Structured Knowledge Units (SKUs)

• Structured Knowledge Units (SKUs) are modular, explicit representations that bundle context, design decision, observation, and insight into a traceable research artifact.
• They are designed to operationalize knowledge by enabling reuse in AI-assisted software development, enterprise guidance, and semantic knowledge-graph applications.
• The SKU framework enhances traceability, reproducibility, and governance by converting implicit design decisions into structured, accumulative insights.

Structured Knowledge Units (SKUs) are modular, explicitly structured representations designed to preserve, reuse, and operationalize knowledge at a granularity smaller than whole documents and larger than isolated tokens or embeddings. In SHAPR, an SKU is defined as “a structured representation of an insight derived from a development cycle that explains how a specific design decision influences artefact behaviour,” while adjacent work describes closely related primitives such as Atomic Knowledge Units, Semantic Units, Action Units, and Knowledge Capsules for enterprise guidance, knowledge-graph organization, actionable infrastructures, and nonparametric LLM memory (Chan, 26 Mar 2026, Bakal, 16 Mar 2026). This suggests that “SKU” now functions both as a specific term within SHAPR and as a broader design pattern for bounded, typed, traceable knowledge objects across structured reasoning systems.

1. Definition and conceptual scope

Within SHAPR, SKUs are the “fundamental unit of knowledge captured within SHAPR,” introduced to prevent development knowledge from remaining implicit, scattered, or lost during iterative AI-assisted research software development (Chan, 26 Mar 2026). Their stated purpose is to convert artefact construction and evaluation into cumulative research knowledge, especially under conditions where conversational interaction environments make design reasoning difficult to reconstruct. The paper explicitly frames the problem as one of tacit knowledge loss, opacity of AI-assisted development, lack of a unit of accumulation, and weakened traceability and reproducibility.

The SHAPR formulation is narrower than a generic note, finding, or design rationale. The paper repeatedly distinguishes SKUs from conversation in the interaction workspace, from cycle records, and from design principles. An SKU records an explanatory relation between context, a design decision, observed artefact behaviour, and an interpreted insight, whereas design principles are higher-order generalizations derived from multiple SKUs (Chan, 26 Mar 2026). In the paper’s own cross-method mapping, “Experiment → Result,” “Case Study → Finding,” “Design Science → Design Principle,” and “SHAPR → Structured Knowledge Unit (SKU)” place the SKU as SHAPR’s preferred unit of accumulated knowledge.

A broader conceptual family appears in knowledge-graph research, where Semantic Units are defined as “semantically significant, named subgraphs in a KG” intended to enhance cognitive interoperability for users (Mustafa, 27 Nov 2025). Semantic Units are not called SKUs in that work, but they share the same core commitments: boundedness, explicit identity, semantic coherence, and reuse. This suggests that the contemporary SKU landscape includes both explanatory units derived from practice and graph-native units designed to make machine-oriented representations meaningful to human readers.

2. Structural schemas and granularity

The most explicit SKU schema in the provided literature comes from SHAPR. Each SKU “typically contains four core elements”: Context, Design Decision, Observation, and Insight (Chan, 26 Mar 2026). Context specifies the design situation or problem addressed; Design Decision specifies the implementation or modification; Observation records what happened during evaluation; and Insight explains why the outcome occurred. The paper is explicit that this is a structured and contextualized insight, but it does not prescribe a canonical syntax such as JSON, XML, or RDF.

A more operationally elaborate SKU-like form appears in Knowledge Activation, where an Atomic Knowledge Unit (AKU) is defined as “the minimal self-contained bundle of intent, procedural knowledge, tool bindings, organizational metadata, governance constraints, continuation paths, and validators” needed for one coherent action (Bakal, 16 Mar 2026). The paper positions AKUs as a specialization of AI Skills and argues that every AKU is a kind of structured knowledge primitive, but one that is explicitly action-oriented, governance-aware, and graph-native. Compared with SHAPR’s four-field SKU, the AKU adds runtime trigger conditions, tool interfaces, organizational metadata, approval requirements, continuation logic, and validator scripts.

