The Augmented Behavioural Capacity (ABC) Model: Large Language Model Assistance and Calibrated Retail-Investor Decision-Making

1. Introduction

Large language models (LLMs) are entering retail-investor decision-making at a point that earlier digital, financial, and decision-support technologies did not reach. They are no longer encountered only as systems to be adopted, configured, and used (Cohn et al., 2024; Colombatto et al., 2025; Heersmink et al., 2024; Mogaji et al., 2024). They are increasingly encountered as conversational reasoning partners with which an investor can structure a problem, compare alternatives, simulate scenarios, request counterarguments, and check the consistency of a candidate course of action - all at the moment of cognitively demanding evaluation, before a trade is placed (Handler et al., 2024; Li et al., 2025; Si et al., 2024; Wang et al., 2025). The phenomenon this paper takes as its starting point is therefore not whether such systems are accepted or used, but how LLM engagement may reshape the evaluative stage (The term evaluative stage refers to the decision phase in which the actor structures the problem, constructs and weighs alternatives, estimates risk, and forms criteria for action. This corresponds to the Design and Choice phases of Simon’s (1960) classical decision-process taxonomy, to the evaluation phase of Kahneman and Tversky’s (1979) prospect-theoretic account of decisions under risk, and to the constructive evaluation processes characterised by Payne, Bettman, and Johnson (1993) as cognitive-effort-bounded strategy selection) of retail-investor decision-making: the phase in which the investor structures the problem, weighs alternatives, estimates risk, and forms criteria for action, corresponding to the design-and-choice phases of Simon’s (1960) decision-process taxonomy.

The dominant inherited frameworks address adjacent but different problems. The Technology Acceptance Model (TAM) and the Unified Theory of Acceptance and Use of Technology (UTAUT) explain whether users judge a system useful, easy to use, socially supported, or controllable, and whether they intend to use it (Davis, 1989; Venkatesh, Morris, Davis, & Davis, 2003; Venkatesh, Thong, & Xu, 2012). The Theory of Planned Behaviour (TPB) explains how attitudes, subjective norms, and perceived behavioural control combine into behavioural intention (Ajzen, 1991). These families remain necessary, because LLMs are still adopted technologies whose use depends on perceived utility, effort, social influence, and control. They are not, however, sufficient: they were not designed to explain what happens to the evaluative content of a user’s reasoning when a generative, dialogic system enters that reasoning at the moment of choice - in effect, they locate the locus of decision-making within the individual user and treat the system as an external object of use rather than as a participant in the evaluation itself. Section 3 develops this gap as a three-layer (temporal, ontological, phenomenological) insufficiency; here it is stated only as the motivation for a domain-bounded theoretical extension rather than a free-standing new construct.

The paper advances that extension as the Augmented Behavioural Capacity Model, hereafter the ABC Model. The label “capacity” is preferred to “control” because the model is not primarily about classical perceived behavioural control in the sense of the TPB. It concerns tool-contingent perceived cognitive capacity: what an investor feels able to understand, compare, simulate, or execute because an LLM is present as a conversational reasoning partner during the decision episode. The central measurable mechanism is Perceived Cognitive Assistance (PCA), the investor’s felt expansion of cognitive capability at the moment of decision (Gimmelberg & Ludviga, 2025, 2026a, 2026c). The model is summarised as a moderated process: LLM engagement may raise PCA; PCA may shift proximal behavioural choice toward greater complexity, risk assertion, or decision-structure elaboration; and Calibration conditions how that movement should be interpreted.

The model foregrounds two linked but distinct claims, stated here in plain terms and specified formally in Section 4. Claim A, the behavioural-migration claim, proposes that LLM engagement may increase PCA and that elevated PCA may move proximal behavioural choice toward greater strategic complexity, risk assertion, or decision-structure elaboration. Claim B, the calibration overlay claim, proposes that Calibration - the match between perceived capacity and objective comprehension, confidence accuracy, verification discipline, reliance, and offloading - conditions whether PCA-driven movement is best read as calibrated enhancement or behavioural distortion. Enhancement in this sense does not mean better realised results or even better ex ante decision quality. It means that elevated perceived cognitive capacity is not merely a self-reported illusion, because it is matched by objective understanding, appropriate confidence, disciplined verification, and retained independent reasoning. Whether such calibrated enhancement also corresponds to higher ex ante decision quality is a separate criterion-validation question for subsequent testing.

The contribution is deliberately bounded in scope and in evidential status, and the two boundaries should be read together. In scope, ABC is offered as a domain-bounded theoretical model of LLM-assisted retail-investor decision-making, not as a universal theory of artificial-intelligence adoption, financial behaviour, or market efficiency. In evidence, the paper does not claim that LLM use improves realised investment returns, which lie outside the model’s explanatory target, and it does not claim full causal validation of every pathway the architecture implies. What is claimed is the theoretical specification of the ABC Model together with partial empirical anchoring of PCA as its central Capacity-layer mechanism. Controlled causal testing, calibration diagnostics, longitudinal refinement, and behavioural-footprint assessment are stated as future validation tasks rather than completed evidence. The remainder of the article follows the journal’s reporting structure, adapted to a model-development study. Section 2 specifies the model-development method and falsification discipline. Section 3, Section 4 and Section 5 develop and present the model, which is the study’s result: Section 3 establishes the phenomenon and the incumbent-theory gap that motivate it; Section 4 sets out the ABC architecture, construct registry, and claims; and Section 5 states the boundary conditions and falsifiable propositions that delimit it. Section 6 discusses and concludes on the contribution, limitations, and evidence.

