. At intermediate levels, representations become actionable with respect to specific operation types, for example by supporting epistemic grounding or enabling transformation processes. At the highest level, knowledge is fully operationalized in the form of intervention action units that integrate contextual information, applicability conditions, procedural knowledge, and explicit goals. The practical implications of this graded and multi-dimensional spectrum for knowledge infrastructure design, including the principle of boundedness and incremental expansion, are developed in Section 8.6.

33

6. Architecture and Operational Perspectives of Actionable Knowledge 6.1 Architectural Synthesis The preceding chapters introduced the representational and conceptual foundations of actionable knowledge, including statement types, contextualization and applicability, operation classes, and action units as integrative structures. Taken together, these elements define a coherent architecture in which knowledge representation and operational use are systematically connected. At the representational level, different types of statements fulfil distinct roles. Assertional statements provide instance-level descriptions of concrete situations, while general statements, including universal, prototypical, directive, and conditional directive statements, capture generalizable and procedural knowledge. Applicability conditions mediate between these layers by specifying the constraints under which general knowledge can be validly related to specific contexts. Action units integrate these components as extended plan specifications, with input–output specifications, plan specifications sensu stricto, and objectives, thereby encoding the structural conditions under which operations can be performed. At the conceptual level, epistemic, transformational, and intervention operations define how knowledge is related to the world, transformed into usable forms, and applied to achieve desired outcomes. Action units serve as the representational counterparts of these operations, specifying how knowledge participates in processes without presupposing a specific situation. This architecture establishes a systematic relationship between: ●​ ●​ ●​ ●​

representation (semantic units and statement types), contextualization (applicability and contextual information), operations (epistemic, transformational, intervention), and integration (action units as extended, structured plan specifications).

Actionable knowledge thus emerges not from individual components, but from their coordinated interaction within this architecture.

6.2 From Applicability to Action: An Operational Continuum A central feature of the architecture is the distinction and dependency between applicability and actionability, which together define an operational continuum. Applicability characterizes the context-specific validity of general knowledge. It is established through epistemic operations that relate applicability conditions to contextual information and determine whether a statement can be meaningfully used or a specific operation should be executed in a given situation. Actionability, in contrast, characterizes the structural capacity of knowledge to support operations. It defines the space of possible operations based on the explicit specification and is encoded in action units, which specify how knowledge can, in principle, be used within processes, including the inputs, procedures, conditions, and objectives involved. The relationship between the two is asymmetric, with applicability constituting an epistemic precondition for interventional actionability. Only when knowledge is assessed as applicable can it serve as a reliable basis for

34 selecting and executing procedures. This distinction resolves a common ambiguity in existing approaches, which often conflate knowing that knowledge applies, and knowing how to act on it. Rather than forming a binary distinction, this relationship defines a progression from structural potential to context-sensitive execution. The framework establishes a direct correspondence between the levels of this progression and their formal representation within the Semantic Units Framework (Table 1). This mapping enables abstract notions such as actionability, applicability, and validation to be expressed as explicit, computable structures, and clarifies how knowledge transitions from abstract representation to situated use. Table 1: Mapping between epistemic concepts and their formal representations Concept

Representation in the Framework

Structurally actionable knowledge

Action Unit

Contextually applicable knowledge

Applicable Action Unit

Empirically supported applicability

Validated Applicable Action Unit

Interventional execution

Procedure within Action Unit

This correspondence highlights that actionable knowledge is not an implicit property of data or models, but is explicitly represented through structured combinations of semantic units. In particular, it shows how the progression from structural actionability to contextual validity and empirical support can be formalized within a unified representation framework.

6.3 Dual Perspective on Actionable Knowledge The proposed framework supports two complementary and symmetrical perspectives on actionable knowledge, reflecting the bidirectional relationship between knowledge and context. These perspectives correspond, at the architectural level, to the same fundamental asymmetry identified at the statement level in Section 2.2 and at the operational level in Section 5.1, i.e., between representing the world and intervening upon it, in Hacking's sense (32), and between the word-to-world and world-to-word directions of fit in Searle's sense (20,21). In the forward (decision-oriented) perspective, a specific situation and objective are given, and the task is to identify relevant knowledge and appropriate procedures. This corresponds to real-world decision-making scenarios, where agents must determine how to act under context-specific constraints. It instantiates the world-to-word direction: the world is given, and the task is to find the directive knowledge and procedures that specify how to change it to match a desired state. In the reverse (knowledge-oriented) perspective, general knowledge and objectives are given, and the task is to identify the contexts in which this knowledge is applicable and operationalizable. This perspective enables systematic validation, comparison across contexts, and the analysis of knowledge transferability. It instantiates the word-to-world direction: knowledge is given, and the task is to determine the real-world situations it correctly describes or to which it validly applies. These perspectives are structurally linked through action units and applicability conditions, which make the conditions of transition between them explicit and evaluable. Together, they reveal the fundamentally bidirectional nature of actionable knowledge: knowledge informs action, while

35 contexts constrain and organize the use of knowledge, i.e., a duality that runs from the philosophy of language through the philosophy of science to the design of knowledge infrastructures. The architectural implications of this framework for knowledge infrastructure design are discussed in Chapter 8.

7. Application to Biodiversity Science: Reinterpreting Failures and Constructing Action Units 7.1 The Knowledge-Action Gap in Practice Biodiversity science provides a particularly instructive domain for examining the knowledge–action gap. Despite decades of intensive research, monitoring programmes, and conservation efforts, the translation of ecological knowledge into effective management action remains persistently difficult. The problem is not a shortage of knowledge. It is the absence of representational structures that make knowledge operationally usable in specific contexts. Two recurring patterns illustrate this gap with particular clarity. Both are applicability failures and thus cases where general knowledge was applied without evaluating whether the conditions for its valid use were satisfied, but they differ in the type of applicability conditions involved and the operation class at which the failure occurs. The first concerns contextual applicability conditions in intervention operations. A restoration procedure was executed without verifying that the site-specific conditions required for its success were present. The second concerns referential and formal applicability conditions in epistemic operations. A predictive model was transferred to a new context without verifying that the conditions required for its predictions to be reliable were satisfied. Together, they show that the applicability problem is not confined to a single operation class but runs across the operational architecture, making the explicit representation of typed applicability conditions a general requirement for actionable knowledge in biodiversity science.

7.1.1 Intervention Failure: Mangrove Restoration A frequently cited example of intervention failure is large-scale mangrove restoration (36). Numerous restoration projects across South and Southeast Asia, East Africa, and Latin America have relied on the general causal principle that planting mangrove seedlings leads to ecosystem recovery and coastal protection. This principle is well-supported by ecological research and constitutes a legitimate universal or prototypical statement within the paper's framework, i.e., a generalizable causal relationship between planting interventions and ecosystem outcomes. However, many of these projects failed. Reported causes include inappropriate hydrological conditions, unsuitable sediment regimes, excessive wave energy, salinity levels outside species tolerances, and inadequate consideration of land-use history (36,37). In several cases, seedlings were planted in locations that mangroves could not naturally colonize under any conditions, resulting in high mortality rates and no measurable recovery. The general causal knowledge was structurally actionable, since it specified a procedure (planting) linked to an ecological goal (ecosystem recovery),

36 but it was applied as if contextually valid without systematic evaluation of whether the conditions under which it holds were satisfied at the target site. From the perspective developed in this paper, these failures share a common diagnostic structure. The intervention action unit was structurally present but applicability-incomplete. The causal hypothesis about mangrove recovery was represented as a universal or prototypical statement. A restoration procedure was available as a directive statement. An ecological goal was specified. But three critical components were missing or underspecified: (i) the contextual applicability conditions defining the environmental constraints under which the restoration procedure can be validly applied (hydrological connectivity, sediment accretion rates, wave exposure thresholds, salinity range), (ii) the contextual information representing the actual site-specific conditions at the target location, and (iii) the epistemic action units required to assess whether those site conditions satisfy the applicability conditions before the intervention is executed. The result was an applicable action unit that was never actually evaluated for applicability—structurally complete but epistemically unjustified.

