실무자 인식론 연구

Practitioners without academic training face a flood of management and self-development frameworks (from SWOT to Lean Startup to Grit), each marketed as evidence-backed. This paper operationalizes a heuristic a practitioner-author proposed independently, later recognized as resembling the classical positivist and behaviorist explain-predict-control schema: a framework's evidential standing can be graded by whether it only explains a phenomenon after the fact, also predicts it prospectively, or has been shown, through causal designs, to improve outcomes when its own prescription is applied. We make this ladder explicit as a four-level codebook (L0: no testable causal claim; L1: explanatory or correlational support only; L2: prospective predictive support; L3: causal or interventional support), with decision rules for retrodictive evidence, single-study results, and cross-study heterogeneity, then apply it to eighteen well-known frameworks. Two independently prompted coders, both instances of the same large-language-model family (Anthropic's Claude) working from a shared structured evidence review, reached substantial agreement (exact agreement = 0.83; quadratic-weighted κ = 0.72; unweighted κ = 0.70 across n = 18), with five disagreements resolved by adjudication. The final classification shows a pronounced popularity-evidence gap: fourteen of eighteen frameworks (78%) receive only L0 or L1 support despite wide adoption, two reach L2 with caveats, and only two reach L3 causal support for their own prescription. We situate the ladder within its philosophical and management-science lineage rather than claim novelty for the three-rung idea, and discuss its narrow contribution as a screening instrument alongside its limits as a substitute for context-specific judgment.

Practitioners without formal training in judgment and decision-making research sometimes describe expertise as an increase in perceptual "resolution," moving from black-and-white categories to graded, multidimensional discrimination. This paper asks whether that practitioner metaphor can be stated precisely enough to be measured. We explicate cognitive resolution (CR) as an ordinal construct that integrates, without superseding, five established literatures: Werner's orthogenetic principle, Bieri's cognitive complexity-simplicity, Suedfeld and Tetlock's integrative complexity, Barrett's emotional granularity, and King and Kitchener's reflective judgment together with Basseches's dialectical thinking. A genealogy table states what each prior construct already measures and what CR adds: a cross-domain visual metaphor and a countable operationalization. We present a four-level scoring rubric (categorical, graded/contextual, multi-dimensional/perspectival, integrated/element-level) for short first-person texts and non-specialist or LLM-assisted coding, distinguished from integrative complexity by object of measurement, scale structure, and intended use. A pilot had three same-model-family large-language-model instances score 24 constructed vignettes against a known key, yielding high nominal agreement, one clear failure pattern at the top of the scale, and a non-trivial length-level correlation; because raters and vignettes share one model family, these figures indicate feasibility of rubric codeability, not reliability or validation. A roadmap sets out the convergent, discriminant, and human-rater tests genuine validation would require on real texts. A practitioner self-audit checklist appears as an appendix. Finally, the staged model that popularized the metaphor was generated by a large language model in dialogue with the author, not by the practitioner, an attribution issue this paper names in its own origin story.

A single sentence from an unelaborated dialogue proposed that sufficient contrast, the capacity to draw distinctions, plus sufficient resolution, the capacity to register fine gradations, yields epistemic clearness, by analogy with image sharpness in optics, where sharpness is a joint function of contrast and spatial-frequency resolution. This paper treats that sentence as a construct-explication problem rather than a finished theory. It first shows that the surface metaphor already exists in popular literature and is not original. It then confronts an unresolved terminology clash: "resolution" already names two distinct technical quantities in adjacent literatures, metacognitive resolution, a Type-2 signal-detection sensitivity that tracks confidence-accuracy correspondence, and forecast-verification resolution, a Brier-score decomposition term that tracks outcome-base-rate discrimination across probability categories, neither of which is the graded, first-order differentiation capacity the optical metaphor requires. The paper explains why its own Resolution axis sits closer to the second sense than the first while remaining distinct from both. On this cleared ground, it gives Contrast and Resolution independent operational definitions, argues for their orthogonality using dichotomous-thinking, integrative-complexity, and construal-level literatures, and crosses the two axes into a 2×2 typology: blur, false sharpness, fine-grained fog, and clearness, the last an emergent joint state rather than a simple sum. The paper closes with an approximately twenty-item pool intended for future scale development, explicitly unvalidated, and a disclosure stating exactly which sentence in the primary dialogue is the author's own.