Knowledge-graph work further extends the structural range. In the Semantic Units literature, statement units are the “smallest, independent propositions” meaningful to a human reader, while compound units group semantically meaningful collections of semantic units and assign them their own UPRI or GUPRI (Vogt et al., 2023). In Knowledge Capsules, the unit is formalized as $C = (S, R, O, I)$, where $S$ is subject, $R$ relation, $O$ object, and $I$ provenance identifier (Ju et al., 22 Apr 2026). Across these formulations, unit granularity varies from four-field explanatory records, to action schemas, to named subgraphs, to relational tuples with provenance. The recurring property is not one universal schema, but bounded structure plus explicit semantics.

3. Production, accumulation, and governance

In SHAPR, SKU creation is tied to the iterative Explore–Build–Use–Evaluate–Learn cycle. The paper states that SKUs are “typically extracted during the Learn stage,” but they depend on the whole cycle: exploration frames a design direction, building embodies a decision, use and evaluation generate observations, and learning converts those observations into structured insights (Chan, 26 Mar 2026). The paper also gives the explicit knowledge hierarchy “Observation … Insight … SKU … Pattern … Design Principle … SHAPR Refinement,” making SKUs both outputs of one cycle and inputs to subsequent cycles.

Enterprise-oriented work presents a different production pipeline. Knowledge Activation defines an end-to-end process of codification, compression, and injection that turns latent institutional knowledge into agent-executable form (Bakal, 16 Mar 2026). Codification extracts tacit and scattered knowledge from experts, runbooks, APIs, policies, postmortems, and platform systems. Compression restructures that knowledge into compact, high-signal units that maximize “knowledge density,” formally defined as

$\rho(k, \tau) = \frac{v(k, \tau)}{c(k)}.$

Injection then delivers the right unit to the right agent at the right moment. In that setting, the SKU-like unit is not primarily a research insight but an action-ready specification.

Continual structured reasoning introduces another production regime. K-DeCore decomposes reasoning into schema filtering and query building, then supports both with stage-specific memory and pseudo-data synthesis over abstract query structures (Chen et al., 21 Sep 2025). The schema-filter stage is treated as relatively task-agnostic after unifying different structured knowledge types into a database-like representation, while the query-building stage remains task-specific. This suggests that SKU production can be organized around transfer properties: reusable schema-selection units on one side, and reusable query-structure templates on the other.

4. Operationalization: graph-native, executable, and action-oriented units

A major contemporary shift is from representational structure to operational structure. Semantic Units were introduced partly because conventional RDF/OWL graphs are difficult to query and often contain many triples of no semantic significance for end users (Mustafa, 27 Nov 2025). By organizing a graph into named subgraphs with their own identifiers, types, and subject links, the framework creates a semantic unit layer over the base graph. The paper shows that these units support alternative query patterns and cleaner visualization, although it also concludes that semantic units do not automatically reduce SPARQL query complexity.

Action Units extend this logic into an explicitly operational register. They are defined as “structured extensions of plan specifications” that make applicability conditions and contextual grounding explicit as first-class typed semantic unit components (Vogt, 2 May 2026). Formally, an Action Unit is a compound unit with input and output units, plan specification units, an applicability conditions unit, and an objective unit. The paper distinguishes epistemic, transformational, and intervention action units, and then introduces conditional action units as executable IF-THEN structures in which the IF clause evaluates applicability conditions and the THEN clause specifies directive actions. The paper’s TripleA Principle—Actionability, Applicability, and Auditability—pushes SKU design beyond representational completeness toward justified action.

Knowledge Capsules and Pandora push operationalization further into LLM systems. Knowledge Capsules replace context-level text injection with External Key Value Injection, compiling symbolic capsules into attention-compatible key–value memory so that external knowledge “directly participate[s] in the model’s attention computation” (Ju et al., 22 Apr 2026). Pandora, by contrast, introduces BOX as a code-based unified knowledge representation $\mathcal{B} = (b, \Phi, \Psi)$, realized as a Pandas DataFrame and manipulated through executable code (Chen et al., 25 Aug 2025). In both cases, the structured unit is not merely stored or retrieved; it is compiled into an execution-compatible form that participates directly in reasoning.

