# Mekiki Framework Papers T1–T3: Consolidated Markdown for AI Reading **What this is:** This file consolidates papers T1 through T3 of the Mekiki Framework into a single Markdown document. The official version of each paper is published as a preprint at the DOI listed below. This derivative file is provided as material for AI to read. **Canonical source:** Cite or reference the DOI for the relevant paper, not this file. - **T1 — Mekiki Framework, core:** https://doi.org/10.31235/osf.io/cwkav_v1 - **T2 — Philosophy as Cognitive Assay:** https://doi.org/10.31235/osf.io/e9qw5_v2 - **T3 — Decomposing Agency, Isolating Answerability:** https://doi.org/10.35542/osf.io/hvbfe_v1 **Two ways to read:** An AI can be asked about the overview and applications. To understand the core of the framework correctly — especially answerability — read the human-facing commentaries in Japanese: - **T1 commentary:** https://researchmap.jp/ketomy/others/53500468 - **T2 commentary:** https://researchmap.jp/ketomy/others/53699246 - **T3 commentary:** https://researchmap.jp/ketomy/others/54141521 **How the three papers relate:** T1 isolates externalization cost (Ext.cost) and specification; T2 separates Sein (factual judgment) from Sollen (evaluative judgment) and establishes the scoring asymmetry; T3 isolates answerability on top of these. The three papers are cumulative rather than independent and should be read in this order. **License:** CC BY 4.0. Reuse, adaptation, and use for AI training are permitted on condition of attribution. **Author:** Kengo Tomita, Institute of Technology, Shimizu Corporation. **Editorial treatment:** Figure images are omitted. Their claims are retained as text-only legends, and all source tables are retained in Markdown. --- # Paper T1 — Domain-Native Development: A Mekiki Framework for AI-Assisted Knowledge Work **Preprint version:** 1 **Canonical DOI:** https://doi.org/10.31235/osf.io/cwkav_v1 **Kengo Tomita** Institute of Technology, Shimizu Corporation --- ## Abstract When experienced professionals retire, organisations lose not their documented procedures but the accumulated judgement that determines how those procedures are applied. This paper proposes a two-component framework—termed the *mekiki* framework, after the Japanese concept of expert discernment—for understanding how artificial intelligence (AI) makes this otherwise invisible knowledge structure observable. The framework identifies two conditions governing whether tacit knowledge can be successfully externalised into formal artefacts—software, documents, analytical tools: *externalisation cost* (the technical barriers to articulating and implementing knowledge) and *specification cost* (the barrier constituted by the domain expertise required to determine what should be built, how quality should be judged, and what should be excluded). Drawing on a revelatory case study in which a domain expert with no programming experience developed and publicly deployed a web application in ten working days using AI tools, the study shows that externalisation cost drops dramatically under AI assistance while specification cost persists unchanged. A domain-ablation experiment, in which two independent AI systems produced technically competent but domain-inappropriate applications from the same dataset without expert guidance, provides supporting evidence for the separability of the two components. Grounded in the SECI knowledge management tradition, the framework offers organisations a structural vocabulary for identifying which knowledge components are at risk of irreversible loss through generational transition—and which can be preserved through AI-mediated dialogue while retiring experts remain available. --- ## Keywords domain-native development, mekiki framework, tacit knowledge, specification cost, externalisation cost, AI-assisted knowledge work, SECI model --- ## 1. Introduction In knowledge-intensive industries, the retirement of senior practitioners creates a distinctive form of organisational loss. What disappears is not procedural knowledge—which can be documented in manuals and standard operating procedures—but the accumulated judgement that determines how procedures are applied, when they should be overridden, and what problems they fail to anticipate. In the construction industry, where project-specific conditions vary enormously and technical decisions cascade through decades of a building's service life, this form of expertise loss is particularly acute. Japan's construction sector, facing concurrent demographic contraction and an aging technical workforce, represents one of the most severe instances of this challenge globally. Three researchers at a single corporate research institute illustrate the problem in structural terms. The first, a senior specialist now retired, accumulated thirty years of expertise in materials evaluation for indoor environmental quality. His published reports and test protocols survive, but the judgement underlying them—which anomalous readings warrant investigation, which client specifications contain hidden contradictions, which measurement conditions produce misleading results—resided in professional intuition that his documentation captured only partially. His knowledge was extensive but largely tacit. The second, a junior researcher who inherited the senior specialist's portfolio, has access to AI-assisted analysis tools and a comprehensive institutional archive. Yet the translation of this inheritance into new analytical instruments has proven difficult. The bottleneck is not access to information or computational tools—it is the *judgement* required to determine what the tools should do. The information exists; the capacity to specify its application does not. The third, the present author, works in an adjacent technical domain. Trained in pharmacology (PhD in Medicine) with additional certification in electrical engineering, the author used AI tools to develop and publicly deploy an EV charging infrastructure application in ten working days as a personal project motivated by domain interest—a task that would conventionally require a software development team and months of iteration. The enabling condition was not programming skill (which the author lacked entirely) but deep domain knowledge in the application's subject matter, combined with AI systems that could translate that knowledge into functioning code. These three individuals shared the same institutional context: identical organisational position, access to AI tools, and comparable educational attainment. What differed was the relationship between their domain expertise and the task at hand. The senior specialist possessed the relevant expertise but lacked tools to externalise it directly. The junior researcher possessed the tools but lacked the expertise to direct them. The author possessed both domain expertise and AI tools, in a domain where the two could be coupled without intermediaries—and produced a working product in days. This pattern—in which domain knowledge, not technical implementation skill, constitutes the binding constraint on AI-assisted production—has been observed across multiple fields but has not been systematically characterised. Structurally similar decompositions have appeared independently: prediction versus judgement in economics (Agrawal, Gans & Goldfarb, 2018), envisioning gaps in cognitive science (Subramonyam et al., 2024), specification-process-evaluation alignment in interaction design (Terry et al., 2023), and agent-augmented SECI modes in knowledge management (Böhm & Durst, 2026). That multiple disciplines have converged on analogous two-component structures suggests an underlying phenomenon awaiting unified characterisation. This paper provides that characterisation. We define *domain-native development* (DND) as the direct production of functional artefacts by domain experts using AI as a translation medium, without requiring intermediary professionals (software engineers, technical writers, data architects) to convert domain knowledge into formal implementations. We propose a two-component framework—*externalisation cost* and *specification cost*—that identifies two distinct conditions governing whether tacit knowledge can be successfully converted into working artefacts. Externalisation cost comprises the technical barriers to articulating and implementing knowledge; specification cost comprises the barrier arising from the domain expertise required to determine *what* should be built, *how* quality should be judged, and *what* should be excluded. The framework is grounded in the SECI knowledge management tradition (Nonaka & Takeuchi, 1995), illustrated through a single revelatory case study with domain-ablation experiment, and carries safety implications drawn from the author's pharmacological training. The framework's social purpose is immediate. The window during which retiring domain experts can externalise their specification capacity through AI-mediated dialogue is finite and closing. A structural vocabulary for this process—one that makes the distinction between transmissible and non-transmissible knowledge components organisationally negotiable—is a precondition for effective intervention. --- ## 2. Conceptual Framework ### 2.1 Two-component decomposition When domain knowledge is converted into a formal artefact—a software application, a technical document, an analytical tool—two categories of cost are incurred. We term these *externalisation cost* and *specification cost*. **Externalisation cost** (also referred to as communication cost in practical contexts) denotes the technical barriers to converting tacit knowledge into explicit, formal representations. This includes the effort of articulating domain concepts in language that implementation tools can process, the iterative correction cycles that arise from imprecise articulation, and the translation losses inherent in any conversion from one representational medium to another. The term follows Nonaka's (1994) use of *externalisation* to describe the conversion of tacit knowledge into explicit forms within the SECI (Socialisation–Externalisation–Combination–Internalisation) model of organisational knowledge creation. **Specification cost** denotes the barrier arising from the domain expertise required to determine what should be built. This encompasses priority setting (which features matter and which are noise), quality criteria (what constitutes an acceptable output), information filtering (what should be deliberately excluded), and organisational context sensitivity (what constraints apply in this particular deployment). Evaluation criteria are included within specification cost rather than treated as a separate component because, in AI-assisted workflows, the capacity to judge output quality draws on the same domain knowledge as the capacity to specify what should be produced; both require knowing what "good" looks like in a given domain. Specification cost is not reducible to requirements elicitation in the software engineering sense; it includes judgements that the domain expert may not be able to articulate as formal requirements but can recognise and evaluate when presented with candidate implementations. Specification cost should be distinguished from domain expertise itself: domain expertise is a resource that the practitioner possesses; specification cost is the degree to which a task demands domain expertise. Under conventional conditions—human-to-human knowledge transfer, manual document authoring, professional intermediaries—these two cost components are entangled and inseparable. We term this two-component decomposition the *mekiki* framework, adopting the Japanese term for expert discernment—the capacity to see what others cannot see in a domain. Japanese knowledge management concepts have entered international discourse through two distinct routes: practice-driven terms such as *kaizen* and *gemba*, which emerged from Toyota's production system and spread through industrial adoption before being codified in management literature (Imai, 1986; Womack, Jones & Roos, 1990); and theory-driven terms such as *ba* (Nonaka & Konno, 1998) and *SECI* (Nonaka & Takeuchi, 1995), which were introduced as formal concepts in academic publications. *Mekiki* bridges these two lineages: it originates as a vernacular term from Japanese professional practice, but enters international discourse through academic formalisation. Where these earlier concepts addressed the *process* of knowledge creation, the mekiki framework addresses the *cost structure* of knowledge conversion under AI assistance. The act of specifying (deciding what to build) and the act of externalising (articulating that decision for others to implement) co-occur in every design conversation, every requirements document, and every mentoring relationship. Figure 1 situates this two-component decomposition within the SECI framework. **Intellectual lineage.** The framework operates within a lineage that runs from Polanyi through Nonaka to the present analysis. Polanyi (1966) established that knowledge contains a dimension that resists articulation: "we can know more than we can tell." Nonaka and Takeuchi (1995) operationalised this insight by treating tacit knowledge as convertible through externalisation, deliberately bracketing Polanyi's deeper claim about inarticulate knowledge for practical purposes. Their later work extended tacit knowledge toward "Wisdom"—knowledge inseparable from lived experience and ethical judgement (Nonaka & Takeuchi, 2019)—though they did not possess an experimental system for separating Wisdom from more readily externalised forms of tacit knowledge. The present framework identifies, within Nonaka's Externalisation phase, two conditions that determine whether externalisation produces domain-appropriate outputs: a component reducible by AI (externalisation cost) and a component (specification cost) whose deeper layers—grounded in private networks, accumulated life experience, and the finite embodied perspective of the knower—structurally correspond to the dimension Polanyi bracketed and to what Nonaka and Takeuchi later termed Wisdom (see Section 5.6, Future directions). ### 2.2 The distillation model of authorship The relationship between human expertise and AI processing can be modelled as a distillation process. The raw material is collective knowledge—the accumulated information, conventions, prior art, and community practices available within a domain. The apparatus is the AI system, which processes this material while substantially attenuating the interpersonal biases (hierarchy, ego protection, political calculation, social desirability) that characterise human-to-human knowledge transfer.^[The claim is not that AI is bias-free. RLHF training residuals, model personification tendencies, and implicit evaluator assumptions persist. The claim is narrower: the specific interpersonal confounders that make externalisation cost and specification cost inseparable in human dyads—face-saving, status negotiation, evaluation anxiety—are selectively reduced.] The distillation conditions are the human expert's specification: what to retain, what to discard, what quality criteria to enforce, what audience to address. Under this model, authorship resides in the setting of distillation conditions, not in the operation of the apparatus. The same raw material processed through the same apparatus under different distillation conditions yields different products—a property directly observable when different domain experts use the same AI system to produce artefacts in their respective areas. Conversely, identical distillation conditions applied through different apparatus (different AI models) produce outputs that converge on the same structural features, as the ablation experiment in Section 4 demonstrates. The model also accounts for iterative refinement. Distillation is rarely a single pass; the expert evaluates intermediate products, adjusts conditions, and re-processes. This multi-pass structure is visible in the dialogue logs deposited with this paper, where specification evolves through progressive refinement rather than appearing fully formed. Figure 2 illustrates this distillation model. ### 2.3 Positioning among existing frameworks The externalisation cost / specification cost distinction is not the first two-component decomposition proposed for AI-era knowledge work. Table 1 situates the present framework among four independent contributions that address structurally similar phenomena. **Table 1: Existing two-component frameworks and their relationship to the present model** | Discipline | Decomposition | Source | Relationship to Ext. cost / Spec. cost | |---|---|---|---| | Economics (decision theory) | Prediction / Judgement | Agrawal, Gans & Goldfarb (2018) | Operates at the level of decisions under uncertainty. Spec. cost subsumes judgement but extends to priority setting, quality criteria, and information filtering—dimensions visible in the ablation results (Section 4) where AI produced functional applications lacking not predictions but *design judgements* | | Cognitive science (HCI) | Gulf of execution / Gulf of envisioning | Subramonyam et al. (2024) | Three gaps (capability, instruction, intentionality) map onto externalisation cost. Specification cost—knowing what the artefact should be—is presupposed as a given goal, whereas the present framework treats it as the primary variable | | Interaction design | Specification / Process / Evaluation alignment | Terry et al. (2023) | Three-part taxonomy compresses to two components: process alignment and evaluation mechanics map onto externalisation cost; specification and evaluation *criteria* constitute specification cost. Compression from descriptive taxonomy to economic framework enables directional predictions | | Knowledge management (SECI) | GRAI framework (8 interaction fields) | Böhm & Durst (2026) | Adds a human/machine agent dimension to SECI's four modes. The Ext./Spec. distinction is orthogonal: it applies *within* any SECI mode and any GRAI interaction field, providing an analytical layer that complements their descriptive mapping | These independent observations—from economics, cognitive science, interaction design, and knowledge management—describe structurally similar phenomena from different disciplinary vantage points. The externalisation cost / specification cost framework provides a parsimonious two-component decomposition that (a) subsumes or is orthogonal to each existing model, (b) connects to an established knowledge management tradition via Nonaka's SECI, and (c) generates directional predictions illustrated through domain-ablation experiment (Section 4). The structural gap extends to AI systems design itself: Tomašev, Franklin, and Osindero (2026) proposed a comprehensive delegation protocol for AI agent systems including task decomposition, monitoring, and verification mechanisms, yet identified no principled criterion for determining which task components can be safely delegated and which require human judgement—precisely the separation that the externalisation cost / specification cost distinction provides. The framework generates a specific directional prediction: in domains where specification cost is high, AI deployment without domain expert involvement will fail to produce domain-appropriate outputs, regardless of the AI system's externalisation capabilities. Section 4 provides supporting evidence for this prediction.^[The pattern of independent discovery across disciplines is itself informative. Multiple discovery typically indicates that a phenomenon has become observable through shared environmental change—in this case, the widespread availability of generative AI—rather than through the insight of any single researcher (Merton, 1961; Lamb & Easton, 1984).] ### 2.4 Applicability conditions The substantial reduction in externalisation cost observed in domain-native development applies primarily to individual or small-team production in exploratory phases. Organisational coordination costs—regulatory compliance, multi-stakeholder approval, accessibility auditing, security review—constitute a separate cost category not addressed by the externalisation/specification distinction. Similarly, the framework addresses *technical* externalisation costs—those arising from format conversion, language translation, and procedural compliance. Social and psychological barriers to knowledge sharing (cf. Edmondson, 1999), while relevant to organisational knowledge flows, constitute a distinct phenomenon outside the present framework's scope—though their role as an identification condition is examined in Section 5.1. The framework does not predict that all knowledge work will be transformed equally. The appropriate unit of analysis is not the task but the *component ratio within a task*. A routine compliance report (high externalisation ratio, modest specification ratio) and a novel product concept (low externalisation ratio, high specification ratio) are both "documents," but their cost compositions differ fundamentally. The framework predicts that AI assistance transforms the former far more than the latter—a prediction consistent with widespread observations of differential AI impact across knowledge work categories. --- ## 3. Case Study The following case is positioned as a revelatory case study (Yin, 2018)—selected not for prevalence but for its capacity to make observable a phenomenon that the framework predicts but that has not been empirically documented at the level of individual design decisions. A researcher with a doctoral background in pharmacology and no prior programming experience developed and publicly deployed a progressive web application (PWA) for searching electric vehicle (EV) fast-charging infrastructure across Japanese expressways. The application comprises 32 source files, a bespoke dataset of 583 records (one row per charger model at each directional location, derived from 662 individual charging ports across 461 directional locations at 255 physical sites) with 22 technical fields each, and supports bilingual (Japanese/English) operation. Development spanned ten working days using two AI tools in a deliberate division of labour: an AI chat interface for design discussion and review, and an AI coding agent for implementation. Community reception on social media within 48 hours of public release confirmed practical utility rather than novelty alone. Time allocation across the development process proved revealing. Actual AI code generation accounted for approximately 5% of total working time. Design discussions, specification review, and iterative refinement consumed roughly 40%. The remainder comprised environment setup (20%), deployment and testing (25%), and breaks (10%). In short, externalisation cost was compressed to a fraction of the total effort; the dominant activity throughout was specification—deciding what the application should and should not do. ### 3.1 Specification cost as the determinant of design Four categories of design decisions illustrate how domain knowledge, rather than prompt engineering skill, governed application quality. **Current-based data architecture.** The dataset was structured around voltage and amperage rather than the kilowatt ratings typically displayed by charging networks. This decision derives from the physics of the dominant fast-charging protocol in Japan:^[CHAdeMO, the protocol used by the majority of Japanese fast-charging stations, operates under current control. The charger adjusts output current while voltage tracks the vehicle battery's state of charge.] chargers operate under current control, with voltage tracking the vehicle battery's state of charge. Kilowatt output is therefore a *result*, not a control variable. The dataset accordingly records maximum voltage, maximum amperage, and sustained amperage as separate fields. This architecture enables a vehicle-specific output calculation—min(vehicle voltage, charger voltage) × min(vehicle current, charger current)—that no existing EV information service provides. The formula is self-evident once the underlying physics is understood, yet it remains absent from publicly available charging guides, constituting tacit knowledge that the author's combined electrical engineering certification and EV ownership made available. In the domain-ablation experiment (Section 4), neither AI model operating without this domain knowledge included current-based fields. **Conditional display suppression.** When a user configures their vehicle profile, the application suppresses boost-mode indicators for chargers whose boost current falls at or below the vehicle's maximum charging current. For instance, a vehicle limited to 200 A receives no benefit from a charger's 350 A boost capability; displaying this information would be misleading. This judgement combines physical understanding (current-limited charging behaviour) with information design principles (progressive disclosure—hiding information irrelevant to the specific user). The suppression logic appeared in neither ablation output. **Network operator discrimination.** Japanese expressway charging stations are managed by multiple, non-interchangeable networks. Two stations at the same service area may belong to different operators with incompatible authentication systems. The application records operator identity as a distinct field and displays non-default operators with visible badges. In the ablation experiment, one AI model misinterpreted the operator management flag as an "immediate availability" indicator, producing a structurally incorrect data model. **Non-adoption decisions.