7.1.2 Epistemic Failure: Species Distribution Model Transferability A second recurring failure pattern concerns the transfer of species distribution models (SDMs) across spatial and temporal contexts. SDMs are among the most widely used tools in biodiversity science, generating predictions of species occurrence or abundance from environmental predictor variables. These predictions are routinely used to inform conservation planning, protected area design, and climate change impact assessments. The causal and statistical relationships encoded in SDMs constitute universal or prototypical statements, i.e., generalizations about how environmental conditions determine species distributions. However, the transferability of SDMs to new spatial contexts, time periods, or environmental regimes is consistently identified as a major challenge (38,39). Models calibrated in one region frequently produce unreliable predictions when applied to novel regions with different species assemblages, altered environmental conditions, or conditions that lie outside the range of the training data. This is not primarily a technical modelling problem, it is an applicability problem. The epistemic operation of applying an SDM to a new context requires that the referential and formal applicability conditions of the model be satisfied in the target context: the predictor variables must remain ecologically meaningful, the species' realized niche must be adequately sampled in the training data, the functional relationships between predictors and species occurrence must be stationary across the transfer extent, and the target context must not lie in environmental space beyond the model's training domain. These conditions are frequently violated in practice. Applications of SDMs in novel geographic or climatic contexts routinely involve extrapolation beyond the training data's environmental range, which is a well-documented source of prediction failure (38). The ecological relationships encoded in the model may be non-stationary across space or time, reflecting locally contingent factors not captured by the predictor variables. Species range-filling may be incomplete, so that absences in the training data do not reliably reflect unsuitable conditions in the target context. From the framework's perspective, these failures represent incomplete epistemic action units. The SDM constitutes a universal or prototypical statement unit encoding a causal and statistical relationship between environmental predictors and species occurrence. Applying it to a new context is an epistemic operation, specifically, a recognition operation that takes contextual information

37 (environmental conditions at the target location) and produces a species distribution estimate (an assertional statement describing the predicted occurrence). This operation requires referential applicability conditions specifying under what environmental and methodological constraints the model's predictions can be validly interpreted. These conditions are rarely made explicit as first-class representational components in current practice. Instead, they are embedded implicitly in modelling choices or reported as caveats in methods sections, rather than represented as evaluable constraints that can be systematically assessed before the model is applied.

7.2 Reinterpreting Failures Through the Framework The two failure patterns described in Section 7.1 are not isolated implementation problems. They are instances of a generalizable structural deficit that the action units framework can precisely characterize and address. Three recurring deficiencies can be identified across both cases: 1.​ Missing or underspecified contextual information​ In the mangrove restoration case, site-specific hydrological, sediment, and salinity conditions were not represented as explicit assertional statement units linked to the intervention action unit. In the SDM case, the environmental conditions at the target location were not systematically compared against the model's training domain. Without structured representations of contextual information that can be queried against applicability conditions, the epistemic step of applicability assessment cannot be performed. 2.​ Implicit or absent applicability conditions​ In both cases, the conditions under which the general knowledge holds were not represented as explicit, evaluable applicability condition units. The causal hypothesis about mangrove recovery implicitly presupposes specific hydrological and sediment conditions, but these were not formalized. The SDM implicitly presupposes environmental stationarity and adequate niche sampling, but these were not represented as formal referential applicability conditions. Without explicit applicability conditions, the question "Does this knowledge apply here?" cannot be systematically asked, let alone answered. 3.​ Incomplete operational grounding​ Both failures reflect a pattern in which action units were used as if they were validated applicable action units and thus as if their validity had been established, when they were at best structurally applicable and at worst not applicable at all. The three-level hierarchy introduced in Section 5.2.5 captures this precisely. The transition from action unit (structural potential) to applicable action unit (contextually validated) requires explicit applicability assessment, and the transition to validated applicable action unit (empirically supported) requires documented evidence from prior applications in comparable contexts. In both cases, the transition was assumed rather than performed. These deficiencies are not idiosyncratic to mangrove restoration or SDM transfer. They reflect a general pattern in which the structural and epistemic preconditions for interventional actions are left implicit in current knowledge infrastructures. The action units framework provides the representational vocabulary to make these preconditions explicit, modular, and evaluable, thereby converting failure patterns from post-hoc diagnostic observations into pre-operational checks that can be performed before action is taken.

38

7.3 Constructing Action Units for Biodiversity Tasks The following three sections develop worked examples of action unit construction for biodiversity tasks, one for each of the three primary operation classes. Each example illustrates how the framework's representational components, i.e., statement types, applicability conditions, contextual information, plan specifications, and objectives, can be assembled into explicit, structured action units that support both human and machine agents in performing the relevant operations reliably.

7.3.1 Epistemic Action Unit: Species Identification in Biomonitoring Biomonitoring programmes generate biodiversity data by identifying and characterising organisms encountered at survey sites. The fundamental epistemic operation underlying biomonitoring is species identification, i.e., given an observed organism or specimen, assign it to the appropriate taxon. This is a recognition operation in the framework's sense, as it takes a material entity (the organism or specimen) as input and produces an assertional statement (a species occurrence record) as output, mediated by diagnostic knowledge. The epistemic action unit for species identification in a standardised biomonitoring context can be constructed as follows. The input unit comprises a physical specimen or field observation and thus a material entity, together with associated observational metadata such as geographic coordinates, habitat description, date, time, and collection method, represented as a compound unit comprising corresponding assertional statement units describing the observation context. The plan specification unit comprises the diagnostic knowledge required to perform the identification, including taxonomic keys, morphological descriptions, molecular identification protocols (e.g., DNA barcoding), photographic guides, or trait-based classification criteria. These are directive statement units specifying how the identification procedure is to be carried out. For machine agents, this component may be implemented as a trained classifier or a SPARQL query over a reference ontology. The referential applicability conditions unit specifies the constraints under which the identification procedure can be validly applied and its output reliably interpreted. These include (i) the specimen or observation must be sufficiently complete and undamaged to allow diagnostic features to be assessed, (ii) the taxonomic reference database used must include the taxon in question and cover the relevant geographic region, (iii) the identification method must be appropriate for the life stage and preservation state of the specimen, and (iv) the observer must possess the referential lexical competence required to apply the diagnostic criteria, i.e., the diagnostic knowledge in the plan specification unit must be operationalised through a combination of training, experience, and method-specific expertise. For machine agents, formal applicability conditions additionally include schema compatibility between the input representation and the classifier's training domain. The objective unit specifies the epistemic goal which is the assignment of the observed specimen or occurrence to a taxon at the appropriate resolution (species, genus, or higher), with an associated confidence level where applicable, producing a validated assertional statement unit that can serve as contextual evidence in downstream action units. The output unit is the species occurrence record and thus an assertional statement unit of the form "taxon X was observed at location Y under conditions Z", which serves as both the product of

39 the epistemic action unit and as contextual information for transformational and intervention action units that operate on aggregated biodiversity data. This construction makes explicit what is typically left implicit in biomonitoring protocols. The validity of an occurrence record depends not only on the observer's identification decision but on the satisfaction of the referential applicability conditions, including adequate specimen quality, appropriate method, relevant reference materials, and sufficient expert competence. Making these conditions explicit as evaluable representational components enables systematic quality assessment of monitoring data and supports the detection of cases where identification reliability may be compromised.

7.3.2 Transformational Action Unit: Deriving Essential Biodiversity Variables Biomonitoring data, including species occurrence records and abundance counts produced by epistemic action units such as the one described in Section 7.3.1, are rarely directly usable for policy-relevant biodiversity assessment. They must first be transformed into standardised, spatially and temporally coherent indicators that can be compared across sites, time periods, and regions. The concept of Essential Biodiversity Variables (EBVs) provides a widely adopted framework for defining these intermediate products (40). EBVs are state variables that stand between raw primary observations and high-level policy indicators, capturing critical aspects of biodiversity change across genetic, species, and ecosystem levels. Deriving an EBV from raw monitoring data is a transformational operation that takes information content entities (occurrence records, abundance counts, trait measurements) as input and produces a derived information content entity (a spatially and temporally explicit biodiversity variable) as output, through formally specified computational and statistical procedures. The corresponding transformational action unit can be constructed as follows. The input unit comprises a set of species occurrence records and associated environmental covariate data as a compound unit consisting of assertional statement units describing specific observations, conforming to a defined data schema (e.g., Darwin Core for occurrence data, standardised covariate grids for environmental data). The plan specification unit comprises the computational methods used to derive the EBV, including species distribution modelling algorithms, occupancy models, abundance estimators, diversity indices, or remote sensing analysis pipelines. These constitute directive statement units at the algorithmic level, specifying the transformation procedure to be applied to the input representations. The formal applicability conditions unit specifies the structural constraints under which the transformation can be validly applied, including that (i) the input occurrence data must conform to the required data schema, including mandatory fields such as species identification, geographic coordinates, date, and sampling effort metadata, (ii) the environmental covariate data must cover the spatial and temporal extent of the occurrence records, (iii) the occurrence data must satisfy minimum sample size requirements for the modelling algorithm, and that (iv) the geographic scope of the analysis must not exceed the spatial domain over which the model's predictors remain ecologically meaningful, that is, the formal applicability conditions include an implicit constraint against extrapolation into novel environmental space, which directly addresses the SDM transferability failure described in Section 7.1.2. For machine agents, these conditions can be

40 operationalised as SHACL validation rules or SPARQL ASK queries applied to the input data before the transformation pipeline is executed. The objective unit specifies the transformational goal that the production of a spatially explicit EBV product, such as species occupancy estimates per grid cell per time period, or functional diversity indices per habitat type, satisfies specified accuracy, resolution, and provenance requirements and can serve as input to higher-level biodiversity indicators or to intervention action units for conservation planning. The output unit is the derived EBV product and thus an information content entity representing aggregated biodiversity state at specified spatial and temporal resolutions, linked through provenance assertions to the input data and processing methods from which it was derived. This construction makes explicit the distinction between the technical execution of a transformation (covered by the plan specification unit and formal applicability conditions) and the epistemic validity of the resulting product (which requires the referential applicability conditions of the upstream species identification action unit to have been satisfied). A transformational action unit alone cannot ensure that its outputs are epistemically valid representations of biodiversity state, as this depends on the reliability of the epistemic action units that produced its inputs. This dependency is precisely what the three-level hierarchy in Section 5.2.5 captures, i.e., a validated applicable action unit for EBV derivation requires not only formal applicability of the transformation but empirical validation of the underlying epistemic operations.