An independent restaurant owner-operator turned AI-assisted researcher has argued, in an unpublished dialogue with a large language model, that because people perceive the world so differently, practitioners should deliberately minimize gray areas and decide sharply, for the sake of resource efficiency, message clarity, and decision completeness. The identical cognitive operation, collapsing graded uncertainty into a binary judgment, is treated in the clinical literature as a diagnostic marker of distortion: all-or-nothing or black-and-white thinking, a validated correlate of rumination, depressive symptoms, and personality-disorder features. This paper treats that contradiction as the object of analysis rather than resolving it by citing only one side. Drawing on essentialism, satisficing, ambiguity tolerance, need for cognitive closure, contested choice-overload evidence, the Cynefin framework, paradox theory and polarity management, strong inference and falsificationism, and the clinical literature on dichotomous thinking, it develops an eight-condition boundary typology distinguishing productive dichotomization, in action, resource, and communication decisions, falsifiable hypotheses, and simple or complicated tasks, from pathological dichotomization, in judging persons, treating unsettled evidence as final, and identity-laden disputes better managed as an enduring polarity than resolved by choosing a side. The paper also examines, rather than accepts, the author's move from perceptual heterogeneity to a general prescription for dichotomization, and distinguishes the resulting typology from the author's related working paper on strict premises in binary trade-offs, which addresses a later stage of the decision process than the categorization question here. The paper is conceptual and reports no new empirical data.

Respecting another person's difference does not require tolerating every factual error they hold. Once a belief has been separated from mere preference and classified as objectively wrong, a further practical question remains for anyone who must share space with the person who holds it: correct the error, or let it stand? This practitioner note develops a three-tier harm-radius rule for that decision in a previously unaddressed setting: a single practitioner, face to face, with no research time and no triage queue, such as an owner-operator in a small workplace. The rule directs the practitioner to act when the error's potential harm is immediate or direct, to weigh intervention when the harm is potential or social, and to refrain when the error is local and harmless. A six-item checklist operationalizes each tier, drawing on the harm principle, error-management research, face-work and politeness theory, psychological safety, correction-efficacy findings, and epistemic paternalism. The note makes two contributions: it treats the decision to correct as analytically separable from the prior question of whether a belief is wrong at all, a distinction the wrong/different test itself does not draw, and it translates existing harm-weighting logic to this practitioner niche, a setting existing frameworks were not built to address. That second, larger move translates harm-weighting logic already present in institutional, media-scale fact-checking frameworks; it is not a new taxonomy of harm. The rule and checklist are illustrated with clearly labeled constructed scenarios; no empirical validation is claimed or reported.

Deliberative-implemental mindset theory holds that people think realistically before committing to a course of action and grow optimistically biased once they have committed to it. Drawing on a founder's own first-person account of launching a small venture, this paper argues that a workable cognitive trajectory for a founder runs the other way. Before launch, near-certain conviction that the venture will succeed, a stance this paper calls Prophet Mode, is what lets a founder cross an action threshold that a calibrated probability estimate would not justify. Once the venture launches, the same founder must deliberately reverse into Paranoid-Survivor Mode, treating failure as the default outcome and actively searching for the scenarios that would produce it. The paper states this as a counter-Rubicon prescription: an event-triggered, one-way switch pinned to the single, unrepeatable act of launching. It distinguishes the switch from three constructs it is easily confused with: the simultaneous holding of faith and fact in the Stockdale paradox, the repeated and reversible peacetime-wartime toggle described by Ben Horowitz, and the recurring paranoia Andrew Grove attaches to strategic inflection points. Six falsifiable propositions specify the trigger, the direction of reversal, the one-way character of the switch, two named failure modes, and switching latency as a moderator, together with boundary conditions, repeated financing thresholds, serial founding, and team-level division of roles, under which the simple one-switch account should give way to repeated or distributed switching. The paper is conceptual; it reports no test of these propositions and rests on a single autoethnographic source.