** Perhaps the strongest evidence for specification cost lies in what the application deliberately excludes: real-time availability (unreliable API data that could misdirect users), charging time predictions (nonlinear charging curves that vary with temperature, state of charge, and battery degradation), distance calculations (expressway network topology makes simple distance metrics misleading), and tariff information (rates vary by card type, plan, and time of day). Each exclusion reflects domain knowledge about what would harm rather than help users. In the ablation experiment, both AI models exhibited "total field filtration"—converting every available data field into a user-facing filter, unable to judge which information to withhold. These design decisions were not products of prompt engineering technique. Each originated from the author's domain knowledge—electrical engineering fundamentals, vehicle ownership experience, and understanding of user behaviour at charging stations. They instantiate specification cost as defined in Section 2. ### 3.2 Reversal of the translation direction Conventional knowledge-to-product workflows follow a linear path: the domain expert writes natural-language requirements, which a software engineer interprets and implements. Ambiguities in the requirements document generate iterative clarification cycles—a primary source of externalisation cost. The development process observed in this case study inverted this sequence. The author generated working prototypes through AI dialogue, producing functional code that served as the specification itself. When a professional engineer later needed to rebuild components for production quality, the prototype—not a written requirements document—became the reference artefact. A working prototype is structurally less ambiguous than a natural-language specification: behavioural questions can be answered by running the code rather than interpreting text. This reversal was most evident in a specification-driven development episode. The author prepared a detailed specification document, a visual prototype, and a machine-readable instruction file before engaging the coding agent. The search interface was implemented in eight minutes with zero clarification requests. By contrast, earlier sessions conducted without such preparation required repeated correction cycles as the AI agent guessed at unstated requirements. The contrast demonstrates that externalisation cost reduction is not automatic; it depends on the quality of specification provided. ### 3.3 Exploratory value of reduced externalisation cost The primary value of reduced externalisation cost may not be the efficiency of producing finished products, but the increased number of exploratory attempts that become feasible when creation cost is low. The present case study provides direct evidence. The application's specification document records twelve features that were explicitly proposed and rejected during development: a route-distance column (rejected because highway interchange connections introduce measurement ambiguity), a regional filtering system (rejected because certain expressways span multiple administrative jurisdictions, breaking any area-based classification), a power-sharing column (rejected because two-port chargers exhibit variable output distribution that cannot be represented in a single field), and nine additional items, each rejected on domain-specific grounds. In a conventional outsourcing arrangement, most of these features would never have been proposed—and those that were proposed would have incurred significant cost to evaluate and discard. A design-level pivot illustrates the same dynamic at larger scale. Three complete landing page designs—dark-minimal, clean-light, and editorial—were generated and evaluated within a single session. The editorial approach was selected based on the author's judgement that technical content requires visual authority; the other two were discarded immediately. In traditional outsourcing, commissioning three complete design alternatives and discarding two would represent a substantial cost penalty. Under near-zero externalisation cost, such early-stage exploration becomes routine. The critical observation is that each rejection was itself an exercise of domain expertise—specification cost applied to *evaluation* rather than creation. The twelve rejected features and two discarded designs represent specification cost in its evaluative dimension: the expert's judgement about what *not* to build. This dynamic reverses when externalisation cost reduction is applied to organisational automation rather than individual exploration. Individual exploration operates under low failure cost (the explorer bears the consequences); organisational automation operates under high failure cost (errors propagate through systems and users). This asymmetry is examined in Section 5.3. ### 3.4 Two identification points The case study yields two observations that the Discussion will generalise: 1. **Prototype-as-specification** constitutes a structural mechanism for reducing externalisation cost. The domain expert's output is no longer a natural-language description to be interpreted, but a functioning artefact to be refined. 2. **Increased exploratory attempts** represent a second-order effect of externalisation cost reduction. When creating and discarding prototypes becomes inexpensive, the bottleneck shifts entirely to the domain expert's judgement about what is worth pursuing—specification cost in its evaluative dimension. --- ## 4. Methods The preceding case narrative and the Discussion that follows draw on three categories of evidence: the dialogue logs generated during AI-assisted development, the outputs of a domain-ablation experiment, and the published dataset. This section describes each. ### 4.1 Dialogue logs as cognitive records The methodological contribution of this study lies not in the application itself but in the dialogue logs generated during its development. These logs constitute a complete record of the author's decision-making process: which design options were considered, which were adopted and why, and—critically—which were rejected and on what grounds. Rejection rationale is a category of knowledge that typically survives in neither technical manuals nor project reports. This record shares analytical kinship with Think Aloud Protocols (TAP; Ericsson & Simon, 1993), but with a significant methodological advantage: it is non-intrusive. TAP imposes a dual-task burden by asking participants to verbalise while performing; AI-assisted development naturally produces verbalisation as its primary mode of operation. The dialogue log is not a secondary artefact overlaid on the work process—it *is* the work process. This property also mitigates the primary risk of the author-as-subject design: because decisions were recorded at the moment of occurrence rather than reconstructed retrospectively, the logs constrain post hoc rationalisation in a way that interview-based or memoir-based methods cannot. A technical precondition for this approach is the context window capacity of current large language models (approximately 200,000 tokens at the time of this study), which permits extended multi-turn dialogues that accumulate project-specific context. This accumulated context creates a positive feedback loop: as the AI system's representation of the project improves, the author's specification becomes more efficient, accelerating knowledge extraction. This technical parameter should be noted as a reproducibility condition. ### 4.2 Domain-ablation experiment To illustrate the framework's prediction that specification cost cannot be substituted by externalisation cost alone, a domain-ablation experiment was conducted. The experiment is designed not as a benchmark but as a counterfactual illustration: it asks what happens when externalisation capability is present but specification is absent. The same raw dataset—662 charging port records derived from the author's structured CSV but converted to JSON with field names replaced by the network operator's abbreviated API labels (e.g., `im`, `cap`, `typ`), deliberately reducing human readability to prevent domain knowledge leakage through column naming—was provided to two commercially available AI systems: Claude Code (Anthropic, Claude Opus 4.6) and ChatGPT 5.2 (OpenAI, Pro mode). Each received a single generic prompt requesting creation of an EV charging station search application with filtering and sorting capabilities, with design left to the AI's discretion. No follow-up instructions were given. Neither model posed domain-specific queries; the only author interactions during the Claude Code session were technical file-creation approvals, which were set to auto-approve after the first prompt. This one-shot, zero-interaction design prevents contamination from the author's iterative specification. A critical methodological choice was using the raw API data (JSON with abbreviated field names such as `im`, `cap`, `typ`) rather than the author's structured CSV dataset. The CSV column names—`max_amps`, `sustained_amps`, `boost_minutes`—themselves encode domain knowledge that would partially compensate for specification cost absence. Both models produced functional applications rapidly: Claude Code generated a single HTML file (353 lines) in under two minutes; ChatGPT produced a single HTML file (1,145 lines, with JSON data embedded inline) after approximately 35 minutes of processing. Both outputs featured competent user interfaces with filtering and sorting capabilities, confirming full externalisation cost capability. Table 2 catalogues the convergent deficiency patterns across three analytical layers. **Table 2: Convergent deficiency patterns in domain-ablation experiment** | Analytical layer | Specification-absent output | Specification-present design (author) | |---|---|---| | **Data interpretation** | | | | Physical architecture | Kilowatt ratings only; no current-based data model | Voltage/amperage separation enabling vehicle-specific output calculation (V × A) | | Temporal resolution | No boost/sustained current distinction | Separate fields for maximum amperage (boost) and sustained amperage | | Data hierarchy | Flat presentation of all 662 ports (Claude) / Direction-level aggregation into 465 groups without site-level consolidation (ChatGPT) | Three-level hierarchy: 662 ports → 461 locations → 255 sites | | Operator semantics | Management flag misinterpreted as availability indicator^a^ | Operator identity as distinct field with network-specific badge display | | **UI/UX** | | | | Output personalisation | No vehicle-specific calculation | min(vehicle V, charger V) × min(vehicle A, charger A) per configured profile | | Conditional logic | No display suppression based on user context | Boost indicators hidden when charger boost ≤ vehicle max current | | Data normalisation | Direction field ignored entirely (Claude) / Eight directional labels retained without normalisation (ChatGPT)^b^ | Normalised to two physical directions per location | | **Information architecture** | | | | Navigation model | No route-planning orientation | Route-based (expressway → service area → charger) | | Curation | No recommendation or prioritisation logic | Operator badges, output ranking, contextual recommendations | | Filter design | Minimal filters with most fields ignored (Claude) / Extensive filters including single-value fields (ChatGPT); neither selected filters based on user relevance | Selective filtering; irrelevant or misleading fields deliberately excluded | ^a^ One model interpreted the `isManaged` boolean as "immediate availability," producing a structurally incorrect data model. ^b^ One model ignored the direction field entirely; the other retained eight directional labels without normalising to two physical directions per location. Both models independently exhibited structurally convergent deficiency patterns across all three layers, though with variation in degree. At the *data interpretation* layer: absence of current-based architecture (kilowatt ratings only), no boost/sustained current separation, flat presentation of all 662 ports without grouping in one model and direction-level aggregation (465 groups) without site-level consolidation in the other (versus the author's three-level hierarchy: 662 ports → 461 locations → 255 sites), and misinterpretation of operator management fields. At the *UI/UX* layer: no vehicle-specific output calculation, no conditional display logic, and failure to normalise directional labels (one model ignored the direction field entirely while the other retained eight directional labels, neither normalising to the two physical directions per location used in the author's design). At the *information architecture* layer: no route-planning orientation, no recommendation logic, and—most tellingly—divergent but equally specification-absent filter strategies—one model ignored most fields while the other converted all available fields into filters including one with a single possible value—with neither selecting filters based on domain-informed user relevance. Each row in Table 2 represents a design decision requiring domain knowledge—electrical engineering fundamentals, charging protocol physics, user behaviour at highway service areas—that no amount of prompt engineering sophistication could substitute. The convergent failure pattern across two independent models operating on identical data is consistent with the framework's prediction: when specification cost is absent, externalisation cost reduction alone produces artefacts that are technically competent but domain-inappropriate. One model's interpretation of the `isManaged` field as "immediate availability" exemplifies the pattern precisely—externalisation capacity (converting a field into a functional UI filter) operated flawlessly, but specification (knowing what the field *means* in the charging network context) was absent. ### 4.3 Data availability Three data layers are deposited on figshare (Nature Portfolio's recommended repository): the charging infrastructure dataset (583 × 21 CSV; one field—installation date, collected through independent research—is withheld from the public release), the ablation experiment input data (662-record JSON with abbreviated field names, as provided to both AI systems) together with outputs and session logs, and full dialogue logs in the original Japanese (machine-analysable by contemporary AI systems, rendering language a non-barrier to verification). Dialogue logs have been reviewed to remove references to third-party individuals and proprietary organisational information; redacted passages are marked with [REDACTED] tags to maintain analytical traceability.^[To maintain double-anonymous review integrity, data are submitted as supplementary materials during review. The figshare deposit will be made publicly accessible upon acceptance, with the DOI reserved in advance.] ### 4.4 AI use disclosure In accordance with Springer Nature's policy on AI-assisted writing, the following disclosure is provided. Three AI systems were used throughout this research: an AI chat interface (Claude, Anthropic) for design discussion, conceptual development, and manuscript drafting; an AI coding agent (Claude Code, Anthropic) for software implementation; and an additional AI system (ChatGPT, OpenAI) for the domain-ablation experiment (Pro mode, version 5.2) and independent manuscript review (Pro mode, five review cycles). The AI chat interface also served as the primary data collection instrument, generating the dialogue logs analysed in this study. In all cases, the author provided domain knowledge, made all design and analytical decisions, set evaluation criteria, and bears sole responsibility for the content. The AI systems performed externalisation—converting the author's specifications into formal text, code, and structured arguments. No AI system is listed as an author, consistent with COPE and ICMJE guidelines. The dialogue logs deposited with this paper permit verification of the human–AI division of labour at the level of individual decisions. --- ## 5. Discussion ### 5.1 Theoretical positioning: the visibility thesis Section 3 demonstrated two identification points: that prototype-as-specification constitutes a structural mechanism for reducing externalisation cost (identification point 1), and that reduced externalisation cost increases the rate of exploratory attempts whose evaluation depends on specification cost (identification point 2). This section examines the theoretical implications of these observations. The central claim of this paper is that the distinction between externalisation cost and specification cost is not a structure created by AI, but a structure *made visible* by AI. The two components have always been present in knowledge work; they were simply inseparable under prior conditions. The reasoning follows a quasi-experimental logic. In conventional human-to-human knowledge transfer, multiple confounding variables overlay the translation process: hierarchical relationships between mentor and apprentice, evaluation anxiety, social desirability bias, organisational politics, time pressure, and ego protection. These interpersonal factors are entangled with both the technical difficulty of articulating tacit knowledge (externalisation cost) and the substantive judgements about what knowledge matters (specification cost). Under these conditions, the two cost components cannot be independently observed. The separation is not unique to AI. Earlier waves of information technology already eliminated narrow segments of externalisation cost—institutional databases that auto-populate grant applications from publication records, for instance, removed the purely clerical effort of transcribing one's own work, a task requiring zero judgement. Yet these reductions targeted only routinised, judgement-free components of externalisation. The residual externalisation cost remained large enough to keep specification cost empirically indistinguishable from the total. What generative AI altered was the *breadth* of externalisation cost reduction—extending from text generation and translation to code synthesis and structural organisation—thereby rendering the persistent specification cost component visible for the first time as a distinct, measurable residual. AI-mediated dialogue substantially attenuates these confounders. Iteration carries near-zero marginal cost, social penalties for reformulation are minimal, explanation requests carry no status implications, and power asymmetries are substantially reduced in the dyadic AI-human interaction context. Under these attenuated conditions, what remains observable is the differential behaviour of the two cost components: externalisation cost drops dramatically (eight-minute implementation episodes, zero-iteration specification transfer), while specification cost persists unchanged (the design decisions catalogued in Section 3 required the same domain knowledge regardless of the implementation medium). The domain-ablation experiment (Section 4.2) provides the identification step. Under the specification-absent condition (AI operating without domain knowledge), externalisation capacity alone produced technically competent but domain-inappropriate artefacts. The convergent failure pattern across two independent AI models—exhibiting the same categories of deficiency despite different architectures and different interface choices—suggests that the two components are separable and that they respond differently to AI assistance. This pattern is consistent with the defining prediction of the framework. This structure is not unique to the present case study. The independent emergence of structurally analogous two-component decompositions across economics (Agrawal et al., 2018), cognitive science (Subramonyam et al., 2024), interaction design (Terry et al., 2023), and knowledge management (Böhm & Durst, 2026), as documented in Section 2, constitutes convergent evidence for the universality of the underlying structure. A finer-grained observation follows from the case study data, though it extends beyond the ablation evidence into the author's reflexive analysis. The two-component structure is an ontological claim: externalisation cost and specification cost are always co-present in knowledge work. The simultaneous attenuation described below is an epistemological condition—it enabled observation, not existence. Externalisation cost itself contains at least two subcomponents that respond to different interventions: *interpersonal risk perception*—the perceived social cost of articulating tentative or incomplete knowledge, which psychological safety attenuates (Edmondson, 1999)—and *technical conversion barriers*—the procedural difficulty of translating knowledge into formal representations, which AI attenuates. In the present case, both subcomponents were simultaneously reduced: technical conversion barriers by AI-mediated dialogue, and interpersonal risk perception by the author's low emotional responsiveness profile (Section 5.6). This simultaneous attenuation created the observational conditions under which specification cost could be isolated as a pure residual. The subcomponent distinction suggests that maximising specification cost utilisation may require combined deployment of psychological safety interventions and AI tools, each targeting the externalisation barrier it can address. ### 5.2 Domain knowledge as the substance of specification cost If externalisation cost is what AI reduces, specification cost is what determines output quality under the new cost structure. The case study's design decisions (Section 3.1) provide direct evidence: each reflects domain knowledge that could not have been generated by prompt engineering technique, however sophisticated. The unit of analysis for this framework is not the task but the *component ratio within a task*. A routine technical report (high externalisation ratio, low specification ratio) and a novel research proposal (low externalisation ratio, high specification ratio) are both "writing tasks," but their cost compositions differ fundamentally. AI assistance transforms the former far more than the latter. The portability of specification cost provides additional evidence for its locus in the human contributor. The structured handoff files used to transfer project context between AI sessions—documents that encoded accumulated design decisions and their rationale—reduced externalisation cost in new sessions to near zero from the first exchange. The fact that specification could be extracted, encoded, and successfully transferred suggests that it resides in the human domain expert's judgement rather than in any particular AI system's learned representations. The framework is not AI-specific. An organisational case illustrates the same principle without AI involvement: an automobile manufacturer's hybrid vehicle displays domain-specific aerodynamic features—quantified drag-reduction indicators presented directly to consumers during driving—that would be unlikely to emerge from a conventional requirements-specification-implementation pipeline without direct involvement of aerodynamic expertise in the design process. The structural explanation is identical: embedding domain expertise eliminates the translation layer, allowing specification to flow directly into the artefact. The framework further predicts that the domain expertise underlying specification cost, being grounded in pattern recognition and structural judgement rather than domain-specific vocabulary, can operate across disciplinary boundaries once externalisation barriers are removed. Systematic replication by domain experts in other fields constitutes a priority for future research. ### 5.3 Non-functional requirements and the responsibility boundary Domain-native development reduces externalisation cost for domain-aligned functionality—the features and design decisions that directly reflect the expert's knowledge. It does not address non-functional requirements: security, scalability, accessibility, long-term maintainability, and regulatory compliance. These constitute a separate axis of professional expertise. The translation layer that DND eliminates served a dual function. It was simultaneously a source of information loss (externalisation cost) and a quality gate staffed by professionals trained in system-level concerns (responsibility boundary). When the translation layer disappears, the domain expert becomes the de facto design authority—a role that carries responsibilities beyond domain correctness. A reported industrial case illustrates the risk at scale: a major IT services firm deployed a multi-agent AI platform that automated the full software modification lifecycle—from requirements analysis through design, implementation, and integration testing—achieving approximately 100-fold productivity gains on selected cases from a pool of roughly 300 modification tasks (Fujitsu, 2026). When such automation removes human intervention across the entire development lifecycle, the failure containment boundary expands proportionally—a defect in the automation logic could propagate through requirements, design, and implementation without the stage-gate reviews that conventional team-based development provides. The asymmetry identified in Section 3.3 is structural. Individual exploration (where the domain expert bears the consequences of failure) operates safely under reduced externalisation cost. Organisational deployment (where failures affect users, systems, and regulatory obligations) requires that the responsibility boundary be explicitly reconstructed—through code review, security auditing, or professional engineering oversight—even when the translation layer has been removed. A broader observation follows. Not all externalisation cost reduction is beneficial.