7.3.3 Intervention Action Unit: Ecosystem Restoration Planning Conservation and restoration practitioners must translate the biodiversity knowledge captured in EBVs and species distribution models into concrete site-level interventions. Ecosystem restoration is the paradigmatic intervention operation in biodiversity science, taking a degraded or modified ecosystem as input, i.e., a material entity in a given state, and aiming to produce a restored ecosystem as output, through a goal-directed process guided by causal and procedural knowledge. The mangrove restoration failure described in Section 7.1.1 represents an incomplete instance of this action unit type. Here we construct the full intervention action unit that the restoration intervention required but lacked. The input unit comprises the target site, which is a material entity representing a degraded coastal or riparian ecosystem, together with contextual information units describing its current state in the form of assertional statement units encoding hydrological connectivity measurements, sediment accretion rates, wave exposure indices, salinity profiles, land-use history, existing vegetation composition, and socio-economic context (tenure arrangements, community presence, resource use). This contextual information constitutes the situational grounding that distinguishes an applicable action unit from a merely structural one. The plan specification unit comprises the procedural knowledge specifying how the restoration intervention is to be carried out, including site assessment protocols, species selection criteria, planting or propagule dispersal procedures, hydrological restoration techniques, and monitoring and adaptive management protocols. These are directive and conditional directive statement units that specify the restoration procedure in actionable, step-by-step form. They may also include conditional directives of the form "IF hydrological connectivity score exceeds threshold T, THEN proceed with passive natural regeneration; ELSE initiate active replanting." This conditional structure corresponds precisely to the conditional action unit formalized in Section 8.5.

41 The contextual applicability conditions unit specifies the environmental and methodological constraints under which the restoration procedure can be validly applied and is expected to achieve the specified goal. For mangrove restoration, these include that (i) tidal inundation frequency must fall within the species-specific tolerance range (typically between approximately 20–75% of the time, depending on species), (ii) sediment accretion rate must be positive or neutral, ensuring that substrate is not eroding faster than plants can establish, (iii) salinity must fall within species tolerance limits, (iv) wave energy must be below the threshold at which seedling survival is critically compromised, and that (v) the site must not be subject to ongoing physical disturbance (e.g., boat traffic, grazing) that prevents establishment. These conditions correspond to the referential and contextual applicability conditions that the restoration failures in Section 7.1.1 neglected to evaluate. The objective unit specifies the intervention goal of establishing a self-sustaining mangrove ecosystem of defined composition, density, and extent within a specified time horizon, measured against explicit ecological indicators (canopy cover, species composition, seedling recruitment, faunal colonisation). The output unit is the restored ecosystem and thus the material entity in its new state, together with monitoring records documenting the degree to which the restoration objective has been achieved, which feed back into subsequent epistemic action units for ongoing assessment. The critical feature of this construction, compared to the incomplete action unit that characterised the failing restoration projects, is the explicit representation of the contextual applicability conditions unit and its evaluation against the contextual information unit before the intervention is executed. In the framework's terms, this evaluation is itself an epistemic action unit that takes the contextual information and applicability conditions as inputs and produces an applicability judgment (a validated assertional statement of the form "the applicability conditions for mangrove restoration are [satisfied / not satisfied / partially satisfied] at this site"). This judgment is the necessary precondition for the intervention action unit to be elevated from an action unit to an applicable action unit, and ultimately, with documented evidence from prior applications, to a validated applicable action unit.

7.4 Multi-Stakeholder Dimensions of Actionable Knowledge Biodiversity knowledge does not serve a single type of agent. Scientists generate and validate general knowledge, including causal hypotheses, statistical models, and monitoring protocols. Policymakers select and prioritise interventions under conditions of uncertainty and competing objectives. Practitioners implement procedures under real-world operational and resource constraints. Citizens and local communities contribute contextual information through participation in monitoring and provide crucial local ecological knowledge that may not be captured in formal databases. Each of these agents interacts with actionable knowledge in a distinct mode, but, and this is one of the framework's central architectural claims, all of them interact with the same underlying action unit structures, accessing and contributing to different components. This multi-stakeholder character maps directly onto the dual perspective introduced in Section 6.3. Scientists primarily work in the reverse (knowledge-oriented) perspective. Given general knowledge (causal hypotheses, SDMs, restoration protocols), they identify the contexts in which it is applicable, refine applicability conditions, and accumulate the empirical evidence required to elevate action units to validated applicable action units. Practitioners and policymakers primarily work in the forward (decision-oriented) perspective. Given a specific situation and management objective, they

42 query available action units to identify applicable procedures and select among them based on contextual validity and empirical support. Citizens contribute primarily to the epistemic layer, as their observations generate assertional statement units that serve as contextual information for applicability assessment and as empirical evidence for validation. Scientists are the primary producers of the general knowledge components of action units, i.e., universal and prototypical statement units encoding causal and statistical relationships, diagnostic plan specification units encoding identification and measurement protocols, and applicability condition units encoding the constraints under which these are valid. Their central epistemic task within the framework is applicability condition specification, by determining and documenting the conditions under which a causal hypothesis, a species distribution model, or a restoration protocol can be reliably applied. This work is currently distributed across methods sections, supplementary materials, and grey literature in ways that make it difficult to retrieve, compare, or machine-process. Action units provide the representational structure for consolidating it. Policymakers interact primarily with the objective units and validated applicable action units. Their central task is goal specification and thus defining what ecological state is desired, at what spatial and temporal scale, under what resource and social constraints, and procedure selection from among applicable action units that have been evaluated for contextual validity. The action units framework supports this by making the applicability assessment explicit. Rather than choosing between restoration approaches based on general reputation or prior experience, a policymaker working within an action unit-structured knowledge infrastructure can identify which approaches have documented evidence of applicability in contexts comparable to the target site. Practitioners interact primarily with the plan specification units and contextual applicability conditions. Their central operational task is the execution of procedures under real-world constraints, including adaptation to conditions that differ from those assumed in the formal protocol. The conditional directive structure of action units supports this by making the decision logic explicit. Practitioners can assess whether conditions have changed in ways that require switching to an alternative procedure or pausing for additional epistemic assessment. Citizens and local community members contribute primarily through the generation of assertional statement units in the form of species observations, habitat descriptions, phenological records, traditional ecological knowledge, thereby enriching the contextual information layer available for applicability assessment. Citizen science platforms such as iNaturalist generate millions of occurrence records that constitute epistemic action unit outputs of the species identification type described in Section 7.3.1. Integrating these records into a structured action unit framework requires that their referential applicability conditions, i.e., observer competence, identification method, data quality thresholds, be explicitly represented so that downstream transformational and intervention action units can assess the reliability of the contextual information they provide. Together, these four modes of engagement show that action units are not merely technical representational structures but organisational interfaces that make the different contributions of different knowledge actors explicit, modular, and composable. The framework does not privilege scientific knowledge over local knowledge or technical expertise over practical experience. Instead, it provides a shared representational architecture within which different forms of knowledge can be integrated, assessed for applicability, and used to support context-sensitive action.