2. Materials and Methods: Theoretical Model Development

2.1. Theoretical Model-Development Approach

This is a theoretical model-development paper, and its method is the disciplined construction of a domain-bounded model. A model becomes theoretically useful only when it specifies the phenomenon it explains, the constructs involved, the logical relationships among them, the proposed mechanism, and the boundary conditions under which the explanation is expected to hold (Bacharach, 1989; Dubin, 1978; Wacker, 1998; Whetten, 1989). Theory and model are not treated as interchangeable: a theory is a broader explanatory system of constructs and propositions, while a model is a domain-bounded representation of that explanation in a specified setting, with explicit constructs, links, mechanisms, assumptions, and limits (Bacharach, 1989; Reynolds, 1971; Wacker, 1998). On this basis the ABC Model is advanced as domain-bounded rather than universal.

This distinction sets the standard for model development: theoretical argument integrates intellectual ancestry, propositions, evidence, and relationship logic into a coherent explanatory account of constructs, mechanisms, and boundaries (DiMaggio, 1995; Sutton & Staw, 1995; Weick, 1989; Whetten, 1989).

Three safeguards follow from this logic and discipline the construction throughout. First, the model fixes its level and unit of analysis, because moving between an individual belief, a decision episode, a repeated behavioural pattern, and a market-level outcome without justification produces category errors and ecological or atomistic fallacies (Bacharach, 1989; Klein, Dansereau, & Hall, 1994; Kozlowski & Klein, 2000; Rousseau, 1985). Second, every model element is classified by its theoretical role before propositions are written, because a construct, latent factor, indicator, antecedent, mechanism, moderator, diagnostic variable, outcome, boundary condition, and extension variable are not interchangeable (Edwards & Bagozzi, 2000; MacKenzie, Podsakoff, & Podsakoff, 2011; Podsakoff, MacKenzie, & Podsakoff, 2016; Suddaby, 2010). Third, each link is assigned a formal status - causal, mediating, moderating, sequential, reciprocal, or diagnostic - rather than asserted as mere association, since these alternatives imply different theoretical claims and different empirical tests (Baron & Kenny, 1986; MacKinnon, 2008; Pearl, 2009; Whetten, 1989). These safeguards are necessary precisely because ABC combines inherited adoption constructs, a validated cognitive-assistance construct, calibration diagnostics, behavioural outcomes, and a downstream footprint extension, which could otherwise be misread as equivalent factors.

2.2. Model-Development Sequence Used in This Article

The article follows the same development sequence used to build the model, because each step depends on the ones before it: a model cannot be defensibly tested before its constructs are defined, its level of analysis fixed, its relationship types stated, and its limits made explicit (Bacharach, 1989; Dubin, 1978; Suddaby, 2010; Wacker, 1998). The sequence proceeds from defining the problem and contribution, to stating the development method, establishing the phenomenon and the theory gap, fixing scope and unit of analysis, classifying constructs and model elements, specifying the architecture and relationship logic, stating boundary conditions and rival explanations, deriving falsifiable propositions, and finally stating the contribution, limitations, and evidence status (Corley & Gioia, 2011; Lakatos, 1970; Popper, 1959; Sutton & Staw, 1995; Weick, 1989; Whetten, 1989). Figure 1 presents this sequence as a nine-step model-development flow.

The flow records the logical order in which the model is built and reported, not a claim of empirical completion at each step: the steps that define, classify, and specify the model are completed to the level required for theoretical anchoring, whereas the proposition-development step separates claims already anchored by completed PCA measurement validation from those that remain open to future causal, calibration, longitudinal, and behavioural-footprint testing. This distinction matters because PCA is validated as the model’s central measurement backbone, while full causal validation of the entire ABC architecture is not claimed. The full step-by-step development workflow and its methodological authorities are documented in Appendix A.

2.3. Falsification Logic and Evidence-Status Discipline

The model is disciplined by a refutation-oriented logic adapted from the philosophy of science. Two commitments are operative. First, core claims are distinguished from auxiliary, developmental, and extension claims, so that what would weaken or refute the model’s centre is not hidden among peripheral commitments; refutation conditions are stated at the level of the architecture as well as at the level of individual propositions (Lakatos, 1970; Popper, 1959; Chiambaretto, Fernandez, & Le Roy, 2025; Rubin, 2025). Second, the relationship logic specifies not only that constructs are related but how, why, and within what boundaries, which is a precondition for determinate theoretical testing rather than open-ended association (Pearl, 2009; Whetten, 1989; Busse, Kach, & Wagner, 2017; von Nordenflycht, 2023; Carton, 2025). Auxiliary modifications that merely absorb disconfirming observations without clarifying the core or generating testable refinement are treated as a failure of discipline rather than a defence of the model (Lakatos, 1970; Rubin, 2025).