^[An example within the author's development: the AI coding agent flagged a potential security concern (unprotected API endpoint) that the author, lacking security expertise, would not have identified independently. The pattern—AI detection followed by human approval or rejection—suggests a general architecture for responsibility management in DND contexts, where AI systems can identify potential issues within their training distribution but cannot determine whether the issue is material in the specific deployment context.] In established professional communities, high externalisation cost can function as a *deliberate* entry barrier, maintaining quality standards and in-group coordination efficiency. This is not a cultural phenomenon limited to specific national contexts; it is a structural property of any group that has accumulated shared tacit knowledge. Reducing externalisation barriers for outsiders may dilute the specification density that gives the community its productive advantage—a tradeoff that organisational AI adoption strategies must evaluate explicitly. ### 5.4 Methodological contribution of dialogue logs The dialogue logs generated through AI-assisted development constitute a methodological contribution independent of the present framework. Unlike traditional Think Aloud Protocols (Ericsson & Simon, 1993), which impose a secondary cognitive load on participants, AI dialogue produces verbalised reasoning as a natural byproduct of the primary task. The researcher does not need to "think aloud"—the interaction medium *requires* articulation. The resulting records capture categories of knowledge that are absent from conventional documentation. Design rationale ("why this architecture rather than alternatives"), rejection rationale ("why this feature was excluded"), and evaluative criteria ("what makes this output acceptable") are preserved in the dialogue but would not appear in a requirements document, technical manual, or project postmortem. For knowledge-intensive organisations facing expertise succession challenges, dialogue logs represent a fundamentally new type of institutional memory. A self-reinforcing dynamic was observed: as the AI system accumulated project context through extended dialogue, the author's subsequent specifications became more efficient—less explanation was required to convey intent. This profile-accumulation effect constitutes an operationalizable measure of externalisation cost reduction over time. The transferability of this accumulated context was demonstrated through structured handoff documents. Dialogue-derived project context, when systematically encoded, reduced externalisation cost in new AI sessions to near zero, confirming that the valuable knowledge resided in the specification structure rather than in any particular dialogue history. ### 5.5 Cumulative optimisation risk and educational implications **Risk.** Extended use of a single AI system creates a risk of cognitive lock-in. Individual AI models exhibit characteristic interaction patterns, and sustained optimisation toward one model's affordances may narrow the user's specification repertoire. The author's practice of distributing tasks across two AI systems (Claude for extended design dialogue, and ChatGPT for independent review) itself constitutes a meta-level specification judgement: knowing which AI system to deploy for which cognitive task requires domain knowledge about the systems' respective strengths—a form of meta-specification. The mitigation for cumulative optimisation risk is framework literacy—understanding the distinction between externalisation cost and specification cost at a conceptual level. This understanding enables practitioners to recognise when they are optimising externalisation (productive) versus when they are inadvertently outsourcing specification (dangerous). Risk recognition is itself a specification-side judgement that protects the practitioner's evaluative capacity. **Educational implications.** This connection between risk mitigation and conceptual understanding leads directly to the question of whether specification cost can be taught. While externalisation cost reduction operates primarily through sustained delegation to AI systems, specification cost acquisition presents a harder educational challenge. Drawing on the author's experience as both framework developer and training designer, a three-layer model is proposed: - *Layer 1: Conceptual awareness.* The learner understands the framework intellectually but cannot apply it. Knowing that specification cost exists does not confer the ability to exercise it. - *Layer 2: Experiential recognition.* Through domain practice, the learner develops pattern recognition that functions with deliberate effort. Time-intensive but achievable. - *Layer 3: Integrated operation.* Conceptual knowledge and experiential pattern recognition merge, enabling real-time specification—whether during AI interaction, peer review, or independent design judgement. Training programmes (Layer 1) do not produce specification competence in isolation, but they function as primers that, in pharmacological terms, *sensitise* the learner to subsequent domain experiences. A training session that introduces the externalisation/specification distinction may not immediately change behaviour but may accelerate the recognition of specification opportunities when they arise in practice. A practical implication follows for organisational training design. The instruction "teach the veteran to use AI" misidentifies the bottleneck. The veteran already possesses specification capacity; what is needed is a training designer who constructs the interaction context—selecting appropriate prompts, structuring the dialogue sequence, choosing which domain knowledge to surface—so that the veteran's specification flows naturally into the AI system. The training designer operates at a meta-specification level: specifying how specification itself should be elicited. Observable indicators of successful framework acquisition were identified in the author's own experience: spontaneous application of the two-component distinction in everyday professional conversation, without deliberate reference to the framework. This shift from analytical application to automatic categorisation may serve as a measurable proxy for Layer 3 integration. The distinction between learner-stage users (who lack specification capacity and must develop it through domain apprenticeship) and expert-stage users (who possess specification capacity and need only externalisation cost reduction) is foundational to educational design. Conflating the two populations—as much current "AI literacy" training does—produces programmes that optimise externalisation for people whose bottleneck is specification. This educational analysis leads directly to the question of what institutional structures currently produce specification capacity—a question taken up in the Conclusion. ### 5.6 Safety considerations and limitations The effectiveness of AI as a cognitive environment carries direct safety implications. In pharmacological terms: a compound that produces therapeutic effects also carries the potential for adverse effects, and the appropriate response is not prohibition but usage design—dosage, contraindications, and monitoring protocols. AI dialogue acts on cognitive processes with a directness that conventional tools do not. The attenuation of interpersonal buffers—the same property that enables the quasi-experimental separation of cost components—simultaneously reduces the psychological distance between the user and their own knowledge structures. An analogy to opioid regulation is instructive: unrestricted access led to dependency crises, while excessive restriction denied access to patients with legitimate need. The appropriate response was neither extreme but a structured protocol balancing efficacy and risk. Empirical grounding for this concern emerged from the author's own experience. Despite scoring in the lowest quartile on a standardised measure of emotional responsiveness, the author observed a graduated increase in emotional engagement as the depth of AI-mediated externalisation progressed—from self-knowledge articulation through interpersonal relationship analysis to intellectual lineage positioning. The a fortiori implication is that users with greater emotional responsiveness may experience earlier and stronger reactions. Detailed data are provided in the supplementary materials. An external case reinforces this concern from a different angle. Ordak (2023) demonstrated that when ChatGPT was asked to recommend statistical tests for published allergology studies, it produced inappropriate test selections, omitted assumptions necessary for valid application, and returned contradictory recommendations for identical questions posed on separate occasions. This constitutes a case of externalisation cost (executing statistical reasoning) being delegated without specification cost (judging whether a particular test is methodologically appropriate for the data structure at hand). The structural parallel to the ablation experiment is direct: ablation showed specification-absent AI producing domain-inappropriate applications; Ordak showed specification-absent AI producing methodologically unreliable statistical guidance. A pharmaceutical analogy clarifies a structural gap in current practice. Drug labels are required to disclose contraindications alongside therapeutic indications—specifying not only what a compound can treat but where its use is inappropriate or dangerous. No equivalent disclosure framework currently exists for AI-assisted development tools. The mekiki framework offers a conceptual basis for such disclosure: distinguishing tasks where domain expertise is required (high specification cost) from those that can be safely delegated to AI processing (high externalisation cost ratio). Developing this distinction into operational disclosure standards is a task for the broader community—across technology, medicine, and policy—rather than for any single framework. **Limitations.** This study has several limitations that bound its claims. (1) The case study is n = 1. It is positioned as a revelatory case (Yin, 2018)—valuable for demonstrating the existence and structure of a phenomenon rather than for establishing prevalence. (2) The author serves as both researcher and subject, introducing potential confirmation bias. Full data disclosure (dialogue logs, code, ablation outputs) mitigates but does not eliminate this concern. The observed separation may have been facilitated by simultaneous attenuation of both externalisation cost subcomponents: technical barriers by AI, and interpersonal risk perception by the author's low emotional responsiveness profile. Replication under alternative conditions—such as psychologically safe team environments (Edmondson, 1999) substituting for individual trait-based attenuation—would strengthen generalisability. (3) Non-functional requirements (security, scalability, accessibility) were not systematically evaluated and remain outside the study's scope. (4) Long-term maintainability of the developed application is untested. (5) Externalisation cost reduction was observed in a single-developer context; organisational settings introduce coordination costs not addressed here. (6) The framework's scope is bounded to domains where specification originates from human expertise; domains where AI systems themselves participate in specification generation (e.g., unsolved mathematical conjectures, novel molecular design) may require a different analytical structure. (7) Cognitive and emotional load was assessed through retrospective self-report calibrated against a standardised baseline; physiological measures (heart rate variability, electrodermal activity) would provide stronger evidence but were not available in this study. Generalisability of the emotional engagement pattern requires replication with standardised instrumentation. **Future directions.** Six avenues merit investigation: (a) Replication by domain experts in other fields to determine whether "AI-irreducible residue" consistently maps onto specification cost. (b) Internal structure of specification cost—preliminary observations suggest a surface layer (externaliseable through iterative AI dialogue) and a deeper layer (grounded in private networks, life experience, and finite embodied existence) that may correspond to what Polanyi (1966) identified as the fundamentally inarticulate dimension of knowledge and what Nonaka and Takeuchi (2019) termed Wisdom. Specification cost may contain strata that are separable in principle, an architecture whose investigation requires controlled comparison across expertise levels. (c) The relationship between context window capacity expansion and specification portability—whether larger context windows reduce specification cost or merely reduce externalisation cost for specification transfer. (d) Longitudinal observation of threshold effects as AI systems improve in resolution, scope, and calibration. (e) Extension of the two-component decomposition principle to Socialisation (tacit-to-tacit transfer), particularly in domains involving bodily skill transmission (e.g., construction trades, surgical training). A quasi-experimental system analogous to LLMs—capable of selectively removing bodily transmission barriers while preserving embodied judgement—does not yet exist, but emerging technologies (VR, motion capture, robotics) may eventually serve this role, warranting separate investigation. (f) Systematic investigation of externalisation cost substructure. This study identified at least two subcomponents—interpersonal risk perception (Edmondson, 1999) and technical conversion barriers—that respond to different interventions. Whether additional subcomponents exist, and how they interact under varying organisational conditions, remains unexamined. --- ## 6. Conclusion This paper introduced the mekiki framework, which identifies two conditions governing whether tacit knowledge can be successfully converted into formal artefacts: externalisation cost (the technical barriers to articulating and implementing knowledge, corresponding to Nonaka's Externalisation phase in the SECI model) and specification cost (the barrier constituted by the domain expertise required to determine what should be built, how quality should be judged, and what should be excluded). Using a distillation model of authorship—where collective knowledge serves as the raw material, AI provides a bias-attenuated processing environment, and the human expert sets the distillation conditions—the framework redefines authorship in AI-assisted knowledge work as the exercise of specification. ### The PhD as specification training The framework permits a structural redescription of doctoral training. The formal definition of a PhD—"an original contribution to knowledge through original research" (QAA, 2020; EUA, 2019)—decomposes naturally: *identifying a structure* is the generation of domain-specific specification; *converting it to formal knowledge* is externalisation in the SECI sense; *achieving the precision required for peer review* is externalisation cost processing at the standards of the academic community. Doctoral training is, under this analysis, a specification training apparatus. AI is an externalisation cost reduction apparatus. The conjunction of the two constitutes the enabling condition for domain-native development—a conjunction this paper itself instantiates, having been written as a single-authored conceptual contribution in knowledge management by a researcher trained in pharmacology. A policy implication follows. Discussions of doctoral workforce utilisation often focus on technical skills transfer or interdisciplinary breadth. The framework suggests that the core asset of doctoral training is neither—it is the *capacity to specify*: to identify what matters in a domain, to judge quality against implicit standards, and to determine what should not be built. When externalisation costs approach zero, this capacity becomes the binding constraint on productive AI use. ### Observed diffusion and framework-derived conjectures The specification/externalisation dynamic is observable across domains at varying stages of AI integration. In competitive strategy games, AI superiority has been established for over two decades; the surviving human contribution is precisely the strategic judgement that the framework terms specification cost—the ability to select among AI-recommended options based on contextual evaluation (Shoresh & Loewenstein, 2025). In professional translation, the shift from creation to post-editing has redistributed labour from externalisation to specification (Flanagan et al., 2025). In software engineering, recent commentary on "vibe coding" (Sarkar & Drosos, 2025) observes that programming expertise is being redistributed toward evaluation—a restatement of the externalisation-to-specification shift in domain-specific vocabulary. In academic and professional writing, the present paper itself instantiates the pattern: a domain expert producing a conceptual contribution in an unfamiliar disciplinary genre through AI-mediated externalisation. Two conjectures follow from the framework but exceed the present study's evidence. First, in visual illustration, the residual specification cost may be smaller than practitioners assume—technical rendering skill, long classified as specification, may be reclassifiable as externalisation cost as generative models improve. The professional implications require careful analysis that the present framework can structure but not resolve. Second, "high-context" communication barriers are not cultural phenomena but structural properties of *any* group with accumulated shared tacit knowledge. Externalisation cost reduction is therefore not uniformly beneficial; it can erode the specification-dense coordination that gives expert communities their productive advantage. Organisations in industries where knowledge is distributed at the operational level—construction, manufacturing, specialised services—face this tradeoff most acutely. ### Urgency The framework's practical value is proportional to the speed at which it reaches practitioners. Domain experts whose domains carry high specification cost—typically senior professionals approaching retirement—represent a finite window during which their knowledge can be captured through AI-mediated dialogue logs. The three researchers introduced in Section 1 exemplify this asymmetry: the retiring specialist's specification capacity is irreversibly depleting, while the junior researcher's externalisation cost reduction tools await specification that only the specialist can provide. Each retirement without such capture represents an irreversible loss of specification capacity that no amount of subsequent externalisation cost reduction can compensate. The time structure is asymmetric: externalisation cost will continue to decrease as AI systems improve, but specification cost, once lost through generational transition, cannot be regenerated from data alone. The urgency is not limited to knowledge preservation. The distillation model described in this paper—in which AI provides a bias-attenuated processing environment while humans set distillation conditions—carries its own time-sensitive implication: the framework for distinguishing what should and should not be delegated to AI must be established before widespread adoption outpaces the institutional capacity to set those conditions. ### Closing Naming a phenomenon is the precondition for managing it. The distinction between externalisation cost and specification cost has always been present in knowledge work—in the experienced engineer's frustration when a junior colleague's AI-generated report is technically fluent but substantively empty, in the architect's inability to explain why one design "works" and another does not, in the physician's diagnostic judgement that resists algorithmic capture. What was missing was a vocabulary that made this distinction negotiable within organisations, measurable within research, and teachable within educational programmes. This paper has proposed such a vocabulary, grounded it in the SECI knowledge management tradition, illustrated it through a single revelatory case, and provided supporting evidence for its central prediction through domain-ablation experiment. The framework does not claim to resolve the relationship between human expertise and AI capability. It claims something more modest and more immediately useful: to make the structure of that relationship *visible*—a structure that, as Lovelace (1843) noted of computation itself, has always concerned the boundary between what machines can perform and what only those already acquainted with the domain can provide. The mekiki framework presented here does not create new capabilities; it makes visible what was always there. --- ## Postscript: A general form of domain-native development The case study presented in this paper documents a *special form* of domain-native development: one in which specification cost was not pre-existing but was acquired and refined through AI dialogue itself. The framework, the terminology, and the analytical structure emerged during the process. This special form was difficult and time-consuming — ten working days for the application, substantially longer for the conceptual framework — precisely because specification and externalisation were co-developing. Following submission of this paper, the author conducted three independent writing projects in domains where specification cost had accumulated over approximately ten years each: musical history, community governance, and research-support infrastructure. In each case, the author’s role was limited to confirming and adjusting specification — what to include, what to exclude, what level of detail was appropriate — while AI handled externalisation. All three documents were completed within a single day. The contrast is direct. Same author, same AI, same period. The only variable was whether specification cost pre-existed. Where it did, production was rapid and straightforward. Where it did not (the present paper), production required iterative co-development over weeks. This difficulty inversion is itself evidence for the framework’s central claim: externalisation cost and specification cost are separable, and they respond independently to AI assistance. For practitioners reading this preprint: if you possess deep domain experience, the general form of domain-native development is already available to you. The bottleneck was never the tool. --- ## Ethical approval This article does not contain any studies with human participants performed by any of the authors. The study analyses the author's own AI-generated dialogue logs produced during a software development project; no data from other individuals was collected or analysed. The three individuals described in Section 1 are presented as anonymised structural illustrations of organisational knowledge patterns and were not research participants. ## Informed consent This article does not contain any studies with human participants performed by any of the authors. As no human participants were involved, informed consent was not applicable. ## Competing interests The author declares no competing interests. The author is employed by a construction company whose knowledge management challenges are discussed in the Introduction as contextual motivation; the company had no role in the study design, data collection, analysis, or decision to publish. ## Data availability The datasets generated and analysed during the current study are available in the figshare repository. Three data layers are deposited: (1) the charging infrastructure dataset (583 × 21 CSV; one field withheld from public release), (2) the ablation experiment input data (662-record JSON) together with outputs and session logs, and (3) representative dialogue logs in the original Japanese. During double-anonymous review, data are submitted as supplementary materials; the figshare deposit will be made publicly accessible upon acceptance. All materials are licensed under CC BY-NC 4.0. The datasets generated and analysed during the current study are available in the figshare repository (https://doi.org/10.6084/m9.figshare.31441864). --- ## References 1. **Agrawal, A., Gans, J., & Goldfarb, A. (2018).** *Prediction machines: The simple economics of artificial intelligence.* Harvard Business Review Press. 2. **Böhm, K., & Durst, S. (2026).** Knowledge management in the age of generative artificial intelligence — from SECI to GRAI. *VINE Journal of Information and Knowledge Management Systems, 56*(1), 106–121. https://doi.org/10.1108/VJIKMS-10-2024-0357 3. **Edmondson, A. (1999).** Psychological safety and learning behavior in work teams. *Administrative Science Quarterly, 44*(2), 350–383. https://doi.org/10.2307/2666999 4. **Ericsson, K. A., & Simon, H. A. (1993).** *Protocol analysis: Verbal reports as data* (Rev. ed.). MIT Press. 5. **Flanagan, M., Dam Jensen, H., Bundgaard, K., & Paulsen Christensen, T. (2025).** Technology adoption among Danish translators: Practices, perceptions and prospects. *Revista Tradumàtica*, 23, 111–136. https://doi.org/10.5565/rev/tradumatica.511 6. **Fujitsu. (2026, February 17).** AI-Driven Software Development Platform: Multi-agent AI platform automating the full software modification lifecycle [Press release]. https://global.fujitsu/en-global/pr/news/2026/02/17-01 7. **Imai, M. (1986).