43

7.5 From Individual Action Units to Composite and Conditional Structures The three action units constructed in Sections 7.3.1–7.3.3 are not independent. They form a natural compositional structure that illustrates both the granularity concept introduced in Section 5.1.4 and the conditional action unit formalized in Section 8.5. The species identification action unit (Section 7.3.1) is an epistemic action unit whose output, i.e., a species occurrence record, is the primary input to the EBV derivation action unit (Section 7.3.2). The EBV derivation action unit is a transformational action unit whose output, i.e., spatially explicit biodiversity state estimates, provides the contextual information layer for the ecosystem restoration action unit (Section 7.3.3). The three action units are therefore constitutively ordered. The restoration intervention action unit depends on the EBV derivation transformational action unit, which depends on the species identification epistemic action unit. Together, they form a composite action unit of the type described in Section 5.2.1, a higher-order action unit whose coarser-grained goal (achieving a defined ecological state through restoration) decomposes into an ordered sequence of lower-order action units of different types. This composite structure also illustrates what can be described as an ecological fingerprint or biodiversity profile of an ecosystem, which is a structured, multidimensional characterisation of its state derived from integrated biomonitoring data, emerging as the operational product of the linked epistemic and transformational processes: species identification generates occurrence records, transformational operations aggregate and normalise these into EBVs and derived indicators, and the resulting composite ecological profile characterises the ecosystem’s state at a specific spatial location and time. The ecological fingerprint integrates measurements and metrics across multiple classification dimensions, including biome type, climate class, soil type, landscape form, species composition, functional and ecological trait composition, and ecosystem functions and services, providing a structured multidimensional representation of ecosystem identity and condition. Crucially, because biomonitoring data is collected repeatedly over time, ecological fingerprints can be constructed as temporal sequences that track how ecosystem state changes across monitoring cycles, supporting both retrospective assessment of ecosystem trajectories and prospective evaluation of restoration outcomes. Consequently, rather than a static classification scheme, the ecological fingerprint is a continuously updatable, evidence-grounded representation that supports both backward-looking assessment (how has this ecosystem changed?) and forward-looking intervention planning (under what conditions is this ecosystem suitable for which restoration procedures?). The conditional dimension becomes explicit when the composite action unit is extended with a conditional action unit structure. Consider the following conditional logic that governs the transition from ecological assessment to restoration action: IF the EBV-based ecological fingerprint indicates that hydrological connectivity at a candidate mangrove restoration site satisfies the contextual applicability conditions for natural regeneration, THEN trigger a passive natural regeneration action unit; ELSE IF connectivity is restorable through engineering intervention, THEN trigger a hydrological restoration action unit as a precondition for planting; ELSE flag the site as currently non-applicable and schedule a reassessment. This IF–THEN structure is a conditional action unit of the form described in Section 8.5. The IF clause is implemented as an executable query evaluating whether the contextual information (EBV-derived site profile) satisfies the applicability conditions, and the

44 THEN clause triggers the appropriate intervention procedure. The knowledge graph containing the composite action unit can evaluate this conditional logic automatically, identifying applicable procedures for each site in a landscape-scale restoration planning exercise without requiring manual applicability assessment for each site individually. This example illustrates the transition toward post-FAIR knowledge infrastructures that Chapter 8 develops in detail. The knowledge graph is no longer merely a repository of ecological facts and restoration procedures but an active decision-support system that evaluates contextual applicability, identifies applicable action units, and orchestrates the operational workflow from monitoring through assessment to intervention planning. More broadly, these examples reframe biomonitoring not merely as a data collection practice but as a structured epistemic and operational process within an actionable knowledge infrastructure. Species identification, EBV derivation, and restoration planning are not three separate activities. Instead, they are constitutively ordered action units that jointly instantiate the three operation classes across the knowledge–action cycle. In this sense, the ecological fingerprint is not only a product of monitoring but an operational interface between knowledge and action, with a continuously updatable, evidence-grounded representation of ecosystem state that makes the conditions for context-sensitive intervention both explicit and evaluable. This is precisely what post-FAIR knowledge infrastructures, as developed in Chapter 8, are designed to support.

8. Implications for Knowledge Infrastructures 8.1 From Knowledge Infrastructures

Representation

to

Actionable

In the previous chapters, we provided a conceptual framework for actionable knowledge. Here, we want to address the question how this framework can be operationalized by translating it into infrastructure design. We argue that the framework developed in this work has direct implications for the design of next-generation knowledge infrastructures. Existing systems are primarily optimized for the storage, integration, and retrieval of data and conceptual knowledge. While these capabilities are essential, they remain insufficient for supporting the use of knowledge in context. Actionable knowledge infrastructures extend this paradigm by explicitly supporting the transition from representation to operation. Rather than treating knowledge as a static resource, they represent how knowledge participates in processes that relate representations to real-world situations, transform them into usable forms, and apply them to achieve desired outcomes. This shift requires that knowledge infrastructures make operational structure explicit. In particular, action units must be represented as first-class objects within the knowledge graph. By integrating semantic content with input–output specifications, plan specifications, applicability conditions, and objectives, action units encode how knowledge can, in principle, be used within epistemic, transformational, and intervention operations. A central implication of this approach is that structural actionability refers to whether knowledge is represented in a form that enables operations by explicitly specifying their inputs, outputs, procedures, and objectives, while contextual applicability is an epistemic property that concerns whether a given operation or action unit is valid in a specific situation, based on contextual

45 information and applicability conditions. For knowledge infrastructures, this distinction implies that supporting action requires more than enabling execution. Systems must also support the evaluation of contextual validity. This introduces an explicit epistemic layer in which applicability conditions are evaluated against contextual information, often under conditions of uncertainty. Within this architecture, different representational components fulfil distinct roles: ●​ Universal, prototypical, directive, and conditional directive statement units provide the causal, statistical, and procedural knowledge basis for reasoning and action. ●​ Assertional statement units provide contextual information describing specific situations. ●​ Applicability condition units define the constraints under which knowledge and procedures can be validly applied. ●​ Plan specification units represent methods, algorithms, and practices that can be executed within processes. ●​ Objective units define desired outcomes. ●​ Action units integrate these components into operational structures that enable the coordination of epistemic, transformational, and intervention operations as well as composite operations thereof. Taken together, these elements enable knowledge infrastructures to support not only what can be done, but also what should be done in a given context. From an operational perspective, this integration can be understood as a structured workflow that connects the three classes of operations (recognition → transformation → decision → intervention; discussed in Section 8.4). More generally, actionable knowledge infrastructures must integrate three complementary capabilities. First, operational linking that connects semantic units to operations and thereby establishes structural actionability. Second, procedural transparency that represents methods and algorithms as explicit, reusable plan specifications. And third, epistemic evaluation that enables the assessment of applicability through the explicit representation and evaluation of contextual information and applicability conditions. Only the integration of these capabilities enables actionable understanding. It transforms knowledge infrastructures from systems that support data access and computation into systems that support context-sensitive, evidence-informed, and goal-directed action.

8.2 Linking Actionability

Semantic

Units

to

Operations:

Structural

A central requirement for actionable knowledge infrastructures is the explicit linking of knowledge representations to operations. This establishes an intermediate operational layer between semantic representation and procedural execution, connecting what knowledge is with what can be done with it. This layer can be realized by associating semantic unit types with formal schema definitions and defining operations in terms of their required inputs, outputs, and associated plan specifications. In such a system, semantic units are typed by schemata, while operations specify how entities of given types can participate in processes through input–output roles, procedural definitions, and applicability constraints.

46 Several prior frameworks have addressed related aspects of linking knowledge representations to processes and workflows. Process ontologies such as the Ontology for Biomedical Investigations (OBI) (35) provide controlled vocabularies for representing biomedical investigations, including planned processes, their participants, and their outputs, grounded in upper-level ontologies such as the Basic Formal Ontology (BFO) (9). The W3C PROV-O provenance ontology (41) and its extension P-Plan (42) represent the provenance of scientific processes and link plan specifications to their execution records. Workflow bundling frameworks such as RO-Crate (43) and the Common Workflow Language (CWL) (44) provide practical mechanisms for packaging research artefacts with their metadata and for specifying portable, reusable computational workflows across heterogeneous execution environments. Together, these approaches constitute important infrastructure for process documentation, provenance tracking, and computational reproducibility. Like plan specifications in OBI and related frameworks, action units specify participants, procedural knowledge, and objectives. They differ, however, in a fundamental and consequential respect. OBI, PROV-O, P-Plan, RO-Crate, and CWL are all primarily concerned with the documentation and execution of processes that have been or will be carried out. They represent what happened, what was planned, or how a workflow should be executed. They do not make contextual applicability conditions explicit as first-class, typed, and evaluable components, nor do they distinguish between the structural capacity of a representation to support an operation (actionability) and the epistemic validity of performing that operation in a specific situation (applicability). Action units address precisely this prior layer. They specify not only how an operation can be performed, but also the conditions under which it should be performed, and what evidence would establish that those conditions are satisfied. In this sense, action units provide pre-operational infrastructure that complements workflow and provenance frameworks by making the representational and epistemic preconditions of process execution explicit and evaluable. Where CWL specifies how a tool should be invoked given conformant inputs, an action unit specifies when the invocation is contextually valid and what contextual information is required to establish this. Where PROV-O and OBI record what an activity used and generated, an action unit specifies the applicability conditions that must hold for the activity to be epistemically justified. Importantly, operations in this context are not limited to data transformations. They comprise all three operation classes. Transformational operations are typically fully formalizable as computational processes, whereas epistemic and intervention operations may involve interpretation, observation, judgment, or physical interaction with the world. The typed operational layer therefore represents the structural components of operations independently of whether their execution is fully machine-, human-, or hybrid-mediated. Within this generalized operational layer, different operation classes are governed by different compatibility relations. Epistemic operations depend on referential compatibility, determined by diagnostic criteria linking representations to material entities or system states. Transformational operations depend on formal compatibility, defined by schema constraints on input and output representations. Intervention operations depend on contextual compatibility, determined by applicability conditions evaluated against situational context. This explicit typing exposes the operational affordances of knowledge representations. It becomes possible to determine which operations are structurally applicable to which entities, and under which conditions they can be executed. Conceptually, it can be understood as an operations meta-graph in which workflows emerge as compositions of compatible operations across this meta-graph, constituting an operational layer over the knowledge graph, in which semantic units