5. Applications and empirical evaluation

The SKU idea has been applied across research software, e-commerce, item knowledge production, and benchmark construction. In product mapping, Q2K reframes same-SKU detection as question-driven knowledge construction: a Reasoning Agent generates disambiguation questions, a Knowledge Agent resolves them with focused web search, and a Deduplication Agent reuses validated question-answer traces (Seo et al., 1 Sep 2025). The five matching dimensions are Brand, Core Product Name, Variant, Specification, and Quantity. On a dataset of approximately 72,250 product pairs, the reported accuracies are 0.6737 for Rule-Based Matching, 0.9034 for Zero-Shot Inference, 0.9268 for Few-Shot Inference, 0.9344 for Web Search Inference, and 0.9562 for Q2K. The reuse mechanism activates in 22% of cases, and the average number of generated questions is 1.4 per product pair.

Benchmarking work has also adopted an explicit knowledge-unit perspective. SKA-Bench defines each instance using a question, an answer, positive knowledge units, and noisy knowledge units, spanning four structured knowledge forms: KG, Table, KG+Text, and Table+Text (Liu et al., 23 Jul 2025). In this benchmark, triples and table rows are treated as individual structured knowledge units, and the benchmark expands them into four ability testbeds: Noise Robustness, Order Insensitivity, Information Integration, and Negative Rejection. The reported results show that even strong models remain sensitive to noise, unit order, and unsupported inference; for example, DeepSeek-R1’s average rejection rate under pure-noise conditions is 78.71, and performance declines as required positive unit count increases.

Industrial-scale item systems provide another concrete realization. JD Oxygen AI Item Center organizes item knowledge around four ontology element types—Category, Attribute Key, Attribute Value, and Scenario Tag—and maps raw item data to normalized item knowledge through ontology engineering, Semantic Search then Discrimination, self-evolving LLMs/VLMs, and a unified item tunnel (AIIC et al., 26 Jun 2026). The platform reports end-to-end AI Item Library production with 94.2% precision and 82.8% recall, covers tens of thousands of JD categories, processes hundreds of millions of item updates per day, and has accumulated hundreds of billions of item-knowledge assets. In a different table-centric setting, HyperG treats cells as nodes and rows, columns, and the full table as semantically meaningful hyperedges, showing that SKU-like design can also be multi-granular and hypergraph-structured rather than tuple-structured (Huang et al., 25 Feb 2025).

6. Limitations, controversies, and methodological implications

Despite convergence around structured units, the literature does not yet provide a single settled SKU theory. SHAPR is explicit about several open issues: it does not specify a formal schema beyond the four fields, a machine-readable ontology, quality criteria for a “good” SKU, conflict resolution, lifecycle governance such as versioning or deprecation, or empirical evaluation of SKU effectiveness (Chan, 26 Mar 2026). Knowledge Activation makes broad claims about reducing correction cascades, accelerating onboarding, and lowering the institutional knowledge tax, but those outcomes are presented as claims and predictions rather than experimentally validated results (Bakal, 16 Mar 2026).

Knowledge-graph work is similarly mixed. The Biodiversity Exploratories Semantic Units implementation argues that named subgraphs improve cognitive interoperability and visualization, but it also states that “SPARQL query complexity when querying over semantic units is likely not lower than the complexity of conventional queries,” and no proper user evaluation was conducted (Mustafa, 27 Nov 2025). Action Units add an operational layer, yet their broader adoption would require typed schemas, workflow bindings, and validation infrastructures robust enough to make applicability conditions executable in production settings (Vogt, 2 May 2026).

Evaluation frameworks and application papers also leave important methodological gaps. Q2K does not specify a confidence model, train/validation/test split, contradiction-resolution procedure, or detailed human-in-the-loop protocol (Seo et al., 1 Sep 2025). SKA-Bench studies only one granularity per modality and measures understanding after linearization into prompts, so some benchmark difficulty reflects serialization effects rather than purely native structured reasoning (Liu et al., 23 Jul 2025). HyperG introduces LLM-generated contextual augmentation for sparse tables but does not specify explicit hallucination controls for the generated row, column, or caption descriptions (Huang et al., 25 Feb 2025). Taken together, these limitations suggest that SKU research is moving from ad hoc structuring toward explicit unit design, but the field still lacks a fully unified standard spanning schema, validation, provenance, execution, and lifecycle management.