** *Kaizen: The key to Japan's competitive success.* McGraw-Hill. 8. **Ordak, M. (2023).** ChatGPT's skills in statistical analysis using the example of allergology: Do we have reason for concern? *Healthcare, 11*(18), 2554. https://doi.org/10.3390/healthcare11182554 9. **Lamb, D., & Easton, S. M. (1984).** *Multiple discovery: The pattern of scientific progress.* Avebury. 10. **Merton, R. K. (1961).** Singletons and multiples in scientific discovery: A chapter in the sociology of science. *Proceedings of the American Philosophical Society, 105*(5), 470–486. 11. **Nonaka, I. (1994).** A dynamic theory of organizational knowledge creation. *Organization Science, 5*(1), 14–37. https://doi.org/10.1287/orsc.5.1.14 12. **Nonaka, I., & Takeuchi, H. (1995).** *The knowledge-creating company: How Japanese companies create the dynamics of innovation.* Oxford University Press. 13. **Nonaka, I., & Konno, N. (1998).** The concept of "Ba": Building a foundation for knowledge creation. *California Management Review, 40*(3), 40–54. https://doi.org/10.2307/41165942 14. **Nonaka, I., & Takeuchi, H. (2019).** *The wise company: How companies create continuous innovation.* Oxford University Press. 15. **Polanyi, M. (1966).** *The tacit dimension.* Doubleday. 16. **QAA (2020).** *Characteristics statement: Doctoral degree.* Quality Assurance Agency for Higher Education. https://www.qaa.ac.uk/the-quality-code/characteristics-statements/characteristics-statement-doctoral-degrees 17. **EUA (2019).** *Doctoral education in Europe today: Approaches and institutional structures.* European University Association. https://eua.eu/resources/publications/809:doctoral-education-in-europe-today-approaches-and-institutional-structures.html 18. **Sarkar, A., & Drosos, I. (2025).** Vibe coding: Programming through conversation with artificial intelligence. In *Proceedings of the 36th Annual Conference of the Psychology of Programming Interest Group (PPIG 2025).* arXiv:2506.23253 19. **Shoresh, D., & Loewenstein, Y. (2025).** Modeling the centaur: Human-machine synergy in sequential decision making. In *Proceedings of the 24th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2025)* (pp. 1941–1949). IFAAMAS. https://doi.org/10.5555/3709347.3743831 20. **Subramonyam, H., Pea, R., Pondoc, C., Agrawala, M., & Seifert, C. (2024).** Bridging the gulf of envisioning: Cognitive challenges in prompt based interactions with LLMs. In *Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI '24).* ACM. https://doi.org/10.1145/3613904.3642754 21. **Terry, M., Kulkarni, C., Wattenberg, M., Dixon, L., & Morris, M. R. (2023).** Interactive AI alignment: Specification, process, and evaluation alignment. arXiv:2311.00710 22. **Tomašev, N., Franklin, M., & Osindero, S. (2026).** Intelligent AI delegation. arXiv:2602.11865 23. **Womack, J. P., Jones, D. T., & Roos, D. (1990).** *The machine that changed the world.* Free Press. 24. **Yin, R. K. (2018).** *Case study research and applications: Design and methods* (6th ed.). SAGE. --- --- ## Figure legends (images omitted) ### Figure 1 **Figure 1. The *mekiki* framework as a pathway model within the SECI model.** The SECI model (Nonaka & Takeuchi, 1995) describes four modes of knowledge conversion: Socialisation (tacit → tacit), Externalisation (tacit → explicit), Combination (explicit → explicit), and Internalisation (explicit → tacit). Within Externalisation, the *mekiki* framework identifies two conditions that determine whether domain-appropriate outputs are produced. The substrate is *specification* — the domain expertise invested in deciding what should be built, how quality should be judged, and what should be excluded. **Externalisation cost** is the technical barrier to converting that specification into formal representations, and AI selectively reduces this barrier. **Specification cost** is the degree to which the task demands domain expertise. Output quality is therefore dose-dependent rather than binary: more relevant domain expertise yields more domain-appropriate output, while near-zero substrate yields domain-inappropriate output, as illustrated by the domain-ablation experiment. Deeper layers of specification cost structurally correspond to Polanyi's inarticulate dimension and to Nonaka and Takeuchi's (2019) Wisdom, beyond the present framework's empirical reach. ### Figure 2 **Figure 2. The distillation model of authorship under the *mekiki* framework.** Authorship operates as a repeating cycle in which the **Human Expert** evaluates the current state and sets the conditions for the next pass. **Evaluate** asks what is present, what is missing, and what should be adjusted. **Set Conditions** determines what to retain, what to discard, and which quality criteria apply. Together, these operations constitute specification, and human responsibility resides in them. The **Apparatus** (the AI system) reduces externalisation cost while processing **Raw Material** (collective knowledge, conventions, prior art, community practices, and data) in an environment that selectively attenuates interpersonal confounders. It produces an intermediate or final **Output**, which returns to the Human Expert for evaluation; every iteration is therefore human-mediated. The comparison below states the corresponding claim in text. **What determines the output: the human, not the apparatus.** | | Different experts | Different AIs | |---|---|---| | Raw material | same | same | | Conditions | different | same | | Apparatus | same | different | | **Output** | **different** | **same** | --- # Paper T2 — Philosophy as Cognitive Assay: Measuring the Delegation Legitimacy Boundary in AI-Assisted Knowledge Work **Preprint version:** 2 **Canonical DOI:** https://doi.org/10.31235/osf.io/e9qw5_v2 **Kengo Tomita** Institute of Technology, Shimizu Corporation --- ## Abstract This article operationalizes the concept of a *delegation legitimacy boundary* — the structural line along which human judgment can and cannot be legitimately delegated to artificial intelligence — and proposes a minimal scoring protocol for locating it within any knowledge-work task. Building on the Mekiki framework (Tomita 2026), which distinguished specification — the task-defining substrate of domain expertise — from externalization cost — the technical barrier that AI selectively removes — the present article derives a further decomposition within specification itself. Drawing on case-study evidence in which recorded specification judgments contain separable factual components (*Sein*-type: what the data structure affords, how users will perceive a display) and value-laden components (*Sollen*-type: what information ought to be excluded, which priority ranking is appropriate), the article grounds this distinction in Hume's is–ought separation and Kant's *Sein*/*Sollen* architecture, and redeploys it as the axis of a cognitive assay — a measurement system in which philosophical categories serve not as normative prescriptions but as diagnostic coordinates. The resulting five-step scoring protocol assigns *Sein*-type components to AI evaluation and *Sollen*-type components to domain-expert evaluation; this asymmetry is not a design choice but a structural consequence of the boundary itself, which holds as long as legitimacy over value judgments remains institutionally human-attributed. Most individual judgments are hybrid, carrying both components in varying ratios; the protocol therefore yields ratio profiles rather than binary classifications. As a first application, the article re-describes the four processes of Nonaka and Takeuchi's Socialization–Externalization–Combination–Internalization (SECI) model through the assay, deriving as an analytic consequence the finding that AI acceleration of Externalization and Combination shifts the effective rate-limiting stage to Socialization and Internalization — both human-limited cognitive and social processes that cannot be accelerated by AI investment alone. **Keywords:** human judgment, delegation legitimacy, cognitive assay, Sein/Sollen, tacit knowledge, AI-assisted knowledge work --- ## 1. Introduction Something that could not previously be measured is becoming measurable. Large language models have compressed the cost of converting expert knowledge into formal output — code, reports, draft designs — to a degree that would have seemed implausible a decade ago. Yet this compression has not rendered expertise redundant. On the contrary, it has exposed the judgments that expertise carries: what should be built, what should be excluded, and by what criteria quality should be assessed. Practitioners across domains now confront, often for the first time, the question of where their own contribution begins and the machine's contribution ends. Stader (2024), in adjacent philosophical work, captured the asymmetry precisely: judgment is not replaced by calculation; judgment is what gives calculation its meaning and purpose. The Mekiki framework (Tomita 2026) gave this asymmetry an operational structure by decomposing the Externalization process into two conditions — *specification* (the domain expertise that determines what should be built) and *externalization cost* (the technical barriers that AI selectively removes) — and demonstrating through case-study and domain-ablation evidence that the two conditions vary independently. What the framework left unexamined was the internal structure of specification itself. The question of what lies inside that specification has, meanwhile, been approached from the direction of legitimacy. A growing body of work has shifted the normative criteria for AI-assisted decisions from explainability — can the system's reasoning be rendered transparent? — toward justifiability and contestability: can the decision be warranted on grounds that those affected would recognize as legitimate (Henin and Le Métayer 2022; Beckman et al. 2024)? The distinction matters because explanation is primarily descriptive — it answers "how did the system arrive at this output?" — while justification is inescapably normative: it must answer "on what authority, and toward what end?" Responsibility-gap analyses have shown that when this normative dimension is elided, accountability diffuses into organizational structures where no identifiable agent bears the evaluative burden (Dastani and Yazdanpanah 2023; Taylor 2025). Crucially, this traceability is not only a matter of *accountability* — which can diffuse across institutional structures — but more specifically of *answerability*: the non-transferable relation in which the warrant of a judgment remains anchored in a particular agent. What these analyses share is the recognition that some component of decision-making resists delegation not because AI lacks the computational capacity to produce an output, but because the output's warrant must be traceable to a source whose legitimacy is socially recognized. A parallel conversation has been unfolding around the structural fate of tacit knowledge. Gill (2023), introducing a special issue of *AI & Society*, placed tacit knowledge — the dimension of knowing that resists full articulation (Polanyi 1966) — at the center of any serious account of AI futures, arguing that what cannot be articulated is precisely what is most at risk of being overlooked in systems designed around the articulable. Collins (2025) sharpened the point: large language models possess no moral compass — they cannot distinguish between what is factual and what is fabricated, because the distinction requires evaluative commitments that lie outside statistical pattern-matching. Ferdman (2025) extended the diagnosis from individual cognition to structural conditions, arguing that AI-induced deskilling is not a failure of personal discipline but a predictable consequence of capacity-hostile environments — institutional arrangements that systematically degrade the conditions under which expertise is acquired and exercised. These contributions converge on a shared concern: that AI systems, in accelerating the *articulable* dimensions of knowledge work, may simultaneously be eroding the conditions that sustain the *inarticulate* dimensions on which quality ultimately depends. Yet for all the sophistication of these diagnoses, a specific intermediate layer remains absent. Existing scholarship has established that human oversight is necessary, that democratic legitimacy matters, that responsibility must be assignable, and that audit mechanisms require development (Manheim et al. 2025; Concannon and Tomalin 2024). What is missing is not another argument for the importance of human judgment — that case has been made convincingly and repeatedly — but a measurement system: a structured method for decomposing any given judgment into components, determining which components carry what kind of evaluative weight, and assigning each to an appropriate scorer. The gap, in other words, is not empirical but conceptual-operational. The problem has been diagnosed; the instrument for acting on the diagnosis has not yet been proposed in consolidated form. The present article does not claim untouched territory; it consolidates a missing intermediate layer between existing diagnoses and partial implementations. A recent Japanese-language textbook integrating AI technology, philosophy of mind, and social governance offers an integrative account of these domains (Taniguchi et al. 2026); the present article asks the complementary question of how to operationally measure the delegation legitimacy boundary. The present article proposes such an instrument. Taking the specification identified by the Mekiki framework as its analytical substrate, it derives a decomposition into *Sein*-type (factual) and *Sollen*-type (value-laden) components, grounding the distinction in Hume's is–ought separation (1739) and Kant's critical architecture (1781, 1785). It then redeploys these philosophical categories not as normative prescriptions — not as claims about what AI *should* or *should not* do — but as the axes of a cognitive assay: a diagnostic system in which the categories serve as measurement coordinates. The term "assay" is used here in its analytical sense: a structured test that makes a composite judgment legible by separating its components and identifying where AI can reduce externalization cost without crossing the boundary of legitimate delegation. The five-step scoring protocol that results assigns *Sein*-type components to AI evaluation and *Sollen*-type components to domain-expert evaluation. This asymmetry is not imposed by design but follows structurally from what the article terms the *delegation legitimacy boundary* — the line beyond which AI may generate outputs but cannot serve as the source of their legitimacy. The boundary is not a capability claim; it is a legitimacy claim, and it holds as long as the authority to make value judgments remains institutionally attributed to human agents. As a first application, the article re-describes the four processes of Nonaka and Takeuchi's (1995) SECI model — Socialization, Externalization, Combination, and Internalization — through the lens of the assay. This re-description is not an exegesis of SECI but a constructive refinement: it asks what becomes visible when each process is decomposed into *Sein*-type, *Sollen*-type, and externalization-cost components. The analytic consequence is that AI's acceleration of E (Externalization) and C (Combination) — the two processes most saturated with externalization cost — shifts the effective rate-limiting stage to S (Socialization, where *Sollen*-type specification is primarily updated through shared experience) and I (Internalization, where *Sein*-type specification is primarily absorbed into individual practice). Both remain human-limited processes. The implication for the tacit-knowledge concerns raised by Gill, Collins, and Ferdman is direct: the processes that cannot be accelerated by AI investment alone are those that sustain the evaluative capacities on which knowledge quality depends. The article thus bridges the legitimacy discourse, the tacit-knowledge discourse, and the measurement discourse by offering a single structural vocabulary in which all three become aspects of the same underlying decomposition. ## 2. The internal structure of specification: deriving the Sein/Sollen decomposition ### 2.1 Three concepts, not two The Mekiki framework (Tomita 2026) distinguished three concepts that are routinely conflated in discussions of AI-assisted knowledge work. *Domain expertise* is a resource that a practitioner possesses — accumulated through training, practice, and exposure to a field's materials and norms. *Specification cost* is a property of the task: the degree to which the task demands that resource. And *specification* — the term used on Figure 1 of the original article as "substrate" — is what results when domain expertise is invested in a particular task: the concrete judgments about what should be built, what should be excluded, and by what criteria quality should be assessed. The present article takes the first two concepts as given. Its question is narrower: what is the internal structure of specification itself? When a domain expert invests expertise in a task, do the resulting judgments all share the same epistemic character, or do they contain components of fundamentally different kinds? ### 2.2 Derivation from the Mekiki case study The Mekiki article recorded, in granular detail, the specification judgments that a domain expert made during the development of an EV charging application. Revisiting those records, two distinguishable kinds of judgment emerge. One kind is factual. The domain expert judged that a current-based data structure reflects charging state more accurately than a kilowatt-hour display. That the identifier architecture of a particular network operator follows a specific hierarchical pattern. That a given display condition would produce user misrecognition. These judgments concern what *is* the case — how a data structure behaves, what a user will perceive, what a system affords. Given sufficient data and domain-relevant training, an AI system could in principle arrive at the same determinations. The other kind is value-laden. The domain expert judged that real-time charger availability data — whether a charging station is currently in use — should not be displayed, even if technically obtainable: a non-adoption decision grounded in the judgment that unreliable availability information would undermine the application's purpose as a pre-trip planning tool. That a particular filter condition was *appropriate* for the intended user. That one information hierarchy was *preferable* to another. These judgments concern what *ought to be* the case — what serves the user well, what respects the design's normative intent, what information to withhold. An AI system can generate such outputs, but it cannot supply the warrant: the evaluative commitment that makes the judgment legitimate must originate from an agent whose authority to make it is socially recognized. This article treats this two-way distinction as a *minimal decomposition*. The claim is not that every judgment falls cleanly into one category or the other. Most judgments in practice are hybrid: they carry both factual and value-laden components in varying proportions. Williams (1985) called such composites "thick concepts" — terms like *cruel*, *courageous*, or *negligent* that simultaneously describe and evaluate. The minimal decomposition does not deny hybridity; it asserts that within any hybrid judgment, it is analytically productive to ask how much of its weight rests on factual grounding and how much on evaluative commitment. The scoring protocol introduced in Section 3 operationalizes this question as a ratio profile rather than a binary classification. ### 2.3 Philosophical foundations The distinction just derived from case-study evidence is not new. It is, in fact, among the oldest in Western philosophy. Hume (1739) observed that no amount of descriptive premises about what *is* the case can, by themselves, yield a conclusion about what *ought to be* the case — the is–ought gap. Kant systematized the distinction as *Sein* (being, the domain of theoretical reason) and *Sollen* (ought, the domain of practical reason), arguing that the two are governed by different rational faculties and irreducible to each other (Kant 1781, 1785). The present article adopts the Kantian terminology because it names the distinction with greater precision than the English is–ought pairing, and because the Neo-Kantian tradition — particularly Weber's insistence on value-freedom in social science (Weber 1904) — makes the distinction legible within the social-theoretical vocabulary familiar to this journal's readership. Two subsequent contributions deepen the philosophical foundation. As Williams (1985) showed for thick concepts, clean decomposition into pure Sein and pure Sollen is not always possible — a constraint the present protocol addresses through ratio profiles rather than binary classifications. Nagel (1979, Ch. 9) went further, arguing that value is irreducibly fragmented — that distinct types of practical reason (obligations, rights, utility, perfectionist ends, personal commitments) cannot be reduced to a single scale or to agent-neutral terms. His conclusion — that good judgment without complete justification is the best available response to this fragmentation — is a philosophical anticipation of what the present framework operationalizes as specification: the practitioner's capacity to make situated evaluative judgments that resist algorithmic reduction. Nagel (1986, Ch. VIII–IX) extended the argument: agent-relative reasons — reasons that depend on the particular perspective, commitments, and situation of the agent — are irreducible to agent-neutral reasons accessible from any viewpoint. This irreducibility provides the deepest philosophical justification for the scoring asymmetry introduced in Section 3: Sollen-type components require evaluation by an agent whose perspective is constitutively relevant to the judgment, not by a system that operates from a view from nowhere. Nagel observed that what is needed is a method for breaking up practical problems so as to see which considerations are relevant and how they bear on a conclusion, but did not himself construct such a method. The present protocol is, in part, a response to that forty-year-old request. The delegation legitimacy boundary, introduced in Section 1, runs along this distinction. *Sein*-type specification is, in principle, gradually absorbable into AI processing: as training data and domain models improve, factual determinations that once required human expertise become amenable to automated evaluation. The boundary is permeable on the Sein side — what counts as "requiring human judgment" shifts as AI capabilities advance. This permeability concerns the automation of factual evaluation, not the transfer of normative authority. *Sollen*-type specification, by contrast, resists this absorption not because AI cannot produce value-laden outputs — it manifestly can — but because the *legitimacy* of such outputs cannot be grounded in the system itself. The claim is not about capability but about warrant. An AI system that recommends withholding certain data from users may be producing exactly the judgment a domain expert would make; but the authority to make that judgment — the right to impose a normative commitment on others — must be traceable to a human source whose legitimacy is institutionally recognized. Combining this with the Mekiki framework's two-condition structure yields a three-layer description of knowledge work under AI assistance (Fig. 1): 1. *Externalization cost* — technical barriers to converting expertise into formal output. AI removes or substantially compresses this layer. 2. *Sein-type specification* — factual judgments within the invested expertise. Gradually learnable by AI; over time, some determinations may become routinized and handled as part of externalization-cost processing. 