Applied here, this discipline produces a precise and limited claim. When a phenomenon is theoretically novel and its constructs are still being stabilised, the defensible contribution is to define the phenomenon, specify the model, validate the central measurement mechanism, and state a falsifiable agenda for subsequent tests, rather than to claim a full causal test of every pathway in a single study (Edmondson & McManus, 2007; Lynham, 2002; Van de Ven, 2007). Scale-development and construct-validation guidance makes the same point: content validation, factor-analytic confirmation, discriminant testing, and early predictive evidence can establish a measurement backbone without by themselves validating an entire causal system (Clark & Watson, 2019; Cronbach & Meehl, 1955; MacKenzie, Podsakoff, & Podsakoff, 2011; Messick, 1995). The article therefore claims theoretical specification and partial empirical anchoring, and explicitly not full causal validation; failed or untested elements are reported as such rather than deferred into vague future work.

3. The LLM Phenomenon and the Incumbent-Theory Gap

3.1. LLMs as Evaluative-Stage Participants

LLMs differ from the technologies against which inherited adoption models were calibrated. They are dialogic rather than static, generative rather than retrieval-bound, explanatory rather than merely informative, context-retaining across turns, and capable of participating in the framing of the problem itself (Handler, Larsen, & Hackathorn, 2024; Sabbah & Li, 2025; Tankelevitch et al., 2024; Wang, Lu, & Pan, 2025). Recent work extends the classical extended-mind and scaffolding perspective specifically to LLMs, treating them as interactive cognitive artefacts and scaffolds rather than tools whose relevant properties are fixed before use (Heersmink, de Rooij, Clavel Vázquez, & Colombo, 2024; Kudina, Ballsun-Stanton, & Alfano, 2025; Smart, Clowes, & Clark, 2025). These properties matter because the locus of theoretical interest is the evaluative stage: interpretation, comparison, risk estimation, perceived capability, and perceived manageability of the action. It is at this stage, and not only at the point of adoption, that an LLM can alter the content of an investor’s reasoning (Jakesch, Bhat, Buschek, Zalmanson, & Naaman, 2023; Si et al., 2024).

This locus is consequential in retail finance specifically, because retail investors are documented to trade on overconfidence, to under-diversify, and to be sensitive to information presentation and financial-literacy constraints (Agnew & Szykman, 2005; Barber & Odean, 2000, 2001; Lusardi & Mitchell, 2014). A system that can elaborate, reframe, and appear to validate a strategy at the moment of evaluation engages exactly the cognitive and affective vulnerabilities that this literature identifies, which is why the evaluative stage rather than the adoption decision is the model’s point of entry.

This contrast is summarised visually in Figure 2 across six evaluative-stage dimensions that distinguish LLMs from ordinary adoption-relevant tools.

3.2. Necessary but Incomplete Inherited Theories

The inherited frameworks remain necessary. TAM, UTAUT explain access, perceived usefulness, ease of use, social influence, facilitating conditions, and continued use (Davis, 1989; Venkatesh et al., 2003, 2012), and TPB explains how attitude, subjective norm, and perceived behavioural control combine into intention (Ajzen, 1991). UTAUT2 is the member of this family closest to the present setting: it extends UTAUT to the consumer context with hedonic motivation, price value, and habit (Venkatesh et al., 2012), and retail investors using a consumer technology fall squarely within its scope. A standing critical tradition has nonetheless questioned how far these frameworks travel: commentators have asked whether TAM has reached the limits of its explanatory yield (Bagozzi, 2007; Benbasat & Barki, 2007; Lee, Kozar, & Larsen, 2003; Straub & Burton-Jones, 2007), debated whether TPB should be retired or retained (Ajzen, 2015; Conner, 2015; Sniehotta, Presseau, & Araújo-Soares, 2014), and more recently asked whether generative artificial intelligence (AI) strains the acceptance paradigm itself (Davis & Granić, 2024; Mogaji, Viglia, Srivastava, & Dwivedi, 2024; Schittko, Planing, & Müller, 2026). The point ABC takes from this tradition is narrow: not that the inherited frameworks are wrong, but that they do not specify the tool-contingent perceived cognitive capacity that operates inside the evaluative episode.

What the critique tradition establishes is that the inherited frameworks are reaching the limits of their original formulations; what it does not supply is a positive specification of the decision-time mechanism that LLMs introduce. The published critiques diagnose strain - in TAM’s explanatory yield, in TPB’s coverage of behaviour beyond intention - but they do not name the construct that would carry the missing explanation. ABC’s contribution begins where those critiques stop: rather than retiring or defending the inherited frameworks, it accepts them at the adoption boundary and specifies the evaluative-stage mechanism they leave unaddressed.