47 define available inputs and outputs, and operations define the transformations between them. Methods, algorithms, and practices are represented as operations with defined input and output constraints, allowing both human users and computational agents to identify compatible procedures and assemble workflows based on structural compatibility. In this role, action units and their associated plan specifications function as interfaces between semantic representations and execution environments, linking knowledge structures to executable operations in external systems. This approach aligns with emerging FAIR service ecosystems (7,12), in which terminology services manage controlled vocabularies and ontologies, schema services define semantic schemata with formal constraints (e.g., via SHACL shapes), operation services expose executable functions over these schemata, and workflow services enable composition of operations into pipelines. Within this ecosystem, transformational operations can become fully machine-actionable. If operations are defined with explicit input and output schemata, tested, and registered as reliable services, then AI agents can, given a goal and input data, construct deterministic or semi-deterministic pipelines over available operations to reach a desired transformation outcome. This enables a form of trusted, reproducible, and traceable agentic execution, particularly for data processing and representation transformation tasks. In this sense, transformational action units can be executed directly in distributed computational environments, provided schema compatibility is satisfied. By providing formal schemata for all semantic units, the components of action units become explicitly typed. This enables structural operation discovery. For a given datum or semantic unit, it becomes possible to query which operations are compatible with its schema-defined type. In a knowledge graph interface, this can be exposed as an operational affordance layer showing which transformations, recognition procedures, or interventions are structurally available for a given type of entity. In parallel, the framework supports a complementary mechanism for epistemic support and applicability assessment. Operations can be linked to empirical evidence units documenting prior successful applications in comparable contexts. This allows systems to surface not only what can be done structurally, but also what has been shown to work under similar conditions, thereby supporting the evaluation of contextual applicability. This dual exposure establishes a tight integration between the structural layer (schema-based compatibility and executable operations), and the epistemic layer (contextual evidence and applicability support). Not all operations are fully machine-actionable. Epistemic operations often involve observation, interpretation, and judgment, while intervention operations frequently involve physical actions on material systems. Such operations may depend on human agents or external systems and cannot, in general, be fully captured as executable services. However, even when not executable, they can be represented structurally through action units specifying inputs, outputs, procedures, and applicability conditions. The framework therefore extends beyond machine-actionability by uniformly representing human-, machine-, and hybrid-actionable operations within a single typed structure (see also Fig. 5).

8.3 Epistemic Actionability and Formalized Statement Schemata Actionable knowledge infrastructures must support not only intervention actions but also epistemic actions, such as recognition and description, which establish the relationship between knowledge representations and real-world situations. These operations constitute the epistemic foundation of

48 actionable understanding, as they determine what is the case and whether knowledge can be meaningfully applied. At the core of epistemic actionability lies the ability to systematically relate representations to observations. At the level of terms, this is achieved through diagnostic criteria that enable recognition and designation of entities. At the level of statements, however, additional structure is required to support consistent interpretation and reuse across contexts. This structure is provided by formalized statement schemata, which define reusable templates for specific types of statements. Such schemata specify (i) the structural organization of statements (e.g., predicate-argument structures such as RDF’s triple syntax of Subject–Predicate–Object), (ii) the semantic roles associated with each syntactic position, and (iii) constraints on admissible values for each position. By making these elements explicit, schemata enable consistent representation, interpretation, and validation of statements across datasets and domains. They thereby provide the foundation for statement-level epistemic operations, including both recognition (representation → world) and description (world → representation) (12). A key aspect of epistemic actionability is the alignment between human and machine interpretability. For human agents, epistemic actionability is grounded in the interpretation of predicate–argument structures and their semantic roles, enabling meaningful understanding of statements in context. For machine agents, epistemic actionability is realized through schema-constrained matching, querying, and validation over formally specified structures (e.g., SHACL-defined shapes). In this sense, human interpretation aligns with semantic roles in predicate–argument structures, whereas machine interpretation aligns with schema-constrained pattern matching. Formalized statement schemata, as they are a constituent component within the Semantic Units Framework, bridge these perspectives by providing representations that are both cognitively interpretable and computationally processable. In this context, assertional statements play a central role. When represented using formalized schemata, they become epistemically actionable by enabling the structured recognition and description of concrete situations. They provide the contextual grounding required for evaluating applicability and for supporting subsequent transformation and intervention operations. In contrast, universal and prototypical statements are not directly epistemically actionable in isolation. Their application to specific situations requires epistemic operations that relate them to contextual information, typically through the evaluation of applicability conditions. In practice, this relation can be understood analogously to instance–type relationships: assertional statements may instantiate or exemplify general statements, but establishing this relationship requires explicit epistemic grounding. Epistemic actionability is formally captured through epistemic action units, which represent the structural conditions under which such grounding can be achieved. These units integrate representations (terms or statements), diagnostic procedural knowledge (e.g., recognition criteria or observational methods), referential applicability conditions, and epistemic objectives. They define how representations can be validly related to entities, system states, or situations, and thereby provide the operational basis for establishing context-sensitive understanding. In summary, epistemic actionability depends on the availability of structured representations that support both human interpretation and machine processing, as well as on the explicit specification of diagnostic procedures and applicability conditions. Formalized statement schemata, in combination with epistemic action units, provide the necessary foundation for systematically

49 linking knowledge to the world and for enabling the epistemic processes on which all subsequent action depends.

8.4 Objective-Driven Action Discovery A central requirement for actionable knowledge infrastructures is the explicit representation of goals as first-class semantic entities. Within the proposed framework, objectives are represented as objective units, which specify desired states of a system or intended outcomes of an operation. These units provide the orientation necessary for moving from knowledge representation to goal-directed use. Objective units are not limited to intervention contexts but are intrinsic to all operation types. Epistemic objectives guide recognition and description (e.g., establishing whether a condition holds), while transformational objectives guide the preparation and derivation of representations (e.g., producing data in a required form). In the context of action discovery, however, intervention objectives play a central role, as they define the desired changes in material systems. Objective-driven action discovery emerges from the structured integration of four core components. Objective units specify what should be achieved, procedural plan specification units define how an objective can be achieved, applicability conditions units specify under which conditions a procedure can be validly applied, and contextual information units describe the current situation through assertional statements. This integration is formalized through intervention action units, which combine these components into coherent, operational structures. Within such units, applicability conditions mediate between general procedural knowledge and context-specific decision-making by determining whether a given procedure is valid in a particular situation. A key enabler of objective-driven action discovery is the operationalization of applicability conditions. When expressed in a form that can be evaluated, e.g., as structured queries or executable constraints (e.g., in the form of executable question units (6)), they allow systems to assess whether the conditions for applying a procedure are satisfied by the available contextual information. This transforms applicability assessment into an explicit and, in part, automatable process and enables the dynamic identification of applicable goal–method combinations for a given situation. On this basis, actionable knowledge infrastructures can support context-sensitive action discovery. Given a specific situation and a set of objectives, the system can identify which procedures are applicable and therefore which actions can be validly performed. Conversely, given a procedure or action unit, the system can identify the contexts in which it is applicable and the objectives it can achieve. This bidirectional capability reflects the dual perspective on actionable knowledge introduced in Section 6.3. Importantly, applicability conditions not only enable the identification of applicable actions but also reveal gaps between current conditions and required conditions. This allows systems to support informed decision-making under uncertainty by indicating why a procedure is not currently applicable, identifying missing or uncertain contextual information, and suggesting epistemic or transformational operations needed to establish applicability (e.g., additional observations, measurements, or data transformations). In this way, objective-driven action discovery is not a simple selection mechanism but part of a broader operational process. It connects epistemic operations (establishing what is the case), transformational operations (preparing relevant representations), and intervention operations

50 (acting on systems) into a coherent workflow of Recognition → Transformation → Applicability Assessment → Intervention. This workflow does not constitute a rigid sequence, but an iterative and interdependent process in which outputs of one stage feed back into others. In particular, intervention outcomes generate new contextual information, which re-enters the epistemic layer and supports continuous refinement of knowledge and procedures. By explicitly linking objectives, procedures, applicability conditions, and contextual information within action units, the framework enables knowledge infrastructures to move beyond static representation toward dynamic, context-aware, and goal-directed support for both human and machine agents.