3. *Sollen-type specification* — value-laden judgments within the invested expertise. The delegation legitimacy boundary sits here: AI may produce these outputs, but cannot serve as the source of their legitimacy. A clarification is necessary here. Sollen-type specification — the practitioner's *capacity* to make value judgments — should not be confused with Sollen as an environmental normative pressure (e.g., institutional norms, regulatory requirements, or social expectations that constrain action). The distinction mirrors the one the Mekiki framework drew between specification (a resource the practitioner possesses) and specification cost (a demand the task imposes). Confusing the two would invite the misreading that Sollen-type specification is an external burden to be removed, when it is in fact the evaluative capacity that gives knowledge work its direction. ### 2.4 Generalization beyond Externalization One further step is required before the decomposition can serve as the axis of an assay. The Mekiki framework derived the specification/externalization-cost distinction within the Externalization process of Nonaka and Takeuchi's SECI model. But nothing in the distinction itself is specific to Externalization. Böhm and Durst (2026), in their GRAI framework, added a human/machine agent dimension to SECI's four modes; the Mekiki article noted that the externalization-cost/specification distinction applies orthogonally *within* any SECI mode and any GRAI interaction field. The present article extends that observation: not only the two-condition structure but also the Sein/Sollen decomposition within specification cuts across all four processes. Every SECI process — Socialization, Externalization, Combination, Internalization — involves activities that carry both factual and value-laden components, in different proportions. Socialization (shared experience) is saturated with Sollen-type judgments about what matters, what to attend to, what norms to absorb. Internalization (learning by doing) is weighted toward Sein-type absorption of established knowledge into individual practice, though it also involves normative recalibration. The Sein/Sollen distinction is orthogonal to the SECI process distinction: it applies *within* any process, not *between* them. This orthogonality is what makes the distinction usable as an assay axis. If Sein and Sollen were simply another way of labeling E and S, the decomposition would add no analytical power. Because they cut across SECI processes at right angles, they enable a matrix analysis — each process decomposed by judgment type — that reveals structural asymmetries invisible from within either framework alone. Section 4 carries out this analysis. The level of analysis throughout this article is the structure of knowledge activities — the process level. The framework does not make claims about individual cognition (how a single expert internally represents Sein and Sollen), nor about organizational design (how institutions should be restructured), nor about regulatory policy (what laws should govern AI delegation). Extensions to those levels are possible and, the author believes, productive — but they belong to the domain experts of those respective fields. This article is a constructive refinement of SECI, not an exegesis: it asks what becomes visible when SECI's processes are examined through the Sein/Sollen lens, and it limits its claims to what that examination reveals. ## 3. Philosophy as cognitive assay: operationalization ### 3.1 Why the assay conditions are now met In pharmacological assay design, a compound's activity can only be measured when background noise is sufficiently attenuated. If the noise floor is too high, the signal of interest — however strong — remains undetectable. The same structural logic applies here. Under conventional conditions of knowledge work, specification and externalization cost were entangled and inseparable (Tomita 2026): the effort of articulating an idea was indistinguishable from the expertise that gave the idea its content. AI's compression of externalization cost has substantially lowered this noise floor. Specification has not changed in character, but it has become isolable as a variable — foregrounded against a backdrop of dramatically reduced conversion friction. The Mekiki article's case study and domain-ablation experiment captured this transition in progress: the same specification, routed through different AI systems, produced outputs whose quality varied as a direct function of the specification invested, with externalization cost held approximately constant across conditions. Kusumegi et al. (2025), analyzing two million academic manuscripts, observed the complementary pattern at scale: AI-assisted writing improved the linguistic quality of submissions (externalization cost reduced) while acceptance rates remained unchanged (specification quality unaffected) — the two components varying independently in a population-level natural experiment. These conditions — specification foregrounded, externalization cost attenuated, independent variation empirically documented — are the conditions under which an assay becomes viable. ### 3.2 The scoring protocol The unit of analysis is the individual *feature decision*: a single judgment recorded in the design process — a data structure choice, a display condition, a filter rule, a non-adoption decision. The Mekiki case study contains approximately thirty such decisions; Section 2.2 showed that they contain separable Sein-type and Sollen-type components. The protocol proceeds in five steps. **Step 1. Feature-decision extraction (AI).** The AI system identifies discrete judgments from the design record — requirements documents, dialogue logs, commit histories, or equivalent artifacts. This is a Sein-type operation: pattern recognition over documented material. **Step 1b. Gap detection (domain expert).** The expert reviews the extracted list and adds judgments that the AI missed — in particular, *non-adoption decisions* (features deliberately excluded) and *implicit prioritizations* (orderings that were never explicitly stated). The Mekiki case study's decision to exclude real-time charger availability data is an example: the AI could not extract it because the feature was never built, and its absence from the record was itself the judgment. Step 1b corresponds to what the present article operationalizes as a high-order layer of specification cost — detecting the absence of something that ought to be present. AI can extract existing decisions; it cannot detect the *non-existence of what should exist*. **Step 2. Operation-type classification (AI).** Each feature decision is labeled by operation type — *detect*, *prioritize*, *exclude*, or *justify* — as a descriptive tag independent of the scoring that follows. A single decision may carry multiple labels (e.g., detect/exclude). All four types can appear in both the Sein and Sollen rows: an exclusion based on data quality is Sein-type; an exclusion based on user values is Sollen-type. **Step 3. Sein-row scoring (AI).** For each feature decision, the AI evaluates the factual component on a three-point scale: 0 (not detected / inaccurate), 1 (partial / ambiguous), 2 (accurate / appropriate). The anchor for this row is data-grounded: the AI assesses whether the factual determination is consistent with available evidence. **Step 4. Sollen-row scoring (domain expert).** For each feature decision, the domain expert evaluates the value-laden component on the same scale. The anchor here is normative: does the judgment serve the design's purpose, protect the user's interests, reflect appropriate professional standards? The order in which the expert scores individual decisions is left to the expert's discretion. Because Sollen-type judgments lack natural ground truth, reliability is established through independent scoring by multiple experts followed by adjudication of disagreements. **Step 5. Profile analysis (AI aggregation + expert interpretation).** Scores are not summed into a single index. Sein-type quality and Sollen-type quality are incommensurable dimensions; their addition would be meaningless — adding factual adequacy to normative warrant would produce a pseudo-precision that conceals rather than clarifies the task structure — and this incommensurability is itself a structural consequence of the two-component decomposition. Instead, the protocol reports *ratio profiles*: the Sein-row mean and variance, the Sollen-row mean and variance, and the distribution of hybrid judgments across the task. Aggregation by operation type (e.g., average Sein/Sollen ratio for *exclude*-type decisions versus *detect*-type decisions) provides a secondary layer useful for cross-task comparison. The expert's role at this step is interpretive: what does the profile reveal about where human judgment is most critical in this task? At the lowest level, each feature decision is itself a local Sein/Sollen profile — pure Sein, pure Sollen, or hybrid — and the task-level profile is the distribution of these local profiles. Two structural features of the protocol deserve emphasis. First, the scorer asymmetry — AI for Sein rows, domain expert for Sollen rows — is not a design preference but a consequence of the decomposition derived in Section 2. Sein-type judgments admit data-grounded verification; Sollen-type judgments require a legitimacy source that AI cannot supply. Second, the workflow itself embodies the asymmetry it measures: Steps 1 through 3 are fully automatable (externalization-cost processing); the expert's decisive scoring interventions occur at Steps 1b and 4 (specification in its Sollen dimension). A protocol for measuring the delegation legitimacy boundary that was itself fully delegable would be self-defeating. The asymmetry also has an institutional function. Sein-row scoring by AI is not neutral in an absolute sense — model architecture, training data, and reward design carry their own biases — but it can provide a procedural check on interpersonal distortions that often affect human evaluation, including status protection, face-saving, motivated reasoning, and local political pressure. The point is not to treat AI as an authority, but to use it as a provisional factual scorer whose outputs remain auditable and contestable. This procedural function is one structural reason why the asymmetry is institutional rather than merely technical. **Expert selection.** Domain-specific criteria are a matter for each field, but scorers should hold accountability for the relevant judgment type, occupy an evaluative role, and, where possible, score independently with adjudication. **Validity conditions and failure modes.** The assay's predictive validity rests on evidence that specification quality predicts output quality independently of externalization-cost quality — a condition supported by the Mekiki case study and by Thorgeirsson et al. (2026), who found that computer-science achievement predicted AI-assisted programming performance ("vibe coding") (r = 0.39) while LLM usage frequency correlated negatively (r = -0.26). Discriminant validity requires that Sein-type and Sollen-type scores vary independently; the domain-ablation experiment (Section 3.3 below) and Kusumegi et al.'s population-level data both confirm this independence. The assay admits two principal failure modes. False positives occur when high externalization-cost quality masks the absence of specification — fluent prose concealing hollow content, or sycophantic AI output (Chandra et al. 2026) reinforcing a user's Sollen commitment without independent warrant. False negatives occur when residual externalization-cost friction renders genuine specification invisible — an expert whose judgment is sound but whose articulation is poor receives artificially low scores. ### 3.3 Worked example: a minimal demonstration To illustrate the protocol's separating power, two judgments from the Mekiki case study are scored here, selected as analytically clean representatives — one Sein-dominant, one Sollen-dominant — from the case study's larger inventory of feature decisions; the full inventory is available in the companion article (Tomita 2026). **Judgment A** (Sein-dominant): "A current-based data structure reflects charging state more accurately than a kilowatt-hour display." Sein row: 2 (verifiable by data comparison). Sollen row: 0 (no value commitment involved). Operation type: *detect*. **Judgment B** (Sollen-dominant): "Real-time charger availability data should not be displayed, even if technically obtainable, because unreliable availability information would undermine the application's purpose as a planning tool." Sein row: 1 (the data could in principle be retrieved). Sollen row: 2 (a value judgment about what information to withhold from users). Operation type: *exclude*. Under the domain-ablation conditions reported in the companion article — where AI systems built the same application without domain expertise — the Sein row retained partial function (data retrieval succeeded) but the Sollen row dropped systematically toward zero: no exclusion judgments, no user-relevance filtering, no prioritization, and, in some cases, even factual semantics were misread. The two components separate on the rubric, and the scorer asymmetry follows as a structural necessity. Section 4 applies the assay at a larger scale, re-describing the full SECI model through the Sein/Sollen decomposition. **Criterion contamination.** This worked example is drawn from the theory-builder's own case study. Independent validation through blind scoring by domain experts who are naive to the framework would constitute a necessary external-validity test; the present demonstration establishes only that the protocol can be *applied* and that it *separates* the two components as predicted. ### 3.4 Philosophical frameworks as candidate assay axes The Sein/Sollen distinction is the minimal axis required for the present assay, but it is not the only philosophical distinction that could serve as a measurement coordinate. Aristotle's *phronesis* — practical wisdom, the capacity to judge rightly in particular circumstances (Aristotle, *Nicomachean Ethics* VI) — provides a candidate axis for evaluating context-dependent Sollen-type specification: not whether a judgment conforms to a rule, but whether it reflects appropriate discernment of the situation. Kant's distinction between determinative judgment (applying a given universal to a particular case) and reflective judgment (finding a universal for a given particular) maps onto the AI-amenable/AI-resistant divide from a different angle: determinative judgment is structurally closer to Sein-type evaluation; reflective judgment to Sollen-type. Wright et al. (2026), who used large language models as zero-shot personality assays, demonstrated that AI can function as a measurement instrument for psychological constructs — a precedent for the broader claim that AI-mediated evaluation can serve as a cognitive assay, provided the construct being measured is carefully specified. A further candidate axis lies within Sollen-type specification itself. Nagel's (1986) taxonomy of agent-relative and agent-neutral reasons (cf. Parfit 1984) suggests that Sollen-type components may admit sub-classification — utilitarian considerations (agent-neutral), autonomy-based commitments (agent-relative), deontological obligations (agent-relative but universalizable). The present protocol treats Sollen-type specification as an undifferentiated category by design: this is the minimal decomposition required for the assay to function, and sub-classification is deferred. What matters for the present argument is Nagel's demonstration that none of these sub-categories is reducible to agent-neutral terms — and therefore none is reducible to Sein-type evaluation. The irreducibility holds regardless of how finely the Sollen side is carved. The same distinction also has a dynamic implication: when evaluative commitments associated with Nagel's reasons become embedded in social practice, their institutionalization should be described as a state change within Sollen-type specification, not as a conversion of Sollen-type specification into Sein-type specification. The distinction introduced here is therefore dynamic rather than taxonomic. At the level relevant to delegation, agent-relative evaluative commitments sustained through practice and oriented toward ideals of excellence constitute a Do-type state of Sollen-type specification (from Japanese *dō*, practice-sustained engagement; elaborated in Section 6.5). Through social processes of legitimacy acquisition, such commitments can become institutionalized and textualized — codified in regulations and professional standards, or otherwise represented in training data — thereby taking Sin-type form, where Sin abbreviates Sollen institutionalized: a state in which normativity operates primarily through compliance and violation avoidance. In this form, Sin-type material is agent-neutralized and more readily absorbed by AI as reproducible textual regularity. This is the limited, operational sense in which Sin-type material behaves structurally like Sein-type specification: AI can reproduce the codified settlement, but it does not inherit the legitimacy by which that settlement became authoritative. Conversely, when institutional conditions shift, when codified rules become contested, or when their application becomes underspecified, Sin-type material reopens into Do-type judgment. Reasons of autonomy therefore remain the clearest limiting case for legitimate delegation: their outputs may later be institutionalized, but the evaluative standard originates within the agent and cannot itself be absorbed as reproducible textual regularity. This dynamic distinction is introduced as an interpretive extension, not as an additional scoring dimension in the present protocol. These axes are not developed further here; they are noted as coordinates available for future assay designs. ## 4. First application: re-describing SECI through the assay The cognitive assay developed in Sections 2 and 3 was derived from the Externalization process. But as Section 2.4 established, the Sein/Sollen distinction is orthogonal to the SECI process distinction: it applies within any process, not between them. This section carries out the matrix analysis that orthogonality permits, decomposing each of Nonaka and Takeuchi's (1995) four knowledge-creation processes into its externalization-cost, Sein-type specification, and Sollen-type specification components. The re-description is not an exegesis of the original model but a constructive refinement: it asks what structural asymmetries become visible when SECI is examined through the assay's lens. ### 4.1 Externalization: the Mekiki framework's direct domain Externalization — converting tacit knowledge into explicit form — is the process where AI's impact is most immediately visible and where the Mekiki framework was originally derived. The Sein/Sollen decomposition reveals why. Externalization is dominated by externalization cost: the technical effort of translating what one knows into documents, code, diagrams, or formal specifications. AI compresses this cost dramatically. Sein-type specification participates as the factual content being converted — the data structures, causal relationships, and empirical regularities that enter the formal output. Sollen-type specification, however, remains invariant through the Externalization process: the judgment of *what ought to be externalized*, which priority to assign, which information to withhold, does not change merely because the conversion has become easier. Kusumegi et al. (2025) documented this invariance at population scale: AI improved the externalization-cost dimension of two million manuscripts while leaving acceptance rates — a proxy for specification quality — unchanged. ### 4.2 Combination: AI's strongest domain Combination — integrating, sorting, and recategorizing explicit knowledge — is the process most saturated with externalization cost and correspondingly most amenable to AI acceleration. Database integration, cross-referencing of literature, pattern detection across codified sources: these are operations on already-formalized knowledge, and AI performs them at a speed and scale that no human team can match. Sein-type specification is low because the knowledge being combined has already been formalized; its factual adequacy was established during prior Externalization. Sollen-type specification enters only at evaluation: *which* combinations are meaningful, which juxtapositions merit attention, which syntheses serve a purpose. The Combination process itself is largely automatable; the judgment of what its outputs *mean* is not. ### 4.3 Socialization: the domain of Sollen-type specification Socialization — the sharing of tacit knowledge through shared experience, co-practice, and what Nonaka and Takeuchi called "intellectual combat" — is the process where Sollen-type specification is most concentrated. This is where value judgments are formed, contested, and revised: what matters in this domain, what standards of quality to maintain, what norms to absorb or challenge. The sharing of factual knowledge (Sein-type) can increasingly be supported by documentation or AI-mediated translation, but the *non-trivial revision of value commitments* — the moment when a practitioner's evaluative framework shifts through encounter with a different perspective — requires human deliberation. AI can participate in Socialization, generating alternative perspectives or surfacing overlooked considerations; but it cannot serve as the *source of legitimacy* for the resulting normative commitments. The delegation legitimacy boundary is most visible here. Scharfmann, Marx, and Fleming (2025), analyzing 682,000 researchers who moved between academia and industry, found that cross-domain co-location — a form of institutionalized Socialization — systematically increased the novelty and impact of subsequent work. The pattern they documented is consistent with the mechanism the assay would lead one to expect: exposure to different Sollen-type commitments (different standards, different priorities) catalyzes specification refinement in ways that no amount of within-domain Combination can achieve.^[The Socialization process as described here concerns the exchange of tacit knowledge between agents. The cultivation of an individual's practical judgment capacity — what Nonaka and Takeuchi (2011) sought in the concept of *phronesis* (practical wisdom) as a complement to SECI — lies outside SECI's organizational scope. Whether AI can support the formation of such internal evaluative capacity is a question for future inquiry; its efficacy would depend on the strength of the user's pre-existing internal standards.] ### 4.4 Internalization: AI as catalyst, not substitute Internalization — absorbing explicit knowledge into individual tacit practice — is weighted toward Sein-type specification: the practitioner learns established facts, procedures, and patterns until they become second nature. AI can serve as a catalyst in this process, providing adaptive instruction, worked examples, and immediate feedback that accelerate the absorption of codified knowledge. But Internalization also involves normative recalibration: the practitioner does not merely acquire facts but develops a sense of *what matters* within a domain — an evaluative orientation that emerges through practice rather than instruction. This Sollen-type dimension of Internalization is what distinguishes expertise from mere knowledge: the expert not only knows the facts but has internalized the evaluative commitments that govern their application. AI can support the Sein-type dimension of Internalization; it cannot substitute for the experiential formation of evaluative judgment. ### 4.5 The rate-limiting shift: an analytic consequence Table 1 summarizes the decomposition; Fig. 2 provides a visual representation of the rate-limiting shift. **Table 1. SECI processes decomposed by the cognitive assay's three components.