3.3. The Three-Layer Gap: Temporal, Ontological, Phenomenological

The insufficiency can be diagnosed at three layers. The temporal layer is a substrate problem: the inherited theories were calibrated against stable, determinate, single-function tools, whereas LLMs are probabilistic, dialogic, and non-stationary across model generations, so constructs designed for fixed-function software do not straightforwardly transfer (Mogaji et al., 2024; Ouyang, Wu, Jiang, Almeida, Wainwright, Mishkin, et al., 2022). The ontological layer concerns where cognition is located: inherited models place reasoning inside the user and treat the technology as an evaluated external object, whereas LLMs partly co-produce the reasoning, a configuration the distributed-cognition and scaffolding tradition anticipates but the acceptance tradition does not encode (Clark & Chalmers, 1998; Hutchins, 1995; Risko & Gilbert, 2016; Wood, Bruner, & Ross, 1976). The phenomenological layer concerns mechanisms that arise in LLM-mediated interaction and have no clean home inside inherited constructs.

Eight such mechanisms can be named in compressed form. Hallucination is the generation of fluent but unfounded content (Ji, Lee, Frieske, Yu, Su, Xu, et al., 2023). Sycophancy is the tendency of dialogue-tuned systems to agree with or flatter the user’s framing, a known by-product of preference-based fine-tuning (Malmqvist, 2024; Ouyang et al., 2022; Sharma, Tong, Korbak, Duvenaud, Askell, Bowman, et al., 2024). Anthropomorphic dialogic interaction elicits social responses to a non-social system (Epley, Waytz, & Cacioppo, 2007; Liew, Lim, Khan, & Tan, 2025; Nass & Moon, 2000). Stochastic non-determinism means the same prompt can yield materially different outputs across runs, so the system the investor evaluates is not fixed. Epistemic opacity with interactional transparency means the system reads as clear and confident while its internal basis is unavailable to the user (Glikson & Woolley, 2020). Asymmetric trust calibration is the mismatch between warranted and actual reliance documented in the automation and advice-taking literatures, where users variously over-rely on or discount machine advice (Dietvorst, Simmons, & Massey, 2015; Lee & See, 2004; Logg, Minson, & Moore, 2019; Parasuraman & Manzey, 2010). Context-window adaptation means the option set shown at a later turn is shaped by framings introduced earlier in the same conversation. Co-construction of the option set is the user-side counterpart: alternatives, criteria, and sometimes attitudes are produced inside the dialogue rather than brought to it, so that an attitude TPB would treat as a stable antecedent may instead be generated within the episode (Jakesch, Bhat, Buschek, Zalmanson, & Naaman, 2023).

The argument here is not that every mechanism lacks intellectual ancestry. Several have clear lineages in metacognition, automation, and distributed cognition (Alter & Oppenheimer, 2009; Lee & See, 2004; Risko & Gilbert, 2016; Rozenblit & Keil, 2002). The argument is that the LLM-mediated configuration activates these mechanisms jointly, inside a single evaluative episode, in a way the inherited models can only patch fragment by fragment by appending one moderator at a time. Perceived ease of use illustrates the cost concretely: calibrated against determinate-output tools, it tracks the friction of obtaining a response but not the difficulty of verifying it, so under LLMs it can register highest in exactly the conditions of greatest over-reliance risk, leaving the user no within-construct signal to separate the two (Buçinca, Malaya, & Gajos, 2021; Heersmink et al., 2024; Vasconcelos et al., 2023). Figure 3 compresses this diagnosis: it aligns the temporal, ontological, and phenomenological layers with the specific feature of LLM-mediated reasoning that inherited adoption and intention theories were not built to carry, and is the basis for treating the gap as structural rather than incremental. The expanded critique tradition and the full mechanism-by-mechanism support are documented in Appendix A.

3.4. Construct Proliferation and the Implication for a Capacity–Calibration–Choice Model

A further symptom supports the diagnosis. The construct-proliferation audit by Gimmelberg and Ludviga (2026b) reinforces this diagnosis by showing that the neighbouring construct space is theoretically active and useful for boundary-setting, while still pointing toward the need for a new theoretical model. The human–AI and GenAI behaviour literature has produced a rapid expansion of partially overlapping constructs - AI self-efficacy and literacy, cognitive absorption, metacognitive demand and workload, task–technology fit, perceived understanding, and dependence or over-reliance, among others (Almeida Lima, Bellei, Ballesteros Martins, & Terlizzi, 2024; Goh, Hartanto, & Majeed, 2025; Huy, Nguyen, Vo-Thanh, Thinh, & Thi Thu Dung, 2024; Sarraf, Kar, & Janssen, 2024; Tankelevitch et al., 2024; Wang, Lu, & Pan, 2025; Wang, Rau, & Yuan, 2023; Wang & Chuang, 2024; Xue, Ghazali, & Mahat, 2025). Read through the construct-clarity lens, this proliferation is itself evidence of theoretical strain: when a phenomenon outruns its inherited constructs, the field tends to multiply neighbouring constructs rather than respecify the mechanism (Shaffer, DeGeest, & Li, 2016; Suddaby, 2010). The implication ABC draws is that the appropriate response is not a further loose construct appended to the inherited frameworks, but a domain-bounded model that names the construct doing the explanatory work, the apparatus that conditions its interpretation, and the proximal outcome family in which the effect appears.