8.5 Executable Conditional Action Units and Graph-Native Decision Systems A central implication of the proposed framework is that knowledge graphs can be extended beyond passive representation systems to function as decision-support and workflow orchestration systems. This is achieved through the formalization of conditional action units (Fig. 5), which operationalize context-sensitive decision logic within the graph. Conditional action units correspond to structured IF–THEN constructs, in which the IF clause represents applicability conditions, and the THEN clause specifies directive actions. Within the Semantic Units Framework, these components can be represented and operationalized as follows: The IF clause is implemented as an executable question unit, typically realized as a graph query (e.g., SPARQL ASK or SELECT). It evaluates whether contextual information satisfies the applicability conditions of an action unit. The THEN clause is implemented as an executable directive unit, which may take the form of a transformation (e.g., SPARQL CONSTRUCT or UPDATE), the invocation of an external service, or the execution of a workflow involving human agents (Fig. 5). This structure establishes a direct and explicit linkage between context evaluation (epistemic operation), data retrieval and preparation (transformational operation), and action execution (intervention operation). In this way, conditional action units provide a unified mechanism for coordinating the three classes of operations within a single, graph-native structure. Importantly, conditional action units can be interpreted as operationalized conditional directive statements, in which applicability conditions are no longer merely descriptive constraints but become computable predicates, and directive statements become executable procedures when aligned with appropriate tooling and human-machine interfaces. This transforms static knowledge representations into active decision logic. A key consequence is that the knowledge graph itself can function as a graph-native decision system. Rather than merely storing knowledge, it can evaluate context-dependent conditions, identify applicable action units, trigger corresponding operations, and orchestrate workflows across internal and external computational components. This capability is particularly powerful for linking causal hypotheses to empirical evidence. For example, conditional action units can be defined such that IF relevant measurements for the variables of a causal hypothesis are available in a given context, THEN construct an evidence unit and link it to the hypothesis. Such patterns enable the systematic integration of observational data with

51 hypothesis representations. They support automated evidence generation and linkage, detection of missing data, and formulation of counterfactual scenarios for testing context dependence. More generally, conditional action units enable the representation of closed-loop workflows within the knowledge graph, in which epistemic evaluation, data transformation, and intervention are tightly coupled. This extends the role of knowledge graphs from static repositories to active, context-aware systems capable of supporting both reasoning and action. In this sense, the framework outlines a transition toward post-FAIR knowledge infrastructures, in which data and knowledge are not only findable and interoperable, but also operationally integrated, executable, and context-sensitive.

Figure 5: Conditional action units in knowledge graphs. The IF-clause is represented as an executable question unit that queries for contextual conditions (e.g., soil moisture below threshold X) within the entire knowledge graph or a specific subgraph, while the THEN-clause is represented as a directive unit that triggers an associated action or workflow (e.g., irrigation). Together, they link context evaluation, data processing, and action execution within a unified structure, enabling the knowledge graph to function as a decision-support and workflow orchestration system.

8.6 Bounded Action Spaces and Incremental Design A key design implication of the proposed framework is that actionable knowledge infrastructures should not aim for exhaustive coverage of all possible actions. Actionability is inherently bounded and graded, reflecting that both human and machine agents operate under cognitive and computational constraints, and real-world systems introduce additional complexity and uncertainty. Representing all possible actions, contexts, and procedures within a domain would require unrealistic levels of formalization and maintenance. For infrastructure design, this implies a shift from completeness to prioritization. Systems should be organized around bounded sets of high-value action units, defined by frequency of use, relevance for decision-making, and availability of sufficiently robust knowledge, data, and procedures. This focus enables tractable implementation in a scalable manner while ensuring practical utility and measurable impact, enabling focused curation of knowledge and procedures. This principle also guides the allocation of resources, ensuring that knowledge infrastructures are both usable and sustainable. At the same time, actionable knowledge infrastructures should support incremental expansion. Action units provide a modular structure that allows new capabilities to be added progressively, as new data become available, applicability conditions are refined, and additional procedures are formalized and operationalized.

52 This results in a layered and evolving action space, in which core, well-supported actions are implemented first, additional action units extend the system over time, and higher levels of actionability (e.g., validated applicable action units) emerge through continued use and evidence accumulation. Boundedness is thus not a limitation but a design principle, as it enables scalable development, targeted resource allocation, and continuous refinement of actionable knowledge systems.

8.7 Relation to Decision-Theoretic, Knowledge-Intensive Process Approaches

Planning,

and

The framework developed in this work is conceptually adjacent to several established research traditions that address the relationship between knowledge, context, and action. Understanding how action units and the actionability/applicability distinction relate to these traditions clarifies both the scope and the distinctive contribution of the proposed framework.

8.7.1 Knowledge-Intensive Processes The knowledge-intensive processes (KIP) literature in business process and service science addresses a closely related problem space. Di Ciccio, Marrella, and Russo (45) characterize knowledge-intensive processes as genuinely knowledge- and data-centric processes that require substantial flexibility at both design- and run-time, precisely because they cannot be fully pre-specified in advance. Motahari Nezhad and Swenson (46) identify case management as a domain in which the work procedure cannot be prescribed into machine programs, since the work is highly variable and must be figured out by knowledge workers each time, with work patterns emerging from the case as more information becomes available. This diagnosis is structurally parallel to the paper's central argument that formal process models and workflow engines assume fully specified procedures, whereas real-world knowledge-driven action requires context-sensitive, judgment-dependent decision-making that resists complete pre-specification. The KIP literature further addresses the role of knowledge workers as agents who combine formal process support with tacit, contextual judgment. This connects directly to the paper's treatment of human-actionability and the distinction between machine-executable and human-mediated operations. In both frameworks, the central challenge is not the absence of knowledge but the absence of mechanisms for evaluating whether and how available knowledge applies in a specific situation. However, the KIP literature and the adaptive case management tradition focus primarily on the management and execution layer and thus on how to support knowledge workers as they navigate incompletely specified processes at runtime. They take as given the knowledge resources that workers draw upon and focus on orchestrating their use. The framework proposed here addresses the prior and complementary challenge of how those knowledge resources should be represented so that their applicability conditions are explicit, evaluable, and available to both human and machine agents before and during execution. Action units can thus be understood as a representational foundation for the structured knowledge that KIP systems require but typically leave implicit or tacit. In this sense, the two frameworks are complementary, as KIP research addresses how to manage and orchestrate knowledge-driven processes at runtime, whereas the present

53 framework addresses how to represent the knowledge that grounds such processes in a form that makes applicability assessment possible.

8.7.2 Decision-Theoretic and Planning Approaches The graph-native decision system enabled by conditional action units shares important conceptual similarities with approaches in decision theory, causal inference, and AI planning, particularly in their use of conditional logic, structured representations of actions, and goal-oriented procedures. However, the proposed framework differs fundamentally in scope, purpose, and level of abstraction. Decision-theoretic and planning approaches typically assume that the relevant elements of a decision problem are already formally specified. These include well-defined system states, action alternatives, transition dynamics, and, in many cases, utility or objective functions. Within such models, the primary task is to select or optimize actions under uncertainty. In contrast, the framework developed in this work addresses the prior and largely underexplored challenge of how heterogeneous, partially formalized, and context-dependent knowledge can be represented, structured, and operationalized so that it can support action in the first place. Real-world knowledge rarely exists in the fully specified form assumed by decision-theoretic models. Instead, it is distributed across datasets, narratives, causal hypotheses, and procedural descriptions, often with varying levels of formalization and contextual dependence. The Semantic Units Framework, together with the notion of action units introduced in this work, provides a representation-level foundation for addressing this challenge. It enables the explicit integration of semantic content (assertional and general statements), contextual information, applicability conditions, and procedural and objective specifications. Through this integration, knowledge is transformed into operationally structured and context-aware representations that define what can be done, under which conditions, and for which goals. In this sense, the proposed framework should be understood as a form of pre-decisional infrastructure. It does not aim to replace decision-theoretic, planning, or KIP approaches, but to complement them by providing the structured inputs they typically presuppose. Action units, in particular, can serve as intermediaries between heterogeneous knowledge sources and formal decision models or case management systems, enabling the translation of real-world knowledge into representations that can be used within optimization, planning, or knowledge-intensive process management frameworks. At the same time, the framework remains applicable in settings where fully specified decision models are unavailable, incomplete, or continuously evolving. By supporting incremental formalization, contextualization, and validation of knowledge, it enables actionable reasoning and decision support even under conditions of uncertainty and partial knowledge. Thus, rather than competing with these approaches, the proposed framework fills a critical gap by providing the missing upstream layer that makes knowledge operational and prepares it for use in downstream decision-making, planning, and knowledge-intensive process management systems. The framework developed across this chapter thus provides the missing upstream layer that makes knowledge operational by preparing it for use in downstream decision-making, planning, and knowledge-intensive process management, and constituting the representational foundation for the post-FAIR knowledge infrastructures discussed in the conclusion.