A learner with no prior literacy in a field cannot use the content of an AI system's answer to judge whether it is true, because that judgment requires exactly the competence the learner lacks. This bind is already named and documented in the AI-literacy literature; this paper does not claim to discover it. Instead, four verification mechanisms drawn from the automated hallucination-mitigation literature, forced external retrieval, per-claim source attribution, cross-source triangulation, and role-separated adversarial self-review, are translated into a four-step manual protocol runnable in an ordinary chat window with no code, then piloted against unassisted answering across six domains unfamiliar to the author. Thirty matched questions per condition, one false-premise item per domain, were checked post hoc by blinded verifiers against external sources, so all figures below are descriptive, not inferential. Both conditions produced high answer accuracy (unassisted mean 0.93, fully correct 86.7 percent; protocol mean 0.98, fully correct 96.7 percent) and identical claim-level contradiction rates, with false-premise correction already at ceiling in both conditions. The protocol's effect concentrated almost entirely in citation validity: the share of cited sources confirmed to exist rose from 0.22 to 1.00, and the share actually supporting the claim attributed to them rose from 0.33 to 0.99. The paper interprets this dissociation as evidence that the protocol converts an unverifiable claim into a checkable one, without making the novice a better judge of truth, and reports a mid-pilot failure and repair, one underperforming domain, and the limits of a harnessed, same-family-verified pilot.

Prompt recipes for learning a new field with AI assistance already converge on a two-move pattern: ask the model to map a field's structure, then ask it to compress the field to its highest-leverage fraction. This practitioner note, built from one author's archived dialogue with an AI system in November 2025, argues that a third move, usually left implicit, should be made explicit and placed first: before mapping a field, the learner should be asked to surface and consciously suspend what they already believe they know, because unexamined prior habits, especially an experienced practitioner's own domain routines, can distort how AI-generated structure and compressed content are absorbed. The Map-Core-Link portion of the resulting four-step sequence is shown operating within the same dialogue across two tracks, learning about AI itself and using AI to learn an unrelated field; the Unlearn step itself is argued for on track-agnostic grounds and is not separately demonstrated on each track. The note states plainly what it does not claim: the Map-Core-Link mechanism is documented prior art in prompt-recipe grey literature and in a peer-reviewed multi-agent system, and one of the dialogue's own turns, which supplied the initial three-step template, was pasted in by the author from a separate AI session and is credited as such rather than as original composition. The note closes with a proposed, unexecuted, within-person evaluation design for testing whether the explicit Unlearn step adds measurable value over skipping directly to Map. No data were collected for this paper.

Experts often resist artificial intelligence not because the tools fail but because adoption is framed, implicitly or explicitly, as an admission that everything they knew was wrong. Drawing on a dialogue in which an independent restaurant owner-operator asked an AI system whether adopting AI required doubting their entire cognitive history, and the system's reply distinguishing experience and judgment, an asset to keep, from methods and tools, a liability to retire, this paper develops a conceptual asset–liability framework for unlearning among non-organizational owner-operators in traditional industries. The organizational-unlearning literature and a fast-growing 2025–2026 literature on AI-induced professional identity threat converge on the same mitigation logic, protect identity and self-efficacy while replacing obsolete methods, but every study in that literature is set inside a formal employing organization. Independent owner-operators unlearn alone, with no human-resources-mediated safety net, and, per craft-identity research, with method and self less separable than in white-collar knowledge work. The paper reconstructs the framework's origin with attribution disaggregated between the practitioner's question and the model's terminology, formalizes a matrix distinguishing practices to keep from practices to retire with stated boundary conditions, derives five testable propositions, and specifies a complete, unexecuted interview protocol and coding frame for a future study of five to eight traditional-industry owner-operators. No interview data were collected; the contribution is the instrument and the population extension, not a new mechanism.