4. Results: The ABC Model: Construct Registry, Architecture, and Claims

4.1. Scope and Unit of Analysis

The unit of analysis is the evaluative decision episode: a self-directed retail investor using an LLM to construct, weigh, or revise a candidate trading strategy, at or near the moment of a cognitively demanding decision. The scope is bounded accordingly. The model addresses retail-investor decisions, not professional, institutional, or fully advised contexts; cognitively demanding tasks, not routine lookup or fully delegated robo-advisory allocation; and proximal choice, not realised returns. Two further bounds are stated as conditions of theoretical applicability rather than as limitations. The model is formulated against the dialogue-tuned, instruction-following LLM capability frontier of 2024–2026; parameters tied to perceived competence - in particular the level and calibration of PCA - are temporally fragile as hallucination rate, sycophancy intensity, reasoning depth, and context-window size move across releases, so findings produced against any one release should be re-anchored as the frontier advances. Parameters may also differ across jurisdictions where LLM-based advisory output is classified or regulated differently, so the present formulation is bounded to its current cultural and regulatory context rather than asserted as cross-jurisdictionally invariant. Empirically, the model is formulated for cross-sectional testing at the episode level, complemented by longitudinal tracking of trust calibration, reliance escalation, and skill atrophy; specific operationalisation choices are a separate empirical question and are not constitutive of the theoretical claim. Fixing the unit at the decision episode locates the model’s elements within a bounded context without collapsing them into a single sequential mediation chain, which matters because the elements occupy different theoretical roles.

4.2. Construct and Element Registry

ABC is organised around five elements, classified by theoretical role rather than treated as five equivalent factors. Adoption and Engagement is the inherited upstream boundary: the conditions of access, uptake, and use carried over from TAM, UTAUT and UTAUT2, and TPB, accepted rather than re-tested here (Ajzen, 1991; Davis, 1989; Venkatesh et al., 2003, 2012). Capacity is the ABC core construct, operationalised as a single latent construct through PCA, the felt expansion of cognitive capability at decision time. PCA measures the user’s perceived expansion of cognitive capability in the presence of an LLM; the extended-mind and scaffolding tradition supplies the motivation for why such a construct should exist as a separate measurement target, but ABC does not claim, at the level PCA measures, that cognition is empirically distributed across the user–system boundary. Calibration is a heterogeneous diagnostic and moderating apparatus, not a unified construct, comprising objective understanding, perceived understanding, overprecision, reliance and deference, and state offloading (Lee & See, 2004; Moore & Healy, 2008; Risko & Gilbert, 2016; Rozenblit & Keil, 2002). Choice is a family of proximal behavioural outcomes - complexity intentions, strategy evaluation, structural complexity of intended trades, and risk assertion - rather than a single dependent variable. Behavioural Footprint is a downstream extension into realised trading behaviour, documented but not part of the defended core.

The structural asymmetry across the three Cs is a substantive feature, not a presentational convenience: Capacity is a construct, Calibration is a diagnostic and moderating apparatus, and Choice is an outcome family. Treating the three as parallel latent factors would misrepresent the architecture and reproduce the construct-confusion the model is meant to resolve. The full operational indicators and the formal treatments are documented in Appendix A.

4.3. The ABC Architecture

The architecture is a five-layer process. The three core stages - Capacity, Calibration, and Choice - are the model’s theoretical claims; the two framing stages - Adoption and Engagement upstream, Behavioural Footprint downstream - are boundary architecture. Defending an asymmetric core follows directly from theory-building practice, which requires a model to mark what it claims, what it inherits, and what it extends, with the boundaries between those zones visible rather than concealed (Bacharach, 1989; Suddaby, 2010; Wacker, 1998; Whetten, 1989). Five layers, rather than three or four, are the minimum that lets the inherited adoption tradition appear upstream and the empirical-extension question appear downstream while keeping both at the boundary: a three-layer core would silently inherit adoption conditions without naming them, and a four-layer version would name one boundary while concealing the other.

Three core concepts, rather than four or five, are retained because each names a distinct kind of theoretical object: Capacity is the construct, Calibration the diagnostic gate, and Choice the manifestation. Other candidate “Cs” that recur in the literature reduce to one of these three or sit outside the core. Confidence and comprehension are not separate stages but components of Calibration’s battery, operationalised as overprecision and as objective understanding (Moore & Healy, 2008; Rozenblit & Keil, 2002). Control is an inherited neighbouring belief from the TPB and UTAUT traditions rather than an ABC core stage; ABC theorises the narrower tool-contingent cognitive-capacity subset that the inherited control construct does not capture (Ajzen, 1991; Venkatesh et al., 2003). Complexity functions as a modulator of Choice and a task-bounding scope condition rather than a stage of its own, and conduct is downstream, at the Footprint boundary. The three Cs are thus the minimum that distinguishes the construct, the gate, and the manifestation; further Cs are not suppressed but located where they belong.

Within the core, the relationship logic is a moderated decision process. PCA captures perceived cognitive capacity; Calibration conditions how that perceived capacity translates into proximal choice. Calibration is specified primarily as a moderator rather than as a second latent construct, because the model’s own language - high PCA matched by objective understanding supports calibrated enhancement, unmatched supports behavioural distortion - is moderation language, and because the moderator form is the most statistically tractable for an early-stage programme and does not foreclose a later latent-class specification if subgroup discontinuities emerge (Aiken & West, 1991; Baron & Kenny, 1986; Hayes, 2018; MacKinnon, 2008; Mathieu & Taylor, 2006; McCutcheon, 1987). Because the five calibration variables are not directionally homogeneous, the apparatus is modelled as a set of sign-aligned miscalibration gaps, with a configural component-specific specification retained as an alternative; a naive additive composite is excluded, because it can mix adequacy and distortion signals and produce an uninterpretable null through cancellation. Figure 4 displays the five-layer architecture, the three-stage core, and the placement of Claim A and Claim B.