54

9. Conclusion The increasing availability of data, metadata, and formally structured knowledge has not eliminated the persistent gap between knowledge and action in biodiversity science and other data-intensive sciences. While existing knowledge infrastructures have substantially improved the findability, accessibility, interoperability, and interpretability of knowledge, they remain limited in their ability to support context-sensitive, goal-oriented action. Addressing this limitation requires a conceptual shift from representing data to meaning and enabling operational use. In this work, we have argued that closing the knowledge–action gap requires a transformation in how knowledge is structured and related to processes. Rather than treating terms or data as primary units, we adopt a statement-centred perspective in which knowledge is represented as semantically coherent, context-sensitive units. Building on the Semantic Units Framework, we introduced action units as a representational layer that integrates semantic content with contextual information, applicability conditions, procedural knowledge, and explicit objectives. This integration enables knowledge to be not only interpretable, but operationally usable. In Searle's (20,21) terms, existing infrastructures have excelled at representing the word-to-world dimension of knowledge and thus what is the case, while providing limited support for the world-to-word dimension and thus what should or can be done and how. A central contribution of the framework is the explicit distinction between structural actionability and epistemic applicability. Structural actionability defines what operations can, in principle, be performed given a particular representation, by linking semantic units to procedures and workflows. Applicability, in contrast, characterizes whether such operations are valid in a specific context, based on the evaluation of context-dependent conditions. This distinction establishes a progression from executable knowledge (what can be done) to contextually justified action (what should be done here), and highlights that actionability alone is insufficient without epistemic grounding. More broadly, the framework distinguishes three complementary forms of operationalization. Epistemic operationalization establishes the relationship between knowledge and real-world situations, enabling recognition, description, and applicability assessment. Transformational operationalization enables the preparation, derivation, and integration of representations required for interpretation, analysis, and action. Interventional operationalization enables the execution of goal-directed actions that modify system states. These three forms are functionally interdependent. Epistemic operationalization provides the basis for assessing validity, transformational operationalization provides the representational means for processing knowledge, and interventional operationalization enables goal-directed change. Actionable understanding emerges only through the coordination of these processes. Accordingly, knowledge use is inherently processual and iterative, linking representation, evaluation, and intervention in continuous feedback cycles. In this sense, the three A’s of the framework for A3ctionable Understanding—Actionability, Applicability, and Action Units—are not independent contributions but a single integrated argument about what it means for knowledge to be usable in practice. Importantly, the framework does not assume universal actionability. Knowledge is often incomplete, context-dependent, and uncertain. Action units therefore define structural potential, while their valid use depends on context-specific applicability and, where available, empirical validation. This perspective emphasizes that actionable knowledge is not an intrinsic property of

55 representations, but emerges from the alignment of knowledge structures, operations, context, and agents. The framework is domain-agnostic and supports the integration of heterogeneous and progressively formalized knowledge. It is particularly relevant for the development of next-generation knowledge graphs and AI-supported systems, where semantic representations must be connected to executable operations, workflows, and context-aware reasoning. At the same time, it complements decision-theoretic and planning approaches by addressing the prior challenge of how knowledge can be structured and operationalized so that it can serve as input to such methods in the first place. From an implementation perspective, actionable knowledge infrastructures must be designed as integrated systems that represent knowledge in structured, semantically explicit forms, link representations to operations through well-defined input–output relations, make applicability conditions explicit and evaluable, and support both human and machine agents in interpreting and executing knowledge. Such systems must also remain bounded and incremental, focusing on high-value action spaces and expanding progressively as knowledge, data, and use cases evolve. The framework developed in this paper points toward a guiding principle for the design of next-generation knowledge infrastructures that extends and complements the FAIR and CLEAR Principles. We propose the TripleA Principle: Actionability, Applicability, and Auditability. Knowledge representations should be (i) actionable, explicitly specifying the operations through which they can be used and the agents and processes they support, (ii) applicable, explicitly specifying the conditions under which they can be validly used in specific contexts and enabling systematic evaluation before action is taken, and (iii) auditable, making the evidence base, validation history, and conditions of validity of actionable knowledge claims explicit, traceable, and assessable by both human and machine agents. Action units, as introduced in this work, constitute a concrete implementation of this principle within the Semantic Units Framework, but the principle itself is intended as a framework-agnostic requirement that future knowledge infrastructure designs can be evaluated against. Ultimately, closing the knowledge–action gap will depend not only on the availability of knowledge, but on our ability to represent and organize it in ways that make its conditions of validity, relevance, and use explicit. This framework also suggests a deeper point about the nature of understanding itself. Understanding is not a static internal state that precedes action or communication but rather an emergent property of engagement. In many real-world settings, understanding is refined and often first achieved through the application of knowledge in concrete contexts. We discover what we do and do not know by trying to act on what we believe we know. But understanding emerges through the act of articulating knowledge for others, i.e., through explanation, instruction, and collaborative reasoning, which forces the externalisation of what was previously implicit and the identification of gaps that abstract possession of knowledge conceals. In both cases, the demand for actionability, i.e., the requirement that knowledge be made operational, either for oneself in a specific context or for another agent in a communicable form, is what drives understanding rather than merely expressing it. By making the conditions of valid use explicit and representable, the framework proposed here supports both dimensions as it structures knowledge for context-sensitive action and organises it in a form that is communicable, assessable, and transferable across agents. In doing so, it provides a structured basis for problem-solving competence—understood here as the capacity to systematically align knowledge, context, and procedures toward a goal—for both human and machine agents. In this sense, actionable

56 understanding is not only a prerequisite for action, but also an outcome of it, and the infrastructure that supports actionable understanding is, at the same time, an infrastructure for shared and cumulative knowledge.

References 1.​ Wilkinson MD, Dumontier M, Aalbersberg IJ, Appleton G, Axton M, Baak A, et al. The FAIR Guiding Principles for scientific data management and stewardship. Sci Data. 2016 Dec;3(1):160018. doi:10.1038/sdata.2016.18 2.​ Vogt L. The CLEAR Principle: organizing data and metadata into semantically meaningful types of FAIR Digital Objects to increase their human explorability and cognitive interoperability. J Biomed. 2024;16(18):1–26. doi:10.1186/s13326-025-00340-7 3.​ Hogan A, Blomqvist E, Cochez M, D’amato C, de Melo G, Gutierrez C, et al. Knowledge Graphs. ACM Comput Surv. 2021;Synthesis Lectures on Data, Semantics, and Knowledge54(4):1–37. doi:10.1145/3447772 4.​ Guizzardi G. Ontology, Ontologies and the “I” of FAIR. Data Intell. 2020 Jan;2(1–2):181–91. doi:10.1162/dint_a_00040 5.​ Vogt L, Kuhn T, Hoehndorf R. Semantic units: organizing knowledge graphs into semantically meaningful units of representation. J Biomed Semant. 2024 May;15(7):1–18. doi:10.1186/s13326-024-00310-5 6.​ Vogt L. Rethinking OWL Expressivity: Semantic Units for FAIR and Cognitively Interoperable Knowledge Graphs Why OWLs don’t have to understand everything they say [Internet]. 2025. Available from: https://arxiv.org/abs/2407.10720 doi:10.48550/arXiv.2407.10720 7.​ Vogt L, Mons B. The Grammar of FAIR: A Granular Architecture of Semantic Units for FAIR Semantics, Inspired by Biology and Linguistics [Internet]. 2025. Available from: https://arxiv.org/abs/2509.26434v1 doi:10.48550/arXiv.2509.26434 8.​ Vogt L. The Semantic Ladder: A Framework for Progressive Formalization of Natural Language Content for Knowledge Graphs and AI Systems [Internet]. arXiv; 2026 [cited 2026 Mar 30]. Available from: http://arxiv.org/abs/2603.22136 doi:10.48550/arXiv.2603.22136 9.​ Arp R, Smith B, Spear AD. Building Ontologies with Basic Formal Ontology [Internet]. Cambridge, Massachusetts: The MIT Press; 2015. 248 p. Available from: http://mitpress.universitypressscholarship.com/view/10.7551/mitpress/9780262527811.001.00 01/upso-9780262527811 doi:10.7551/mitpress/9780262527811.001.0001 10.​ Gangemi A, Guarino N, Masolo C, Oltramari A, Schneider L. Sweetening ontologies with DOLCE. In: Gómez-Perez A, Benjamins VR, editors. European Knowledge Aquisition Workshop (EKAW-2002), Siguenza, Spain [Internet]. Springer; 2002. p. 166–81. Available from: http://www.ncbi.nlm.nih.gov/pubmed/18487833 PubMed PMID: 18487833. 11.​ Vogt L, Mikó I, Bartolomaeus T. Anatomy and the type concept in biology show that ontologies must be adapted to the diagnostic needs of research. J Biomed Semant. 2022 Dec;13(18):27. doi:10.1186/s13326-022-00268-2 12.​ Vogt L, Strömert P, Matentzoglu N, Karam N, Konrad M, Prinz M, et al. Suggestions for extending the FAIR Principles based on a linguistic perspective on semantic interoperability. Sci Data. 2025 Apr;12(1):688. doi:10.1038/s41597-025-05011-x 13.​ Ehrlinger L, Wöß W. Towards a Definition of Knowledge Graphs. In: Joint Proceedings of the Posters and Demos Track of the 12th International Conference on Semantic Systems — SEMANTiCS2016 and the 1st International Workshop on Semantic Change & Evolving Semantics (SuCCESS’16) [Internet]. Leipzig, Germany: CEUR Workshop Proceedings; 2016. Available from: http://ceur-ws.org/Vol-1695/paper4.pdf 14.​ Cyganiak R, Lanthaler M, Wood D. RDF 1.1 Concepts and Abstract Syntax; W3C Recommendation 25 February 2014 [Internet]. 2014. Available from: https://www.w3.org/TR/rdf11-concepts/