Why does the same verified, correct piece of information change one person's decisions and leave another's untouched? This paper argues that accurate information has no fixed value: realized value depends on the recipient, not on the information's correctness. Using a thought experiment built on the historically contested anecdote of Gauss summing the integers from 1 to 100, we separate the objective, decision-theoretic value correct information carries for an idealized rational agent from its realized value for a specific person, and propose a three-stage serial filter governing the gap between the two: receptivity (humility and openness to revise prior beliefs), inquiry (active pursuit and personal reconstruction of the information), and enactment (converting reconstructed understanding into action). We integrate three established individual-difference literatures, learning agility, intellectual humility and actively open-minded thinking, and growth mindset, under this filter, and draw on the absorptive-capacity, knowing-doing-gap, and implementation-intentions literatures to explain why even receptive, inquisitive individuals routinely fail at the final, enactment stage. We do not claim a new psychological construct: the three-stage structure closely parallels the learning-agility construct family, and our contribution is limited to two things: the accuracy/realized-value distinction motivated by the Gauss illustration, and a re-operationalized, individual-level self-diagnostic checklist adapted from these literatures for practitioners outside organizational leadership-assessment settings. A short illustrative section, non-evidentiary by design, applies the filter to two independently documented historical cases and one first-person practitioner episode. The paper closes with boundary conditions and the empirical validation future work must supply.

Practitioners who reach durable conclusions through lived experience have no established, low-cost way to check those conclusions against the scholarly record short of formal academic training. This methods paper formalizes a three-stage protocol, externalizing tacit claims as text, reverse-matching them against candidate academic theories with a large language model, then adversarially searching for prior art, that a solo, non-institutionally affiliated practitioner-researcher can run alone and, in principle, reproducibly. The protocol is not presented as a discovery: its central operation is structurally close to Argyris and Schön's audit of espoused theory against theory-in-use, and it inherits the validity problem Schön identified in that tradition, namely that self-report is not a reliable route to one's own theory-in-use. Layered onto this classical threat are risks specific to using a language model as the matching instrument, among them sycophantic confirmation of the practitioner's framing, fabricated or misattributed citations, and the silent relabeling of model-generated structure as the practitioner's own insight. We specify Protocol v1's countermeasures, namely counter-search and falsification prompts, multi-model cross-verification, blinded re-matching, and an explicit attribution audit, and demonstrate the pipeline on 27 claims extracted from eight archived practitioner-LLM dialogues, checked against 146 located prior-art items (0 novel, 25 incremental, 2 redundant; mean self-reported model confidence 0.62). The demonstration's own attribution audit catches four cases of an archiving note misattributed as authorial metacognition and two cases of model-authored structure later adopted as the practitioner's — a failure inside the protocol's own operation of exactly the kind it is designed to detect.

When a practitioner develops a conceptual framework through sustained dialogue with a large language model, what the practitioner remembers building and what the transcript shows can diverge. This paper double-codes every concept unit (N = 107) across eight archived dialogues (84 turns, June 2025 to January 2026) between an independent AI consultant and restaurant owner-operator and Google Gemini, using a five-category origin codebook applied by two independent large-language-model coders and adjudicated where they disagreed (29 of 107 units; exact agreement .73; Cohen's kappa = .63 on the five-category scheme, .58 collapsed to three). The resulting distribution of 15 user-originated, 35 user-seed/model-elaborated, 3 co-constructed, 18 model-originated-and-adopted, and 36 model-originated-and-unused units, against a corrected model-to-user character ratio of 13.98 to 1, shows that the model originated roughly half of all coded units (50.5 percent, whether or not the practitioner used them). A case study of the practitioner's four-book series makes the mechanism concrete: the practitioner asked the model to recall prior themes from its memory of earlier sessions; the model retrieved and sequenced them, including the governing curriculum sentence the practitioner later treated as their own; the practitioner's only substantive input was a one-item correction. No existing authorship-disclosure framework has a category for this pattern, which differs from AI ghostwriting, sycophancy, and cryptomnesia. The paper is reflexive on two fronts: the researcher is the coded subject, and this manuscript was drafted with heavy AI assistance. It closes with a reusable codebook, a disclosure-paragraph template, and a call for replication with human coders.