4.4. Claim A and Claim B

Claim A (behavioural-migration claim) specifies the theoretically proposed Capacity→ Choice path: LLM engagement may increase PCA, and elevated PCA may shift proximal behavioural choice toward greater strategic complexity, risk assertion, or decision-structure elaboration.

Claim B (calibration overlay claim) specifies the overlay sitting above that path: the match between perceived cognitive capacity and its objective referents across the five calibration diagnostics - objective understanding, perceived understanding, overprecision, reliance and deference, and state offloading - conditions whether the same PCA-driven movement is interpreted as calibrated enhancement or behavioural distortion. Calibrated enhancement means that elevated PCA is matched by objective understanding, confidence accuracy, verification discipline, and retained independent reasoning. Behavioural distortion means that elevated PCA is paired with weak understanding, overprecision, uncritical deference, perceived understanding without objective comprehension, or unverified offloading.

The two claims are linked but distinct. Claim A concerns movement toward complexity; Claim B concerns the calibration-contingent interpretation of that movement. Claim B does not claim that enhancement produces better decisions. It claims that elevated PCA can be classified as calibrated enhancement rather than illusion or distortion when the relevant diagnostic referents are present. Whether calibrated enhancement then corresponds to higher ex ante decision quality is a separate criterion-validation test, not a premise of the classification itself. The model therefore does not posit two separate causal mechanisms so much as a single Capacity → Choice mechanism whose behavioural meaning changes across calibration conditions. Claim B is, at the moderator level, construct-general: whatever raises PCA, its onward effect on Choice is moderated by calibration, so Claim B takes elevated PCA as given rather than tracing its origin.

4.5. Current Evidentiary Status

The two claims are anchored asymmetrically, and the asymmetry is reported rather than smoothed. The Capacity layer is the most developed: PCA has been content-validated against perceived usefulness, perceived ease of use, trust in the LLM, and trading self-efficacy, confirmed as a one-factor scale with adequate reliability and scalar invariance across trading-experience and recency strata, and shown to predict complexity intentions incrementally over inherited proximal comparators, with the PCA–usefulness boundary close but defensible on confirmatory and discriminant criteria (Gimmelberg & Ludviga, 2026a, 2026c). Claim A is therefore partly anchored at the component level - through PCA measurement, one-step incremental prediction, and discriminant validity - but its end-to-end directed path is not confirmed as a mediation chain. Claim B is theoretically specified, with a stated architectural refutation condition, but is not operationally tested, because the matched-domain distortion-diagnostic battery it requires - objective comprehension, overprecision, perceived understanding, reliance and deference, and state offloading at the resolution of the decision episode - is not yet stabilised as a coherent instrument package in retail-investor research. The test of Claim B is therefore a calibration-classification test: whether elevated PCA is matched by objective and behavioural diagnostics of understanding, confidence accuracy, verification, and retained reasoning, or instead by diagnostics of overprecision, deference, and offloading. A later link from calibrated enhancement to higher ex ante decision quality would provide criterion-validity evidence for the classification, but the model does not claim that link in the present article.

5. Boundary Conditions and Falsifiable Propositions

5.1. Boundary Conditions

Boundary specification is part of theoretical precision rather than supplementary limitations prose: a model that holds for all values of all variables is empirically empty, so the conditions under which propositions are asserted must be stated explicitly (Bacharach, 1989; Busse, Kach, & Wagner, 2017; Dubin, 1978; Whetten, 1989). Five boundaries fix the value-ranges across which the ABC propositions are asserted. The number five is model-derived rather than method-prescribed: these are the five points at which ABC’s propositions can be misapplied - phenomenon, population, task context, mechanism, and outcome - while cultural/regulatory context and model-generation era are treated as generalisability questions rather than additional proposition boundaries. The phenomenon boundary restricts the claims to LLM-assisted evaluative reasoning, not technology adoption in general; findings bearing only on adoption, intention, or use fall outside the range. The population boundary restricts them to self-directed retail investors, not professional, institutional, or fully advised actors. The task-context boundary restricts them to cognitively demanding trading decisions, not routine lookup or fully delegated allocation; that low-complexity tasks show weaker effects is itself a proposition rather than a free parameter. The mechanism boundary restricts them to perceived cognitive capacity via PCA, moderated by Calibration, and not to objective ability, skill, or performance; the model’s distinctive predictions concern PCA conditional on calibration, not PCA alone. The outcome boundary restricts them to proximal behavioural choice, not realised execution or return. A contradicting finding outside these ranges is not a refutation; a contradicting finding inside them is.