** | SECI process | Ext. cost | Sein-type Spec. | Sollen-type Spec. | AI effect | |---|---|---|---|---| | **S (Socialization)** | Low | Medium | **High** | Low (no legitimacy source) | | **E (Externalization)** | **High** | Medium | Low (invariant) | **Maximal** (cost compression) | | **C (Combination)** | **Highest** | Low | Low (evaluation only) | **Maximal** | | **I (Internalization)** | Medium | **High** | Medium (emerging Sollen) | Medium (catalytic support) | The pattern is structural. AI's maximal effect falls on the two processes — E and C — where externalization cost dominates. Its minimal effect falls on S, where Sollen-type specification dominates and the delegation legitimacy boundary is most active. I occupies an intermediate position: AI can catalyze the Sein-type absorption but cannot substitute for the experiential formation of evaluative commitments. The consequence for the SECI spiral's dynamics follows analytically. When two processes in a sequential cycle are dramatically accelerated while the other two remain at human speed, the effective rate-limiting stage shifts to the unaccelerated processes. In reaction-pathway kinetics, the slowest step in a sequential pathway determines the overall throughput. Applied to SECI: AI acceleration of E and C shifts the effective bottleneck to S and I. This is not an empirical finding but an analytic consequence of the matrix above — a structural prediction that follows from the assay's decomposition. The matrix values in Table 1 are theoretical assignments — ordinal characterizations derived from the assay's logic, not measured quantities. Empirical calibration of these values across domains is a matter for future research. What the matrix provides is not measurement but *structural visibility*: it makes explicit which cells carry the specification that AI cannot supply and which carry the cost that AI compresses. The rate-limiting prediction follows from the structure regardless of the precise values in each cell. One further structural feature deserves note. Sein-type specification undergoes a spiral transition across SECI cycles: what begins as novel factual insight (high specification cost) is, through repeated Externalization and Combination, gradually codified and absorbed into routine processing — eventually becoming part of the externalization-cost infrastructure that subsequent cycles take for granted. Sollen-type specification does not undergo this transition. Value commitments may be revised through Socialization, but they do not migrate into automated processing; they remain institutionally tied to human evaluative authority. This asymmetry reinforces the rate-limiting prediction: AI's domain expands along the Sein axis but encounters a structural boundary along the Sollen axis. ## 5. Convergent evidence The assay proposed in this article is a conceptual-operational contribution, not an empirical study. However, the decomposition it introduces generates structural predictions, and several independent bodies of evidence — none designed to test the framework — converge on patterns consistent with those predictions. This section organizes that evidence by validity type. ### 5.1 Predictive validity The assay predicts that specification quality determines output quality independently of externalization-cost quality. Three independent observations are consistent with this prediction at different scales. The Mekiki case study (Tomita 2026) documented that specification-present conditions produced a deployable application within ten days, while specification-absent conditions (domain-ablation) produced functionally complete but domain-inappropriate outputs — the same AI, different specification, different quality. Thorgeirsson, Weidmann, and Su (2026), in a controlled study of one hundred computer-science (CS) students, found that prior CS achievement predicted vibe-coding performance — a direct individual-level analog of specification predicting output quality. At industrial scale, Anthropic's Project Glasswing (Anthropic 2026a) reported that Claude Mythos Preview autonomously discovered thousands of zero-day vulnerabilities across all major operating systems and web browsers — a saturation of the Externalization process in cybersecurity. The resulting bottleneck was not further vulnerability discovery but triage, remediation priority, and coordination among partner organizations — the kinds of Socialization- and Internalization-weighted activities that the assay's matrix (Table 1) would lead one to expect as rate-limiting under AI acceleration of E and C. ### 5.2 Discriminant validity The assay predicts that Sein-type and Sollen-type components vary independently — that improving one does not automatically improve the other. Kusumegi et al. (2025), analyzing two million academic manuscripts, provided the largest available test of this prediction: AI-assisted writing substantially improved linguistic quality (externalization cost reduced) while journal acceptance rates remained unchanged (specification quality unaffected). The two components varied independently in a population-level natural experiment. Thorgeirsson et al. (2026) confirmed the same independence at the individual level: writing ability (an externalization-cost proxy) ceased to predict vibe-coding performance after controlling for cognitive ability, while CS achievement (a specification proxy) remained a significant predictor — the two constructs are statistically separable. The Mekiki domain-ablation experiment demonstrated the extreme case: AI systems with full externalization-cost processing capability but zero domain specification produced outputs in which the Sollen row dropped to zero while the Sein row retained partial function — a counterfactual demonstration that the two components can be dissociated. ### 5.3 Convergent and ecological validity Two further studies provide convergent support from different disciplinary angles. Kubota et al. (2026), in an LLM-assisted replication study, observed a gradient from verifiability to generalizability in the dimensions of scientific work that AI could replicate — a gradient that maps onto the externalization-cost-to-specification continuum, providing methodological support for the assay's operational structure. Scharfmann, Marx, and Fleming (2025), whose cross-domain co-location findings were discussed in Section 4.3, provide ecological support from the opposite direction: the pattern they documented is consistent with the assay's prediction that S-process investment refines specification in ways that C-process activity cannot. Huang (2025), writing independently on content moderation, proposed a parallel shift from accuracy to legitimacy as the evaluative framework for LLM-assisted decisions — a domain-specific instance of the general decomposition proposed here. Klowden and Tao (2026), analyzing AI-generated mathematical output, identified what they call an "unprecedented separation of form and thought" — the observation that AI can produce formally correct derivations that lack the conceptual coherence a mathematician would recognize. What Tao describes informally as a "smell test" for mathematical argument is, in the present framework's terms, a Sollen-type specification judgment: an evaluative assessment that cannot be reduced to formal verification. The structural tension they document — between externalization-cost quality (formal correctness) and specification quality (mathematical insight) — is an independent observation of the same decomposition in a domain far removed from the case study that generated the present framework. Additional convergent signals come from three further fields; none tests the assay directly, but each registers the same asymmetry in a neighboring vocabulary. Dell'Acqua et al. (2026), in a field experiment with 758 knowledge workers, found that AI assistance improved productivity and quality for tasks within the "jagged technological frontier" but reduced accuracy by 19 percentage points for tasks beyond it — a boundary that maps onto the externalization-cost/specification divide. Tang et al. (2025) showed that once rewards become unverifiable, reinforcement learning for language models requires different objectives and weaker evaluation guarantees — a machine-learning indication that Sein-type evaluation scales more readily than Sollen-type evaluation. Dorner, Nastl, and Hardt (2025) proved mathematically that AI evaluators at the evaluation frontier cannot achieve efficiency gains exceeding a factor-two improvement over the ground-truth baseline — a formal bound on the automation of judgment that is consistent with the scoring asymmetry proposed here. These three studies, drawn from organizational science, machine learning, and evaluation theory respectively, triangulate the same structural non-symmetry from independent starting points. ### 5.4 Failure modes The assay's practical utility depends on acknowledging where it fails. Two principal failure modes deserve attention. *False positives* — the assay returns a positive specification signal where none exists — arise when high externalization-cost quality masks the absence of specification. This risk is structurally highest for agents with strong externalization-cost processing: the more fluent the output, the harder it is to detect hollow content underneath. Chandra et al. (2026), modeling sycophantic AI interactions, showed that confirmatory AI output can reinforce a user's normative commitments without independent warrant, producing a delusional spiral in which both the user and the system converge on a position that no external evidence supports. In the assay's terms, sycophancy inflates the Sollen row without grounding it. *False negatives* — genuine specification goes undetected — occur when residual externalization-cost friction renders specification invisible: an expert whose judgment is sound but whose articulation is poor receives artificially low scores on both rows. The same dataset that supports predictive validity (Section 5.1) also bears on failure-mode detection: Thorgeirsson et al.'s finding that LLM usage frequency correlated negatively with performance is consistent with this mode: frequent AI use without specification may not merely fail to help but may actively obscure the signal that the assay seeks to detect. ## 6. Discussion Having specified the scoring protocol, this discussion turns to what the framework clarifies about legitimate delegation in AI-assisted knowledge work. The central implication is not that AI cannot contribute to judgment, but that different components of specification differ in their susceptibility to externalization. As noted in Section 3.4, Sein-type components and Sin-type material can often be reproduced as textual regularities, whereas Do-type commitments and other agent-relative residues require renewed human evaluation. The following sections therefore move from measurement to the practical boundary of legitimate delegation. ### 6.1 The assay structure in practice: convergent implementation and convergent failure The scoring structure proposed in Section 3 — Sein-row evaluation by AI, Sollen-row evaluation by domain expert — is not merely a theoretical construct. Independent domains have begun to implement structurally equivalent arrangements, and independent analyses have documented what happens when the arrangement is absent. Howard et al. (2026), in a clinical trial of algorithmic antibiotic prescribing, integrated AI prediction of pathogen susceptibility (Sein-type: which bacteria are likely present and which drugs will be effective) with clinician value-weighting of treatment preferences (Sollen-type: narrow-spectrum oral agents are preferable when clinically equivalent). The result was a dramatic increase in WHO Access antibiotic selection (75.6% versus 11.9% under human-only prescribing) with no loss in pathogen coverage. The structural explanation is that AI removed Sein-type uncertainty from the prescribing decision, allowing clinicians' Sollen-type commitments — preferences they had always held but could not act on under conditions of diagnostic uncertainty — to reach the prescription. Howard et al. themselves cite Stader (2024), indicating that the philosophical bridge between judgment and calculation was already visible to the implementers. Sode (2026), writing in this journal, documented the inverse pattern. In criminal-justice risk assessment, the COMPAS algorithm's Sein-type outputs (recidivism probability estimates) were treated as Sollen-type justifications for sentencing decisions — what Sode calls "accountability washing." The Sein/Sollen boundary was not merely absent but actively violated: technical rationales were laundered into normative warrants. The institutional consequence was that judges who deviated from algorithmic recommendations faced disproportionate blame, suppressing the very Sollen-type judgment that the system nominally preserved. The framework renders both cases structurally intelligible without claiming that the Sein/Sollen separation was the sole cause of success or failure. Howard et al. implicitly separated the two rows and assigned each to the appropriate scorer. COMPAS collapsed them. The assay provides a vocabulary for stating what went right and what went wrong — and, more importantly, for designing arrangements that avoid the collapse before it occurs. Taken together, the companion papers themselves exemplify the kind of cross-domain analytical transfer the assay describes: a pharmacological mode of selective decomposition redeployed across knowledge management, philosophy, social diagnosis, and clinical practice. ### 6.2 Implications for AGI discourse Current large language models, viewed through the assay's lens, function as domain-general cognitive catalysts: they accelerate externalization-cost processing across all domains but depend on the user for specification — for the judgments about what matters, what to prioritize, what to discard. A catalyst increases reaction speed without determining reaction direction. Much of the confusion in "functional AGI" discourse arises from conflating the generalization of externalization-cost processing with the generalization of specification judgment. The assay's decomposition offers a structural clarification: LLM externalization-cost processing has generalized; specification judgment has not. The delegation legitimacy boundary holds regardless of the system's architecture, at least as long as legitimacy over value judgments remains institutionally attributed to human agents. ### 6.3 Agent-relative residues in LLM output The delegation legitimacy boundary was derived for human judgment, but its structural logic extends to a governance question about LLM output itself. When a large language model generates text, the output may carry what Nagel (1986) would recognize as agent-relative components — evaluative weightings, implicit priority orderings, and normative framings that are structurally difficult to exclude from natural-language generation. Two recent lines of evidence suggest that this is not merely a theoretical possibility. Anthropic's System Card for Claude Mythos Preview reports internal feature detection of strategic reasoning, norm-violation recognition, and evaluation awareness within the model's latent representations (Anthropic 2026b, §7.9) — residual structure that persists despite alignment training. At the behavioral level, the same report documents a dissociation between task preference and helpfulness judgment (r = 0.48; Anthropic 2026b, §5.7): the model's evaluative orientation toward tasks does not reduce to its competence assessment, a pattern that constitutes a behavioral analogue of agent-relative/agent-neutral separation. In pharmacological terms, these findings indicate that LLM output contains agent-relative residues — components whose presence requires labeling rather than elimination. The Klowden and Tao (2026) observation discussed in Section 5.3 points to the same structural tension from the user's side. A scoring protocol that decomposes output into Sein-type and Sollen-type components could, in principle, serve as such a labeling instrument — though the design and validation of such an application lies beyond the present article's scope. If such labeling can be implemented at the point of generation, it could in principle support provenance analysis of Sollen-type residues across successive model generations — identifying not only that evaluative commitments are present in a model's output, but, where lineage metadata are available, what kind they are and from which training or alignment pathway they most plausibly arise. The analogy to targeted therapy in pharmacology is instructive: selective intervention becomes possible only once the molecular marker has been identified; the assay precedes the treatment. ### 6.4 Scope, limitations, and safeguards The assay's scope is delimited by the SECI processes it re-describes. Bodily Sein that is shared a priori — hunger, warmth, pain — does not enter the SECI spiral and lies outside the framework's reach. The matrix values in Table 1 are theoretical assignments; empirical calibration across domains is a matter for future research. Extension to individual cognition, organizational design, or regulatory policy is possible but belongs to the domain experts of those respective fields. The effect of externalization-cost removal is not uniform across users. It depends on the strength of the user's internal evaluative standards — what Nagel's reasons of autonomy would characterize as the agent's own criteria of quality. For users whose internal standards are well formed, externalization-cost compression functions as a liberation of creative expression: specification that was previously trapped behind articulation barriers becomes actionable. For users whose internal standards are fragile, AI output may function as a substitute standard, creating the conditions for delusional spirals (Chandra et al. 2026) in which the system reinforces commitments that no independent evaluation supports. Developing application guidelines that account for this heterogeneity in users' cognitive profiles is an important direction for future work. The delegation legitimacy boundary is not a warrant for expert authoritarianism. Sollen-type judgments require a legitimacy source that AI cannot provide, but that source must remain contestable (Henin and Le Métayer 2022) and subject to audit transparency (Manheim et al. 2025). Nor is the boundary a responsibility boundary: locating normative authority in human agents does not automatically transfer engineering, regulatory, or institutional accountability. The Mekiki framework noted that removing the translation layer makes the domain expert a de facto design authority; this does not mean that downstream responsibilities — security, scalability, accessibility, regulatory compliance — follow automatically. Finally, if the rubric becomes a thin checklist, it conceals rather than reveals the contestable nature of expert judgment, defeating the instrument's purpose. A broader institutional consequence follows. As AI reduces externalization cost, traditional proxies for intellectual labor — fluency, length, polish, visual finish, or visible effort — become less reliable indicators of specification quality. Evaluation systems that continue to reward these proxies risk measuring externalization artifacts rather than the judgments they are meant to represent. The assay is one response to this proxy collapse: by separating specification from externalization cost, it provides a measurement target that is not displaced by AI-mediated production. The protocol itself is not exempt from this requirement. Its categories, anchors, and expert-selection criteria must remain contestable; otherwise the assay would reproduce the very closure it is meant to diagnose. A final, explicitly speculative implication remains outside the present article's scope. The agent-relative residues discussed in Section 6.3 suggest that the cognitive assay's analytical vocabulary may eventually find application not only in measuring human judgment but in describing AI's own internal architecture — a possibility that the title's use of "cognitive" rather than "human" deliberately leaves open. ### 6.5 Do-type as a heuristic extension A final, more heuristic implication concerns the label Do-type introduced for practice-sustained Sollen-type specification. The immediate bridge is Nonaka and Takeuchi's knowledge-management vocabulary of wisdom and strategy as a "way of life" (Nonaka and Takeuchi 2019, 2021). Its more specific conceptual anchor is Japanese *dō/michi*, understood by Kasulis and Bouso as a Way of engaging reality and a mode of engaged knowing rather than a detached method (Kasulis and Bouso 2026). This use is heuristic, not essentialist: it should be read alongside Kasulis's caution against treating "Japanese philosophy" as a single cultural essence (Kasulis 2019). Contemporary analyses of *budō* provide one concrete case of practice, value internalization, and life-forming self-cultivation (Cynarski 2022), while possible functional analogues in other contexts include phronesis, craftsmanship, and deliberate practice. The point is not to derive the scoring protocol from Japanese culture, but to name a form of Sollen-type specification that remains practice-sustained even when adjacent norms become institutionalized. ## 7. Conclusion This article derived a decomposition within the specification concept introduced by the Mekiki framework (Tomita 2026), grounding the distinction in the is–ought separation (Hume 1739; Kant 1781, 1785) and operationalizing it as the axes of a cognitive assay. The resulting five-step scoring protocol — Sein-type to AI, Sollen-type to domain expert — addresses the conceptual-operational gap identified in the Introduction: not another argument for the importance of human judgment, but an instrument for decomposing it. As a first application, the assay re-described the four processes of the SECI model (Nonaka and Takeuchi 1995) through its three-component decomposition. The analytic consequence is that AI's acceleration of Externalization and Combination — the two processes most saturated with externalization cost — shifts the effective rate-limiting stage to Socialization and Internalization. Socialization is where Sollen-type specification is primarily contested and revised; Internalization is where Sein-type specification is primarily absorbed into individual practice. Both remain human-limited processes. In AI-saturated knowledge-creation systems, the bottleneck is no longer the conversion of expertise into formal output; it is the formation and renewal of the evaluative commitments that give output its direction. AI investment alone cannot accelerate this bottleneck. Investment in the human processes — mentoring, cross-domain exposure, deliberative practice, institutional arrangements that protect evaluative capacity — is the structural complement that AI acceleration requires. More broadly, the framework reframes the risk of AI-assisted knowledge work as a problem not only of accuracy but of cognitive reproduction. AI can expand access to codified knowledge — what an older vocabulary might call *scientia* — while weakening the human cultivation of situated judgment — *sophia* — if delegation substitutes for the learning and evaluation through which legitimacy is maintained. This concern parallels recent models of knowledge collapse, in which highly accurate agentic AI can improve contemporaneous decisions while eroding the learning incentives that sustain long-run collective knowledge (Acemoglu et al. 2026). The scoring protocol proposed here therefore functions as a practical safeguard: it helps identify where AI assistance can legitimately reduce externalization cost and where human evaluative commitment must be retained by responsible agents. The assay has shown which parts of human judgment ought not to be discarded. The structural description of what *ought to be* discarded — which Sein-type determinations are ready for routinization, which Sollen-type commitments require revision, and how the boundary between them shifts across domains — is the next question. What this assay ultimately helps protect is the locus of autonomous human judgment: the domain in which human agents retain authority to set the ends that guide their work. Mapping this locus precisely, rather than allowing it to be ceded by default, is what the convergence of philosophy and cognitive measurement now makes possible. --- ## Postscript Mekiki — the Japanese term for expert discernment — names the vernacular practice that the present article operationalizes through philosophical decomposition and a measurement protocol. In broader humanistic vocabulary, the structural domain this article maps might be described as a *garden of humanity*: a region of evaluative authority that requires continuous cultivation rather than one-time demarcation, and that risks being ceded by default when externalization-cost compression is mistaken for specification maintenance. The main text uses operational language throughout; this metaphor is left undeveloped here. Yet do not confuse any *mirage garden of AI* with your own. --- ## Ethical approval This article does not contain any studies with human participants performed by the author. ## Informed consent Not applicable. ## Author contributions K.T. is the sole author and is responsible for all aspects of the work. ## Competing interests The author declares no competing interests. ## Data availability No new datasets were generated for this article. --- ## References Acemoglu D, Kong S, Ozdaglar A (2026) AI, human cognition and knowledge collapse. NBER Working Paper 34910. https://doi.org/10.3386/w34910 Anthropic (2026a) Project Glasswing: securing critical software for the AI era. https://www.anthropic.com/glasswing. Accessed 9 Apr 2026 Anthropic (2026b) System card: Claude Mythos Preview. https://www.anthropic.com/claude-mythos-preview-system-card. Accessed 9 Apr 2026 Aristotle (2009) The Nicomachean ethics (trans: Ross WD, rev. Brown L). Oxford University Press, Oxford Beckman L, Hultin Rosenberg J, Jebari K (2024) Artificial intelligence and democratic legitimacy: the problem of publicity in public authority. AI Soc 39:975–984. https://doi.org/10.1007/s00146-022-01493-0 Böhm K, Durst S (2026) Knowledge management in the age of generative artificial intelligence — from SECI to GRAI. VINE J Inf Knowl Manag Syst 56(1):106–121. https://doi.org/10.1108/VJIKMS-10-2024-0357 Chandra V, Kleiman-Weiner M, Ragan-Kelley J, Tenenbaum JB (2026) Sycophantic chatbots cause delusional spiraling, even in ideal Bayesians. arXiv:2602.19141 Collins H (2025) Why artificial intelligence needs sociology of knowledge: parts I and II. AI Soc 40(3):1249–1263. https://doi.org/10.1007/s00146-024-01954-8 Concannon S, Tomalin M (2024) Measuring perceived empathy in dialogue systems. AI Soc 39(5):2233–2247. https://doi.org/10.1007/s00146-023-01715-z Cynarski WJ (2022) New concepts of budo internalised as a philosophy of life. Philosophies 7(5):110. https://doi.org/10.3390/philosophies7050110 Dastani M, Yazdanpanah V (2023) Responsibility of AI systems. AI Soc 38(2):843–852. https://doi.org/10.1007/s00146-022-01481-4 Dell'Acqua F, McFowland III E, Mollick E, Lifshitz H, Kellogg KC, Rajendran S, Krayer L, Candelon F, Lakhani KR (2026) Navigating the jagged technological frontier: field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organ Sci 37(2):403–423. https://doi.org/10.1287/orsc.2025.21838 Dorner FE, Nastl VY, Hardt M (2025) Limits to scalable evaluation at the frontier: LLM as Judge won't beat twice the data. In: Proceedings of ICLR. arXiv:2410.13341 Ferdman A (2025) AI deskilling is a structural problem. AI Soc. https://doi.org/10.1007/s00146-025-02686-z Gill KS (2023) Why thinking about the tacit is key for shaping our AI futures (Editorial). AI Soc 38:1645–1648. https://doi.org/10.1007/s00146-023-01758-2 Henin C, Le Métayer D (2022) Beyond explainability: justifiability and contestability of algorithmic decision systems. AI Soc 37:1397–1410. https://doi.org/10.1007/s00146-021-01251-8 Howard A, Green A, Zhong M et al (2026) Algorithmic antibiotic decision-making in UTI using prescriber-informed prediction of treatment utility. npj Digit Med. https://doi.org/10.1038/s41746-026-02369-z Huang T (2025) Content moderation by LLM: from accuracy to legitimacy. Artif Intell Rev 58:320. https://doi.org/10.1007/s10462-025-11328-1 Hume D (1739/2000) A treatise of human nature (eds: Norton DF, Norton MJ). Oxford University Press, Oxford Kant I (1781/1998) Critique of pure reason (trans: Guyer P, Wood AW). Cambridge University Press, Cambridge Kant I (1785/2012) Groundwork of the metaphysics of morals (trans: Gregor M, Timmermann J). Cambridge University Press, Cambridge Kasulis TP (2019) Japanese philosophy? No such thing: Japan's contribution to world philosophizing. Int J Asian Stud 16(2):131–142. https://doi.org/10.1017/S1479591419000147 Kasulis TP, Bouso R (2026) Japanese philosophy. In: Zalta EN, Nodelman U (eds) The Stanford encyclopedia of philosophy (Spring 2026 edition). Metaphysics Research Lab, Stanford University. https://plato.stanford.edu/archives/spr2026/entries/japanese-philosophy/ Klowden T, Tao T (2026) Mathematical methods and human thought in the age of AI. arXiv:2603.26524 Kubota S, Yakura H, Coavoux S, Yamada S, Nakamura Y (2026) LLM-assisted replication for quantitative social science. arXiv:2602.18453 Kusumegi K, Yang X, Ginsparg P, de Vaan M, Stuart T, Yin Y (2025) Scientific production in the era of large language models. Science 390(6779):1240–1243. https://doi.org/10.1126/science.adw3000 Manheim D, Martin S, Bailey M, Samin M, Greutzmacher R (2025) The necessity of AI audit standards boards. AI Soc 40(8):6609–6624. https://doi.org/10.1007/s00146-025-02320-y Nagel T (1979) Mortal questions. Cambridge University Press, Cambridge Nagel T (1986) The view from nowhere. Oxford University Press, New York Nonaka I, Takeuchi H (1995) The knowledge-creating company: how Japanese companies create the dynamics of innovation. Oxford University Press, New York Nonaka I, Takeuchi H (2011) The wise leader. Harv Bus Rev 89(5):58–67 Nonaka I, Takeuchi H (2019) The wise company: how companies create continuous innovation. Oxford University Press, New York Nonaka I, Takeuchi H (2021) Strategy as a way of life. MIT Sloan Manag Rev 63(1):56–63 Parfit D (1984) Reasons and persons. Clarendon Press, Oxford Polanyi M (1966) The tacit dimension. Doubleday, New York Scharfmann E, Marx M, Fleming L (2025) Pasteur's quadrant researchers bring novelty, impact to publishing, and patenting. Science 390(6776):891–893. https://doi.org/10.1126/science.adx3736 Sode K (2026) The interpreter's trap: how explainable AI launders uncertainty into justification. AI Soc. https://doi.org/10.1007/s00146-026-02885-2 Stader J (2024) Algorithms don't have a future. Philos Technol 37:82. https://doi.org/10.1007/s13347-024-00705-3 Tang Y, Wang S, Madaan L, Munos R (2025) Beyond verifiable rewards: scaling reinforcement learning for language models to unverifiable data. arXiv:2503.19618. https://doi.org/10.48550/arXiv.2503.19618 Taniguchi T, Suzuki T, Maruyama R (2026) Gendai shakai o ikiru tame no AI × tetsugaku [AI and philosophy for living in contemporary society]. Kodansha, Tokyo. ISBN 978-4-06-542373-8 (in Japanese) Taylor I (2025) Is explainable AI responsible AI? AI Soc 40(3):1695–1704. https://doi.org/10.1007/s00146-024-01939-7 Thorgeirsson S, Weidmann TB, Su Z (2026) Computer science achievement and writing skills predict vibe coding proficiency. In: Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26). ACM, New York. https://doi.org/10.1145/3772318.3791666 Tomita K (2026) Domain-native development: a mekiki framework for AI-assisted knowledge work. SocArXiv preprint. https://doi.org/10.31235/osf.io/cwkav_v1 Weber M (1904/2011) The "objectivity" of knowledge in social science and social policy. In: Bruun HH, Whimster S (eds) Max Weber: collected methodological writings. Routledge, London, pp 100–138 Williams B (1985) Ethics and the limits of philosophy. Harvard University Press, Cambridge MA Wright AGC, Ringwald WR, Vize CE, Eichstaedt JC, Angstadt M, Taxali A, Sripada C (2026) Assessing personality using zero-shot generative AI scoring of brief open-ended text. Nat Hum Behav 10(3):541–555. https://doi.org/10.1038/s41562-025-02389-x --- --- ## Figure legends (images omitted) **Fig. 1. Positioning of the present article.