57 15.​ Frey J, Müller K, Hellmann S, Rahm E, Vidal ME. Evaluation of metadata representations in RDF stores. Semantic Web. 2019;10(2):205–29. doi:10.3233/SW-180307 16.​ GO FAIR Initiative [Internet]. Available from: https://www.go-fair.org/go-fair-initiative/ 17.​ Schultes E. FAIR digital objects for academic publishers. Duine M, editor. Inf Serv Use. 2023 Dec;44(1):15–21. doi:10.3233/ISU-230227 18.​ Weiland C, Islam S, Broder D, Anders I, Wittenburg P. FDO Machine Actionability - Version 2.1 FDO Forum Proposed Recommendation 19 August 2022 [Internet]. 2022. p. 10. Report August. Available from: https://docs.google.com/document/d/1hbCRJvMTmEmpPcYb4_x6dv1OWrBtKUUW5CEXB2gqsR o/edit# 19.​ Marconi D. On the Structure of Lexical Competence. Proc Aristot Soc. 1995 Jun 1;95(1):131–50. doi:10.1093/aristotelian/95.1.131 20.​ Searle JR. Speech acts: an essay in the philosophy of language. Cambridge University Press; 1969. 203 p. 21.​ Searle JR. A Taxonomy of Illocutionary Acts. Minn Stud Philos Sci. 1975;07:344–69. 22.​ Schulz S, Jansen L. Formal ontologies in biomedical knowledge representation. IMIA Yearb Med Inform 2013. 2013 Jan;8(1):132–46. PubMed PMID: 23974561. 23.​ Vogt L, Farfar KE, Karanth P, Konrad M, Oelen A, Prinz M, et al. Rosetta Statements: Simplifying FAIR Knowledge Graph Construction with a User-Centered Approach [Internet]. 2025. Located at: arXiv. Available from: https://arxiv.org/abs/2407.20007 24.​ Groth P, Gibson A, Velterop J. The Anatomy of a Nano-publication. Inf Serv Use. 2010;30(1–2):51–6. 25.​ Heath T, Bizer C. Linked Data: Evolving the Web into a Global Data Space. Springer Cham; 2011. XIII, 122. (Synthesis Lectures on Data, Semantics, and Knowledge). doi:10.1007/978-3-031-79432-2 26.​ Hartig O. Foundations of RDF* and SPARQL*. In: CEUR Workshop Proceedings 1912 [Internet]. 2017. Available from: https://ceur-ws.org/Vol-1912/paper12.pdf 27.​ Vogt L. Bona fideness of material entities and their boundaries. In: Davies R, editor. Natural and artifactual objects in contemporary metaphysics: exercises in analytical ontology. London: Bloomsbury Academic; 2019. p. 103–20. 28.​ Vogt L. Spatio-structural granularity of biological material entities. BMC Bioinformatics. 2010 May;11(289). doi:10.1186/1471-2105-11-289 PubMed PMID: 20509878. 29.​ Vogt L. Levels and building blocks—toward a domain granularity framework for the life sciences. J Biomed Semant. 2019 Dec;10(4):1–29. doi:10.1186/s13326-019-0196-2 30.​ Vogt L, König-Ries B, Alamenciak T, Brian JI, Arnillas CA, Korell L, et al. A Framework for FAIR and CLEAR Ecological Data and Knowledge: Semantic Units for Synthesis and Causal Modelling [Internet]. arXiv; 2025 [cited 2026 Mar 30]. Available from: http://arxiv.org/abs/2508.08959 doi:10.48550/arXiv.2508.08959 31.​ Hacking I. Representing and Intervening: Introductory Topics in the Philosophy of Natural Science. Cambridge: Cambridge University Press; 1983. 32.​ Genesereth MR, Nilsson NJ. Logical foundations of artificial intelligence. Los Altos, California: Morgan Kaufmann Publishers Inc.; 1987. 405 p. 33.​ Brooks RA. A Robust Layered Control System for a Mobile Robot. IEEE J Robot Autom. 1986 Mar;2(1):14–23. 34.​ Bandrowski A, Brinkman R, Brochhausen M, Brush MH, Bug B, Chibucos MC, et al. The Ontology for Biomedical Investigations. PLoS ONE. 2016;11(4):1–19. doi:10.1371/journal.pone.0154556 PubMed PMID: 27128319. 35.​ Kodikara KAS, Mukherjee N, Jayatissa LP, Dahdouh-Guebas F, Koedam N. Have mangrove restoration projects worked? An in-depth study in Sri Lanka. Restor Ecol. 2017;25(5):705–16. doi:https://doi.org/10.1111/rec.12492

58 36.​ Dickson I, Jones JPG, Paterson S, Trevelyan R, Butchart SHM, Catalano A, et al. Introducing a common taxonomy to support learning from failure in conservation. Conserv Biol. 2022;37(1):e13967. doi:10.1111/cobi.13967 37.​ Lebo T, Sahoo S, McGuinness D. PROV-O: The PROV Ontology; W3C Recommendation 30 April 2013 [Internet]. 2013. Available from: https://www.w3.org/TR/prov-o/ 38.​ Garijo D, Gil Y. Augmenting PROV with Plans in P-PLAN: Scientific Processes as Linked Data. In: LISC@ISWC [Internet]. 2012. Available from: https://api.semanticscholar.org/CorpusID:5826456 39.​ Soiland-Reyes S, Sefton P, Crosas M, Castro LJ, Coppens F, Fernández JM, et al. Packaging research artefacts with RO-Crate. Data Sci. 2022;5(2):97–138. doi:10.3233/DS-210053 40.​ Crusoe MR, Abeln S, Iosup A, Amstutz P, Chilton J, Tijanić N, et al. Methods included: standardizing computational reuse and portability with the Common Workflow Language. Commun ACM. 2022 May;65(6):54–63. doi:10.1145/3486897 41.​ Ciccio CD, Marrella A, Russo A. Knowledge-Intensive Processes: Characteristics, Requirements and Analysis of Contemporary Approaches. J Data Semant. 2015;4:29–57. 42.​ Nezhad HRM, Swenson KD. Adaptive Case Management: Overview and Research Challenges. 2013 IEEE 15th Conf Bus Inform. 2013;264–9.

Acknowledgements The author acknowledges the use of an AI-based language model to assist with language editing, phrasing, and manuscript structuring. The AI system was not used to generate scientific content, and all ideas, concepts, analyses, and conclusions presented in this work are those of the author.

Related documents

• Databases with Missing Values that are Governed by Missingness Mechanisms#1043177
• Proof-of-Retention: A Framework for Auditable Cross-Organization Data Sharing#1043178
• AkasicMEM: Governed Enterprise Memory for Agents#1043179
• Spend Classification Without Leakage: An Evaluation Harness and What It Changed in a Deployed System#1043180
• A Functional Pilot for Certified Freshness-Aware Semantic--Spatial Range Retrieval#1043181