5.2. Rival-Explanation Logic

The boundary step also asks whether neighbouring accounts could make ABC redundant, mis-scoped, or empirically misleading, following the convergent–discriminant and falsificationist traditions (Campbell & Fiske, 1959; Edwards & Bagozzi, 2000; Lakatos, 1970; Popper, 1959). A construct-proliferation audit (Gimmelberg & Ludviga, 2026b) classified candidate rivals into four threat families - direct substitutes, antecedents and confounders, distortion mechanisms, and background theories - because each requires a different empirical response rather than a single test. Seven external rival families remain on record for boundary treatment: AI self-efficacy, cognitive absorption, cognitive load and workload, illusion of explanatory depth, AI dependence and over-reliance, task–technology fit, and explanation quality or perceived understanding (Agarwal & Karahanna, 2000; Goh, Hartanto, & Majeed, 2025; Goodhue & Thompson, 1995; Leppink et al., 2013; Rozenblit & Keil, 2002; Wang & Chuang, 2024). A substantive result of the audit is that several of these are currently blocked from direct same-level psychometric contest, because matched-domain instruments do not yet exist for retail-investor LLM decision episodes; stating that blockage openly, rather than presenting only the deployable contests, is the discipline the falsification logic requires (Lakatos, 1970). The full audit and rival matrix are documented in Appendix A.

5.3. Architecture-Level Refutation Conditions

Four refutation conditions are stated at the level of the architecture, so that what would refute the model’s core is visible rather than buried in a longer list (Lakatos, 1970; Popper, 1959; Bacharach, 1989; Wacker, 1998). First, construct redundancy: the model fails as a contribution if PCA cannot be distinguished from its proximal comparators - perceived usefulness, ease of use, trust, and trading self-efficacy - on factor-analytic or predictive grounds under stronger designs; closeness is expected, absorption is fatal (Cronbach & Meehl, 1955; Campbell & Fiske, 1959; MacKenzie et al., 2011; Shaffer et al., 2016). Second, misclassification of rivals: the model is weakened both by ignoring the external rival families and by testing any of them as a direct comparator before a matched-domain instrument exists, because rival explanations must be classified by theoretical role, level, and measurement parity before they can adjudicate construct boundaries (Bacharach, 1989; Suddaby, 2010; MacKenzie et al., 2011; Podsakoff et al., 2016). Third, failure of the calibration overlay: the model is weakened at the architectural level if PCA predicts the same proximal-choice pattern regardless of calibration diagnostics - that is, if calibration does not moderate the PCA→Choice relationship - a condition that depends on the miscalibration-gap specification, since a naive additive composite could manufacture a null through cancellation (Aiken & West, 1991; Hayes, 2018; Mathieu & Taylor, 2006; Diamantopoulos & Winklhofer, 2001; Jarvis et al., 2003). This condition is currently programmatic rather than deployable, because the required battery does not yet exist. A later failure to link calibrated enhancement with independently scored decision quality would limit the criterion validity of the classification, but it would not by itself refute the calibration classification claim. Fourth, failure to survive calibration-only alternatives: if calibration diagnostics jointly predict choice while PCA adds no residual role and does not interact with them, PCA is at most a surface appraisal accompanying calibration rather than the mechanism through which engagement becomes capacity; this condition follows from treating calibration-only specification as an architectural rival and testing whether the focal mechanism survives competing structural explanations (Pearl, 2009; MacKinnon, 2008; Whetten, 1989).

5.4. Falsifiable Propositions and Validation Agenda

The audit is translated into seven propositions in three classes - Core (P1–P5), Developmental (P6), and Extension (P7) - each with a determinate refutation condition and a required empirical operation (Lakatos, 1970; Popper, 1959). P1 holds that interactive LLM support raises PCA above static or non-dialogic tools and is the antecedent precondition for Claim A. P2, the component-level test of Claim A, holds that PCA predicts proximal choice incrementally over inherited TAM/UTAUT/TPB comparators. P3 holds that task complexity moderates the PCA→Choice path. P4, a discriminant prerequisite shared by both claims, holds that PCA remains distinguishable from perceived usefulness, dependence, and passive offloading. P5, the direct test of Claim B, holds that elevated PCA is interpreted as calibrated enhancement when matched by objective understanding, confidence accuracy, verification discipline, and retained independent reasoning, and as behavioural distortion when paired with weak understanding, overprecision, uncritical deference, perceived understanding without objective comprehension, or unverified offloading. P6 holds that proficiency and adoption maturity strengthen the productive pathway and reduce distortion. P7 extends the claims into realised trading behaviour at the Footprint boundary (Barber & Odean, 2000, 2001).

The evidence status differs across propositions and is reported separately rather than aggregated. P1 and P3 are open; P2 and P4 are partly anchored by completed PCA validation work; P5 is theoretically specified but instrument-blocked because the matched-domain calibration-diagnostic battery does not yet exist; P6 is developmental; and P7 is an extension. A later decision-quality test may provide criterion-validity evidence for P5, but it is not treated as a current theoretical claim.