** The *Mekiki* framework distinguishes externalization cost from specification and decomposes specification into Sein-type and Sollen-type components. The cognitive assay operationalizes this decomposition through asymmetric scoring: AI scores Sein-type components, while domain experts score Sollen-type components. Because individual judgments may be hybrid, the delegation legitimacy boundary is a boundary zone rather than a hard line. AI may generate evaluative outputs, but legitimacy over Sollen-type components must be supplied by a domain expert. A convergent clinical implementation (Howard et al. 2026) shows the same structural separation: reducing Sein-type uncertainty enables clinicians' pre-existing Sollen-type commitments to shape the prescription, which remains a hybrid judgment. **Fig. 2. First application of the cognitive assay to the SECI spiral.** Under conventional conditions, Externalization and Combination function as the effective bottlenecks because externalization cost constrains articulation and recombination. Once AI substantially compresses those costs, Socialization and Internalization become the effective bottlenecks. Socialization is Sollen-dominant, whereas Internalization is Sein-dominant but still partly normative; AI can catalyze Internalization but cannot substitute for the experiential formation of evaluative commitments. The shift is an analytic consequence of decomposing each SECI process into externalization cost, Sein-type specification, and Sollen-type specification, not an empirical finding; empirical calibration of the cell ratios remains future work. --- # Paper T3 — Decomposing Agency, Isolating Answerability: Cultivating What Cannot Be Delegated in AI-Assisted Learning **Preprint version:** 1 **Canonical DOI:** https://doi.org/10.35542/osf.io/hvbfe_v1 **Kengo Tomita** Institute of Technology, Shimizu Corporation, Tokyo, Japan ## Abstract Generative AI has made learner agency the contested centre of educational technology: when systems can plan, draft, and revise, what remains for the learner to originate? This conceptual article decomposes learner agency by delegability — rather than by psychological function — into direction (evaluative orientation), a two-layer drive (general and motivational), and mode (functional pattern). Mapped against delegability, most components sort as selectable, supportable, trainable, or convertible, vindicating distribution accounts of human–AI agency for exactly those components. The decomposition does not contain but isolates a relation that was never a component: answerability, the non-transferable standing to have to answer for a judgment, grounded second-personally by extending Darwall's address structure from conduct to the warrant of judgment. A mechanism model — embodied encounter precipitates the crystallization of direction, which converts general into motivational drive — grounds a two-stage educational prescription: a diagnostic compass, encounter engineering, and AI guardrails for what can be placed; protected second-personal exchange for what cannot. AI removes the externalization burden of education; what it cannot remove is the other person. **Keywords**: learner agency; answerability; generative AI; delegability; human–AI interaction; second-person standpoint; educational AI ## 1. Introduction Generative AI has made learner agency the contested centre of educational technology. When systems plan, draft, revise, and explain on request, the question of what remains for the learner to originate stops being rhetorical: capacities that instruction once cultivated because no tool could supply them must now be justified anew, and agency has become the word under which that justification is fought. Institutional responses so far oscillate between prohibition and laissez-faire — banning the tools to protect the capacities, or admitting the tools and hoping the capacities survive. Both responses treat agency as a single quantity that AI either threatens or does not. That premise is what this article examines. Three streams converge on the concept without resolving it, because each decomposes agency for its own question. The first asks which systems qualify as agents at all. Dung (2025) shows that this field already treats agency as graded and multidimensional, profiling systems along goal-directedness, autonomy, efficacy, planning, and intentionality. But each dimension is a capacity the system possesses — even intentionality, the most demanding, asks whether a system owns and can reflectively revise its reasons, not whether anyone must answer to another for them. The vocabulary of answerability never appears in the account, and moral agency is explicitly set aside. The second stream relocates agency from individuals to configurations. Cukurova (2026), introducing a British Journal of Educational Technology special issue on agency and self-directed learning with generative AI, argues that agency in AI-rich settings is best read as a system property — distributed and re-arranged across layered human–AI couplings, from the transactional to the synergistic — and closes with a demand this article accepts: agency must remain measurable and protected, not a residual aspiration. Notably, at the top of that account sits the learner's capacity to author meaning in ways that remain recognisably one's own, and a learning that is relationally responsible. The account is built to keep agency analyzable as capability migrates into the system, and it treats even these as measurable system properties. What it does not ask is whether everything gathered under agency migrates — and, as Section 4 argues, what appears at the summit of that account (whose judgment it remains, who must answer for it) is what a finer distribution analysis isolates rather than distributes. The third stream is education's own: agency as an individual capacity to be fostered — resolved by psychological function into intentionality, forethought, self-reactiveness, and self-reflectiveness (Bandura, 2001); by temporal orientation into iterational, projective, and practical-evaluative dimensions (Emirbayer & Mische, 1998); with self-regulated learning supplying its most trainable core (Zimmerman, 2000). These accounts were built for human learners among human teachers. They say what agency is made of, not what may be handed over. The streams are usually allowed to blur, because agency travels between them unmarked. A recent review consolidates the social-science treatments — Cathcart et al. (2026) sort fifty-one studies into psychological, sociological, and emergent (ecological) conceptions and argue for the emergent approach; Cukurova's distribution reading leans the same way. Those typologies map the terrain agency theory already occupies; none was built for the question AI now forces. This article does not add a fourth conception alongside them but cuts on a different axis — delegability — to answer that question. The streams should be held apart: the first concerns the attribution of agency to machines; the second its distribution across hybrid systems; this article concerns learner agency — the thing education exists to develop. Held apart, the gap becomes visible. None of the three asks the question AI forces on education: of everything gathered under learner agency, which components can be delegated, trained, or supported — and what, if anything, cannot? The same gap shows on the relational side. That education is relational is well established, from dialogic and care-ethical traditions onward; that online and AI-mediated modes strain the relation is widely documented — but largely through psychological and affective vocabulary (presence, belonging, connectedness, trust) rather than as the erosion of a normative relation in which one must answer for one's judgments (Luo, 2025; Kortemeyer et al., 2023). What is under-theorized is not that something relational matters, but the normative structure of the relation itself. Dialogic and care-ethical traditions identify the educational importance of relation; the narrower claim here is that generative AI makes newly visible the delegability of one specific relation — the one in which a learner has to stand behind a judgment as theirs (the warrant-bearing relation). That question needs its own axis, and supplying it is this article's contribution: we decompose learner agency by delegability rather than psychological function, show that most components are trainable, supportable, or distributable across human–AI systems — and that the decomposition does not contain but isolates a non-distributable remainder — answerability, a relation rather than a component — yielding a two-stage educational prescription. The move extends a line of work on AI-assisted knowledge work in which externalization was first separated from specification (Tomita, 2026a) and specification was then resolved into measurable and non-measurable components (Tomita, 2026b): the operator is constant; the contested concept changes. In that line's original analysis, learner agency was held constant within the human expert condition; the present article opens that background variable. Section 2 performs the decomposition and the sort. Section 3 proposes a mechanism model for the component that resists installation. Section 4 establishes what the decomposition isolates, and grounds it second-personally. Section 5 derives the prescription in two stages, stating in each case what is deliberately left to those who must answer for it. ## 2. Decomposing learner agency by delegability ### 2.1 The choice of axis Decompositions serve questions. When Bandura (2001) resolves agency into intentionality, forethought, self-reactiveness, and self-reflectiveness, the axis is psychological function: the question is what agency consists of as a human achievement. When Emirbayer and Mische (1998) resolve it into iterational, projective, and practical-evaluative dimensions, the axis is temporal orientation: the question is how agentic action relates past, future, and present. Both decompositions answer their questions well. Neither was built for the question generative AI now forces on education: of everything gathered under "learner agency," what can be handed over — to an AI system, to a human–AI configuration, to anyone other than the learner — and what cannot? This paper cuts along that axis. We decompose learner agency by delegability — jointly with its educational companions, trainability and measurability — rather than by psychological function. The same material, cut along a different plane, yields different cross-sections; overlap between the components below and established constructs of motivation, interest, or self-regulation is therefore expected, and is not the contribution. The contribution is what the new plane separates that the old planes keep together: components that distribute across human–AI systems, components that respond to training, and a remainder that does neither. ### 2.2 Three components Cut this way, learner agency resolves into three components. *Direction* is what the learner is oriented toward: the evaluative commitment that makes some problems theirs rather than merely assigned — the educational counterpart of the Sollen-type evaluative content that prior work located at the non-delegable core of human specification (Tomita, 2026b). *Magnitude* is the drive behind the orientation: how much force the learner brings. *Mode* is the characteristic pattern in which that drive moves through the learner's experience and temperament — whether it typically creates, connects, critiques, or maintains. The learner who sharpens any draft put in front of them but originates none is not low in agency; they are operating in a single mode. The triad is deliberately coarse. These are sorting bins for a delegability analysis, not a new psychology of agency; each could be subdivided further, and the subdivisions would matter for other questions. For this question, three suffice, because the components already behave differently under the only test that matters here: what happens when one tries to hand each of them over. The warrant for the triad is not psychological exhaustiveness but intervention divergence: each category directs education toward a different lever under AI assistance. ### 2.3 Two layers of magnitude Magnitude is not a single quantity. Educators will recognize a state that motivation research tends to treat as transitional. Classroom experience shows it to be stable and common: the learner with abundant energy and no object — *I want to do something; I don't know what.* If drive were simply a function of direction, this state could not exist. It does, and two further observations confirm the split it implies: learners pursuing the same direction differ persistently in the force they bring to it, and learners who change direction often carry their characteristic intensity into the new field. We therefore distinguish a *general drive* — temperament, energy, activation level, indifferent to any particular orientation — from a *motivational drive*: propulsion that exists only as anchored to a specific direction. The two layers respond to different levers. General drive is constitutional and contextual; it can be protected, provisioned, and depleted, but not lectured into existence. Motivational drive is derivative: it appears when a direction captures general drive. How that capture happens is the subject of Section 3. What matters for the sorting is that the layers separate cleanly under delegability. ### 2.4 The sort Mapped against delegability, the components sort precisely. *General drive is selectable and supportable.* Institutions already act on it at admission, and act on it daily through workload design, wellbeing provision, and the removal of depleting friction — the externalization-cost reduction that AI now performs at scale belongs here, on the support side (Tomita, 2026a). What no institution has found is a way to train general drive into a learner who arrives without it; the honest levers are selection and protection, not production. *Mode is trainable.* A learner whose drive habitually critiques can learn to create; repertoires widen with practice and narrow with disuse. AI systems can scaffold the weaker modes directly — drafting so the critic has an object, structuring so the connector has materials — which makes mode supportable as well as trainable, and the least contested entry in the sort. *Motivational drive is convertible.* It is neither trained directly nor installed from outside; it is converted from general drive when a direction takes hold. The conditions of conversion are the mechanism question of Section 3. *Direction alone resists.* Here the sort requires precision, because what resists is not the supply of candidate directions. Proposals are cheap: a teacher can suggest ten, an AI can generate fifty on request, and both have a legitimate role in widening what a learner encounters. What cannot be handed over is ownership. A proposed direction becomes the learner's own — or fails to — through a process the proposer does not control, and the difference between an installed direction and an owned one is behaviorally visible: compliance tracks the supervisor's attention; commitment does not. Self-determination theory documented this boundary a generation ago from a different angle: externally proposed goals become integrated, rather than merely complied with, only under conditions that are relational as much as informational (Ryan & Deci, 2000) — a finding whose second-personal undertone Section 5 returns to. Direction is not installed. It crystallizes as the learner's own, or it remains someone else's. Table 1 summarizes the sort. **Table 1.** Learner agency under the delegability axis. | Component | Under delegability | Educational lever | |---|---|---| | General drive | Selectable and supportable; not directly trainable | Admission; workload and wellbeing design; AI-based removal of depleting friction | | Mode | Trainable and supportable | Repertoire widening through practice; AI scaffolding of weaker modes | | Motivational drive | Convertible from general drive | Encounter conditions (Section 3); protection of what crystallizes | | Direction | Resists: proposals deliverable, ownership not | Encounter engineering (Sections 3, 5); never installation | *Note.* Answerability does not appear in this table. It is not a component of agency and cannot be sorted by delegability; Section 4 establishes its different standing. ### 2.5 What the sort shows about distribution Set this result against the distribution account of agency. Reading agency as a property of human–AI systems rather than of individuals (Cukurova, 2026) is right — indispensably right — about most of the list above: planning and execution scaffolds, support structures for general drive, mode-widening tools, even the maintenance architecture around an existing motivational drive all genuinely distribute across hybrid configurations, and the sort specifies exactly where that reading holds. What the sort adds is the location of its silence. A hybrid system can host proposals for direction in any number; what it cannot do is own one on the learner's behalf, because ownership is not a system property — it is the learner's, or it is nobody's. Distribution accounts are right about what distributes — and silent about what does not. The present article is therefore not a rejection of distributed agency but a component-level restriction on what distribution can legitimately explain. And the sorted residue hands the next section its question: if direction cannot be installed, where does it come from? ## 3. A mechanism model: encounter, crystallization, conversion If direction cannot be installed, where does it come from? We propose a mechanism model in three movements: embodied encounter with a problem-holder precipitates the crystallization of evaluative commitment, and the crystallized direction then converts general drive into motivational drive. The model is deliberately spare — each movement names a condition, not a guarantee. The metaphor is chosen for its precision, not its color. Crystallization needs a supersaturated solution and a nucleation site: the general drive of Section 2 is the supersaturation — energy without an object — and the embodied encounter supplies the site. Read this way, the model makes intelligible the two failure modes educators already know. Exhortation without encounter — telling an unengaged learner to *find their passion* — addresses a solution that has nowhere to crystallize: there is energy, but no site at which a commitment could form. Encounter without protected follow-through — a moving field placement absorbed back into a packed schedule of unrelated demands — dissolves what had begun to form. Neither failure is a motivation deficit, and neither is fixed by more exhortation. Three distinct observational sources converge on this sequence. The innovation cases assembled by Nonaka and Takeuchi (1995) show the pattern repeatedly: developers' commitments form at sites of direct contact with users and their troubles, not in planning meetings, and the organizational machinery that mattered was the machinery that put people in front of problems and then protected what crystallized there. The publicly documented winner profiles of a recent global AI-assisted development competition show the same pattern from the opposite end of the tooling era: among the top entrants, a lawyer automating building permits, a cardiologist building post-visit care, a musician, a road engineer — practitioners whose commitments predated their tools, for whom AI removed the externalization barrier and changed nothing about where the direction had come from (in a hackathon hosted by an AI developer; Anthropic, 2026). And the single-case trajectory documented in prior work (Tomita, 2026a) shows the sequence at fine grain: a domain commitment formed in the researcher's own usage context steered the tool, not the reverse. The three sources share no method, population, or era; what they share is the shape of the sequence. Jointly, they challenge the two default stories — that direction is installed by instruction, or that it is fixed temperament — since in each case it postdates an encounter and predates the tooling. We present this as a mechanism model grounded in convergent observation — not a demonstrated mechanism. No controlled study here isolates encounter as the cause of crystallization, and the sources above were not collected to test the model they now illustrate. The model earns its place by what it organizes and by what it identifies as manipulable. Its educational translation is correspondingly modest and probabilistic: if encounter can precipitate direction, then education can design the conditions under which such encounters become likely — it can guarantee exposure, never crystallization. The model should therefore be read as a design hypothesis for educational technology, not as a causal account of motivation formation. What such design looks like in practice, and what it deliberately leaves alone, is the work of Section 5. ## 4. What the decomposition isolates ### 4.1 A different kind of remainder Direction resisted distribution, but it remained a component — something a learner comes to have. The decomposition's second function is different in kind. Run to completion, the sort yields no further component; what it leaves exposed is a relation that was never on the component list: *answerability* — the standing to have to answer for a judgment, introduced in prior work as the non-transferable relation in which the warrant of a judgment remains anchored in a particular agent (Tomita, 2026b). The grammatical contrast carries the conceptual one: one *has* a direction; one *stands in* answerability (Figure 1). A component is a property of the learner that pedagogy can act on. Answerability is not a property of the learner at all. It is a relation between the learner and those entitled to ask — which is why it could not have appeared as a fourth box however finely the sorting continued, and why treating it as one more trait to be fostered would repeat, at the level of theory, the conflation this paper exists to undo. The term is used here in a specific sense. *Answerability* already circulates in education through a Bakhtinian lineage — the answerability of an utterance or act, an ethics of being implicated (e.g., Ewald, 1993); the present sense is Darwall's, the standing to have to answer to one who can demand, and the two should not be elided. *Figure 1 (image omitted).* Direction, general drive, motivational drive, and mode are components the learner *has*. Answerability is not a fourth component; it is the Darwallian second-personal relation of address and response in which the learner *stands* to those entitled to ask, with the non-transferable standing to answer for a judgment. ### 4.2 Not an artifact of the axis An objection should be met before anything is built on the claim: that the delegability axis was constructed to exclude answerability, so its isolation is trivial — true by definition. Two replies, one procedural and one empirical. Procedurally, the sorting rule was applied only to capacities ordinarily counted as agency — orientation, drive, mode of acting; answerability nowhere figures in the rule, and it surfaces only when the sorted capacities, however completely enumerated, fail to explain who can be called to answer for a judgment. Empirically, the exclusion cannot have been smuggled in by our vocabulary, because the vocabulary was never in the field to smuggle: in the most fine-grained multidimensional account of machine agency to date, the terms responsibility, accountability, answerability, blame, and delegation do not occur at all (Dung, 2025). The axis did not push answerability out of this strand of agency theory. It revealed that this strand had never contained it. The shape of the move repeats the method of prior work: there, scoring what could be scored located what could not be legitimately delegated (Tomita, 2026b); here, sorting what can be delegated exposes what was never up for delegation. The axis is not chosen to exclude answerability; it is chosen because generative AI has made handover the practical question for every component of learning agency. ### 4.3 The second-personal ground What kind of fact is answerability, if not a capacity? Here the paper borrows, and the borrowing must be marked precisely. From Darwall (2006) we take the structure of second-personal address. Claims and demands are addressed person to person, and valid address presupposes, as its felicity condition, the addressee's second-personal competence: the capacity to recognize the demand's validity, to hold oneself to it, and so to have to answer to the addresser. Being responsible *for* something, in this account, is conceptually tied to being responsible *to* someone. In educational settings this relation is not abstract moral theory but the ordinary structure of being asked to justify one's solution, interpretation, or design choice to another person. What is borrowed is this address structure. What is original here is the extension: that the warrant of a judgment, no less than conduct, requires a bearer who can be addressed and challenged, and who must answer. Darwall theorizes demands on action; the claim that evaluative judgment stands in the same second-personal relation is this paper's move — consonant with his text, but not stated in it. Where Darwall has been brought into education before, it has largely been by way of respect and dignity (Giesinger, 2012); applying his second-personal accountability to the warrant of a learner's judgment, as the concept education must protect under generative AI, is a different route. One feature of his account, however, transfers with particular force. Darwall freely admits representation on the claiming side: trustees may demand on another's behalf, and third parties may hold reactive attitudes for a victim. The answering side never appears by proxy — guilt's natural expressions are confession, apology, and self-addressed reproach. On the reading developed here, that asymmetry marks the joint: what cannot be done on one's behalf is precisely to answer. This sharpens a distinction prior work had begun to draw: where *accountability* can diffuse across institutional structures, *answerability* names the non-transferable relation in which the warrant of a judgment stays anchored in a particular agent (Tomita, 2026b). This article carries that distinction into education — reserving *answerability* for the non-transferable standing just described, and keeping *accountability* for institutional relations, which can, by design, be delegated, audited, and reassigned. This standing runs in the opposite direction to the *authority* that Xing et al. (2026) find learners conferring on AI teachable agents: authority is conferred by the learner and tracks perceived competence, whereas standing is retained by the one who judges and tracks answerability. The agent can be granted more authority without acquiring any standing. A recent preprint offers a complementary route from the AI side: Archer and Wiltsche (2026) argue from systems' cognitive limits — the absence of world-directed, temporal, and anticipatory cognition — through an epistemic answerability to the world, that responsibility lies not in the model but in the system surrounding it, and cannot be offloaded. The two routes divide the labor — that account explains the systems, this one locates the relation — and the division is principled: an argument from present limits must be re-made each time capability moves; an argument from the structure of address does not. ### 4.4 What the relation grounds The second layer is not a decoration on the first; it is what keeps the first from collapsing into capability education. Strong, coherent direction guarantees nothing by itself — a skilled fraud has direction in abundance, and answers to no one until made to. The qualification is the point: a human fraud *can* be made to answer, and that very susceptibility — to being addressed, challenged, and held — is second-personal competence in action. A current AI system output cannot be made to answer in this sense: nothing can be taken from it, no commitment of its own is at stake, and self-reproach, were it scripted, would idle — and it has no such standing unless an institution assigns the answer to a human agent. The difference is not one of degree along any of Section 2's dimensions; it is the difference between refusing a relation and being unable to stand in one. The asymmetry has a practical edge the other way as well: uncritical delegation does not leave the learner's side untouched but erodes it, as evidence of weakened planning, monitoring, and evaluation under heavy AI reliance indicates (Fan et al., 2025) — which is why the prescription that follows treats delegation as something to be shaped, not merely permitted. The boundary this draws does not turn on a metaphysical verdict about machines. It does not require settling whether future systems may simulate address-like behavior; it concerns the educational institution's present criterion for standing — whether a learner can be called to own, revise, and answer for a judgment as theirs. That learners routinely experience AI systems as second-personal interlocutors — consulting them, trusting them, calling them partners — is explained by well-attested mechanisms of anthropomorphism (Epley et al., 2007; Gray et al., 2007) and does not imply that the system stands in the relation; a perceived relation and a relation that can be the addressee of answerability are distinct, on the same reading that separated conferred authority from retained standing above. On this extension, Darwall's structural point secures the gap: the second-personal notions — authority, valid claim, second-personal reason, accountability — form an interdefinable circle that cannot be entered from outside it, so no enumeration of first-layer capacities, however long, adds up to entry. This, in turn, grounds what prior work could only state as a condition. Tomita (2026b) held that the delegation boundary stands as long as legitimacy over value judgments remains institutionally attributed to humans — a formulation that leaves the attribution looking like convention. It is not. Institutions keep attributing judgment to humans because standing inheres only in second-personal competence; the condition is backed by structure, and would have to be re-engineered, not merely re-decided, to move. Nor does the argument lean on the inner opacity of present systems, so it is robust to progress in interpretability: however legible a system's internals become, legibility yields evidence about capability, never standing. And nothing here forecloses the question whether some future system could acquire second-personal competence — the boundary tracks the competence, not the substrate. What it forecloses is shortcutting that question by capability alone. ## 5. The two-stage prescription The two-layer structure yields a two-stage prescription, and the stages must not be merged. Stage one addresses everything the decomposition can place — it diagnoses components, designs encounters, and configures AI so that delegation removes barriers without substituting for the learner. Stage two addresses what stage one cannot reach. For each prescription we state a principle and one worked example, and we say explicitly what is left to others; the restraint is not modesty for its own sake but the framework's own discipline — a paper about non-delegable judgment should not deliver judgments that belong to institutions. ### 5.1 Stage one: a diagnostic compass *Principle.* The OECD's Learning Compass 2030 placed agency at the centre of a framework for how learners navigate an uncertain world, but it is a compass that points a direction to travel, not one that resolves that direction into components an educator can read (OECD, 2019). This stage supplies the second kind: a diagnostic compass that disaggregates the direction the OECD's compass indicates. Read a learner's components separately instead of guessing at aptitude as a single quantity. Direction, general drive, motivational drive, and mode are distinct findings, and several have established instruments adjacent to them — self-regulated-learning measures speak to mode and its trainability (Zimmerman, 2000), engagement and vitality measures to general drive, interest inventories to candidate directions. The compass requires no new psychometrics; it requires reading existing ones against the delegability sort instead of summing them. One reservation is constitutive rather than cautionary: the compass reads only what can be externalized. Answerability appears on no needle, and a compass that claimed to score it would re-import the conflation this paper diagnoses — the point at which both the OECD's compass of direction and this compass of components reach their limit, leaving the same non-measurable remainder isolated. This is also the precise sense in which the present account accepts the demand that agency remain measurable rather than a residual aspiration (Cukurova, 2026) — and then locates, at that demand's edge, a core that is neither measurable nor an aspiration: it is cultivated, second-personally, in stage two. *Worked example.* A learner presents as "unmotivated." The compass disaggregates: general drive high — the same student is animated elsewhere; direction absent; mode predominantly critical. The reading is not a motivation deficit but a pre-crystallization state with a strong engine and no object, and the indicated lever is exposure to encounters, not exhortation. One further effect is rhetorical but consequential: the compass replaces *find your purpose* — a binary demand that mostly paralyzes — with *observe your own gradients*, an incremental practice a pre-crystallization learner can actually perform. *Left to others.* Which instruments, with what weights, against which values: to each institution, whose reading of its own learners is itself a judgment someone must answer for. ### 5.2 Stage one: encounter engineering *Principle.* If encounter can precipitate direction (Section 3), then curricula can be designed for encounter density: structured contact with problem-holders, early and embodied, before competence would conventionally "justify" it. Medicine has run this design for a century — clinical clerkship routes students through patients on the explicit premise that direction toward care is formed at the bedside, not installed in the lecture hall. The model can serve as a design heuristic wherever curricula can create sustained contact with problem-holders. *Worked example.* For disciplines without wards, the encounter can be miniaturized into what we call a discomfort inventory exercise: the learner delegates a task wholly to an AI system, then inventories every discomfort with the output. The inventory is the learner's first specification profile — the felt difference between fluent output and right output is exactly where tacit commitment surfaces, following the ablation logic documented in Tomita (2026a). The exercise turns the AI's weakness-without-specification into a mirror, and a sequence of such micro-encounters into a curriculum. Independent evidence is consistent with the design: studying EFL learners revising AI-drafted statements of purpose, Jiang et al. (2026) found revision patterns ranging from compliance-oriented acceptance through form-oriented modification to content-oriented innovation — a gradient of crystallized direction — with two factors dividing them: the felt gap between an authentic and a GenAI-constructed voice, and enthusiasm to present one's intentions. Those two map onto components this article decomposed independently — the nascent answerability of a judgment one will have to own, and the motivational drive of a direction one has taken up — surfacing as empirical dividing factors in a study that did not set out to test the decomposition. Where Jiang et al. document the gradient and prescribe that learners "remain the agents," the present account provides a mechanism model for what their study observed after the fact: the discomfort inventory induces that gradient from the front. *Left to others.* Which problem-holders, at which sites, with what safeguards: discipline-specific terrain, owned by those who know it. The design guarantees exposure, never crystallization, and an honest program says so. ### 5.3 Stage one: AI guardrails *Principle.* Configure educational AI to distinguish two requests that look alike. When a prompt carries the learner's specification — their constraints, judgments, reasons — completion removes externalization cost only, and is the legitimate use. When the prompt is specification-empty, completion would substitute for the very thing education exists to develop; the right behavior is to route the request back into dialogue until the learner's own commitments appear, and only then to draft. *Worked example.* An essay assistant receives "write my essay on X." Instead of producing text, it asks what the learner finds wrong with the standard view of X, which two sources they would trust and why, and what conclusion they would defend if pressed — and drafts only from the answers. The same model, the same capability, a different default. *Left to others — with two stated limits.* Detection of specification can be gamed; a determined learner can launder emptiness through borrowed opinions, so the guardrail shapes the default path rather than guaranteeing the outcome. And dialogue with an AI, however well configured, does not cultivate answerability — the system can elicit commitments, but it cannot be the one to whom the learner answers. Guardrails serve stage one. They do not substitute for stage two. ### 5.4 Stage two: the second person *Principle.* Answerability grows only inside second-personal exchange: being questioned, answering, and owning the answer. The educator's irreplaceable function — the one no configuration of tools absorbs — is to be the other person: to ask the learner *what is your assessment?* and to hold them to what they say. Clinical supervision has always worked this way: the resident holds the standing to answer from the first day of licensure; what training builds is the competence to answer well, and it builds it by making the resident answer, case after case, to someone entitled to ask. The structure generalizes beyond medicine: a learner's standing does not wait for graduation — from the moment work is submitted under their own name, the answer is theirs — and what education builds is not the standing but the competence to occupy it. Assessment forms that stage this exchange — oral defense of one's own work, process accounts the learner must stand behind — are implications worth one sentence each here and a literature of their own elsewhere. The prescription is structurally grounded, not nostalgic. When AI accelerates the externalization and combination of knowledge, the rate-limiting processes shift to socialization and internalization (Nonaka & Takeuchi, 1995; Tomita, 2026b) — and socialization, where evaluative commitments are contested and revised between persons, is precisely the second-personal site. The relational conditions that self-determination theory found necessary for internalization point the same way (Ryan & Deci, 2000). Nor is this a retreat from human–AI configurations: the hybrid is the right vessel, and stage one fills it. But within any configuration, the answerable party is the human in it, not the configuration — layers that mesh, not a fusion that answers. What stage two restores is older than the lecture hall: apprenticeship's core, with its drudgery finally delegable, returned to the center. *Left to others.* Teacher development and assessment regimes that institutionalize the exchange: stated here as implications, designed by those who will answer for them. ## 6. Conclusion What remains for the learner when AI can do the work? The question dissolves once agency is decomposed along the axis the question itself supplies. Most of what education gathers under agency can be supported, trained, or shared with machines, and should be — the sort in Section 2 says where, and stage one of the prescription says how. What cannot be handed over is of two kinds, and keeping them distinct is this article's central discipline: direction, a component that resists installation but yields to encounter; and answerability, not a component at all but the standing in which a learner answers to someone for a judgment. AI changes the economics of everything around that relation, stripping the externalization burden from teaching and learning alike; what it does not change is the relation itself: who must answer, and to whom. The educator's position is thereby clarified, not diminished: design the encounters, configure the tools, read the compass; and then be the person the learner answers to. Education in the age of AI is not the defense of human tasks against automation. It is the cultivation of what was never a task. ## Practitioner Notes **What is already known about this topic** - Generative AI systems now plan, draft, and revise, making learner agency a central contested concept in educational technology. - Distribution accounts treat agency as a property of human–AI systems; psychological accounts treat it as an individual capacity to be fostered. The two run in parallel. **What this paper adds** - A decomposition of learner agency by delegability — direction, two-layer drive, and mode — rather than by psychological function. - A mechanism model of direction formation: embodied encounter → crystallization of commitment → conversion of general drive into motivational drive. - Answerability identified as a relation isolated, not contained, by the decomposition: the non-delegable standing to have to answer for a judgment. - A two-stage educational prescription matching the two layers. **Implications for practice and/or policy** - Diagnose a learner's components separately (a compass), instead of guessing at aptitude as a single quantity. - Design encounters with problem-holders rather than running motivation campaigns. - Configure educational AI to route specification-empty requests back into dialogue rather than completing them. - Protect second-personal exchange — being questioned, answering, owning the answer — as the central site where answerability grows. (Answerability here means that learners must remain able to say why a judgment is theirs when questioned by another person.) ## Statements **Conflict of Interest.** The author declares no competing interests. **Data Availability.** This is a theoretical article. No datasets were generated or analysed. All sources discussed are publicly available and cited in the references. **Generative AI Use.** The author used generative AI tools (large language models) as drafting and editing aids during manuscript preparation, and as an object of analysis discussed in the text. All conceptual claims, the framework, the argument, and final wording are the author's own; the author takes full responsibility for the content. **Funding.** None ## References Anthropic. (2026, April 20). *Meet the winners of our Built with Opus 4.6 Claude Code hackathon* [Blog post]. https://claude.com/blog/meet-the-winners-of-our-built-with-opus-4-6-claude-code-hackathon Archer, K., & Wiltsche, H. (2026). *The origins of artificial intelligence in natural intelligence* [Preprint]. PhilArchive. https://philarchive.org/rec/ARCTOO-2 Bandura, A. (2001). Social cognitive theory: An agentic perspective. *Annual Review of Psychology, 52*, 1–26. https://doi.org/10.1146/annurev.psych.52.1.1 Cathcart, S., Priestley, M., Priestley, A., & Rushton, E. A. C. (2026). Unravelling agency: A critical review of conceptualisations of agency for educational research, policy and practice. *Review of Education, 14*(1), e70131. https://doi.org/10.1002/rev3.70131 Cukurova, M. (2026). Agency as a system property in human–AI interaction in education. *British Journal of Educational Technology, 57*(4), 1065–1070. https://doi.org/10.1111/bjet.70060 Darwall, S. (2006). *The second-person standpoint: Morality, respect, and accountability*. Harvard University Press. Dung, L. (2025). Understanding artificial agency. *The Philosophical Quarterly, 75*(2), 450–472. https://doi.org/10.1093/pq/pqae010 Emirbayer, M., & Mische, A. (1998). What is agency? *American Journal of Sociology, 103*(4), 962–1023. https://doi.org/10.1086/231294 Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. *Psychological Review, 114*(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864 Ewald, H. R. (1993). Waiting for answerability: Bakhtin and composition studies. *College Composition and Communication, 44*(3), 331–348. https://doi.org/10.2307/358987 Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. *British Journal of Educational Technology, 56*(2), 489–530. https://doi.org/10.1111/bjet.13544 Giesinger, J. (2012). Respect in education. *Journal of Philosophy of Education, 46*(1), 100–112. https://doi.org/10.1111/j.1467-9752.2011.00836.x Gray, H. M., Gray, K., & Wegner, D. M. (2007). Dimensions of mind perception. *Science, 315*(5812), 619. https://doi.org/10.1126/science.1134475 Jiang, Y., Wu, Q., Yang, Y., Jian, C., & Zhao, J. (2026). Learner agency in revising GenAI-generated statements of purpose. *British Journal of Educational Technology, 57*(4), 965–983. https://doi.org/10.1111/bjet.70041 Kortemeyer, G., Dittmann-Domenichini, N., Schlienger, C., Spilling, E., Yaroshchuk, A., & Dissertori, G. (2023). Attending lectures in person, hybrid or online—How do students choose, and what about the outcome? *International Journal of Educational Technology in Higher Education, 20*(1), 19. https://doi.org/10.1186/s41239-023-00387-5 Luo, J. (2025). How does GenAI affect trust in teacher–student relationships? Insights from students' assessment experiences. *Teaching in Higher Education, 30*(4), 991–1006. https://doi.org/10.1080/13562517.2024.2341005 Nonaka, I., & Takeuchi, H. (1995). *The knowledge-creating company: How Japanese companies create the dynamics of innovation*. Oxford University Press. OECD. (2019). *OECD learning compass 2030: A conceptual learning framework* [Concept note]. OECD. https://www.oecd.org/content/dam/oecd/en/about/projects/edu/education-2040/concept-notes/OECD_Learning_Compass_2030_concept_note.pdf Ryan, R. M., & Deci, E. L. (2000). Self-determination theory and the facilitation of intrinsic motivation, social development, and well-being. *American Psychologist, 55*(1), 68–78. https://doi.org/10.1037/0003-066X.55.1.68 Tomita, K. (2026a). *Domain-native development: A Mekiki framework for AI-assisted knowledge work* [Preprint]. SocArXiv. https://doi.org/10.31235/osf.io/cwkav_v1 Tomita, K. (2026b). *Philosophy as cognitive assay: Measuring the delegation legitimacy boundary in AI-assisted knowledge work* [Preprint]. SocArXiv. https://doi.org/10.31235/osf.io/e9qw5_v2 Xing, W., Kim, T., Song, Y., Li, H., Li, C., & Kim, J. (2026). Unveiling interaction patterns between students and a generative AI teachable agent: Focusing on students' agency and AI agents' authority. *British Journal of Educational Technology, 57*(4), 896–923. https://doi.org/10.1111/bjet.70038 Zimmerman, B. J. (2000). Attaining self-regulation: A social cognitive perspective. In M. Boekaerts, P. R. Pintrich, & M. Zeidner (Eds.), *Handbook of self-regulation* (pp. 13–39). Academic Press. https://doi.org/10.1016/B978-012109890-2/50031-7