The discipline of the set is shown by what it declines to claim: a free-standing mediation proposition of the form LLM use → PCA → choice is deliberately omitted from the core list, because that chain was tested cross-sectionally and not detected, with the bias-corrected indirect-effect interval including zero (Gimmelberg & Ludviga, 2026c). The model treats this non-detection as a constraint on Claim A’s status rather than absorbing it through auxiliary moderation; whether the path is misspecified or merely attenuated in early-stage adoption is left open (MacKinnon, 2008; Mathieu & Taylor, 2006; Quintana, 2024), and P3 and P6 are advanced as independently motivated propositions, not as recovery routes for the unconfirmed mediation. The full proposition-and-refutation matrix is documented in Appendix A.

6. Discussion and Conclusions

The Results sections should be read together as theoretical results of the model-development process rather than as empirical outcome results. The first result is diagnostic: LLM-assisted retail-investor decision-making exposes a gap in inherited adoption and intention frameworks because the system enters the evaluative stage of judgement rather than remaining only an object of use. The second result is architectural: ABC specifies that gap through a Capacity–Calibration–Choice structure, with Perceived Cognitive Assistance as the Capacity-layer mechanism and Calibration as the diagnostic overlay that gives behavioural meaning to PCA-driven movement. The third result is disciplinary: the model is bounded by explicit scope conditions, rival explanations, refutation conditions, and propositions. Taken together, the results position ABC not as proof that LLMs improve investor decisions, but as a specified architecture for studying how dialogic AI assistance may alter perceived decision-time capacity and how that alteration should be interpreted.

The ABC Model contributes a domain-bounded theory of LLM-assisted retail-investor evaluative decision-making. Its theoretical move is to name the mechanism that inherited adoption and intention frameworks reach but do not specify — tool-contingent perceived cognitive capacity, measured as PCA — and to make the interpretation of that capacity conditional on calibration rather than monotonically beneficial. In doing so it reformulates an assumption shared across TAM, UTAUT, UTAUT2 and TPB, that more positive perception of a technology is uniformly beneficial, into a calibration-contingent distinction between calibrated enhancement and behavioural distortion (Ajzen, 1991; Davis, 1989; Venkatesh et al., 2003; Venkatesh et al., 2012). Motivating that theoretical move is a separate diagnostic contribution: a three-layer specification - temporal substrate, ontological cognition-locus, and phenomenological mechanism set - of why inherited adoption theories cannot host LLM-mediated evaluative reasoning. Existing taxonomies of LLM-specific phenomena are risk-organised (Weidinger et al., 2022) or determinant-organised (Eigner & Händler, 2024), and do not group mechanisms by the measurement-boundary crossings of inherited adoption constructs that the eight-mechanism phenomenological layer maps.

This interpretation has two implications for inherited theory. First, ABC does not reject TAM, UTAUT, UTAUT2, or TPB. It relocates them. They remain appropriate for explaining access, adoption, use, usefulness, effort, social influence, control, and intention. Their limitation appears when the theoretical object shifts from adoption of a tool to evaluative reasoning with a dialogic system. Second, ABC treats positive perception of LLM assistance as theoretically ambiguous. Higher perceived cognitive capacity is not assumed to be beneficial by itself. Its interpretation depends on whether perceived capacity is matched by objective understanding, confidence accuracy, verification discipline, and retained independent reasoning. The same apparent movement toward greater strategy complexity can therefore have different behavioural meanings under different calibration conditions.

The practical relevance follows from that distinction. By separating calibrated enhancement from miscalibrated confidence, the model supplies a vocabulary and a falsifiable frame for product, platform, and regulatory audiences deciding how dialogic AI assistance should be embedded in retail-trading interfaces, and PCA offers a deployable instrument for that work. The methodological contribution is to state boundaries, rival explanations, and architecture-level refutation conditions explicitly, rather than to add another loose construct to an already proliferating field (Corley & Gioia, 2011; Shaffer et al., 2016; Suddaby, 2010; Whetten, 1989).

The limitations are intrinsic to the contribution and are stated as such. The model is domain-bounded and does not transfer, without further work, to professional populations, non-financial tasks, or different regulatory and cultural settings (Bacharach, 1989; Busse et al., 2017; Wacker, 1998). It claims no improvement in realised returns or investor welfare, which lie outside its defended core. Its empirical anchoring is moreover a snapshot in a non-stationary landscape, because both the LLM substrate and retail-investor adoption are evolving rapidly (Heersmink et al., 2024; Mogaji et al., 2024; Ouyang et al., 2022); the architecture is intended to be substrate-agnostic, but specific findings will require re-anchoring as the environment matures.

The evidence status of the two claims is deliberately asymmetric. Claim A is partially anchored through PCA measurement and one-step incremental prediction, but is not confirmed as an end-to-end mediation; this non-detection is reported openly rather than explained away. Claim B is theoretically specified with a stated refutation condition, but is not yet operationally tested; its immediate test is whether elevated PCA can be classified as calibrated enhancement or behavioural distortion using matched-domain calibration diagnostics, while any link to ex ante decision quality remains a later criterion-validation question. The validation agenda that follows is determinate: controlled causal tests of the Capacity→Choice path, stabilisation of the calibration battery and a direct test of the moderation claim, longitudinal tests of proficiency effects, and, only if the model is asked to speak to realised behaviour, assessment of the behavioural-footprint extension. The model is offered in a form built to be tested, refined, and where warranted refuted, rather than defended against evidence.