Online assessment (OA) has been advanced as a means of enhancing language proficiency through immediacy, flexibility, and formative feedback (Sumaharani et al., 2023). With the growing integration of artificial intelligence (AI), OA now includes automated pronunciation scoring, fluency monitoring, and adaptive feedback (Alshewiter et al., 2024; Liu et al., 2025), enabling personalized support at scale. These features align with the principles of Learning-Oriented Assessment (LOA), which emphasizes feedback, learner autonomy, and iterative practice as central to language development (Carless, 2014).

In vocational EFL contexts in China—where large classes, limited authentic speaking opportunities, and exam-driven curricula restrict individualized instruction—AI-augmented OA offers a practical mechanism for repeated, low-stakes oral practice. Systems such as ETS SpeechRater™, iFlytek Smart Learning Engine, and Chaoxiang AI modules provide real-time formative feedback that helps learners self-monitor pronunciation, fluency, and coherence (Gong, 2023; Thi-Nhu Ngo et al., 2024). For vocational learners, whose confidence is often fragile, these tools reduce the anxiety associated with face-to-face performance and promote continuous self-evaluation (Marjerison and Yang, 2022).

However, the evidence regarding OA’s impact remains inconclusive. Some studies demonstrate gains in fluency and self-confidence (Cao et al., 2025; Ebadi et al., 2025), while others highlight weak transfer to authentic communication when feedback is generic or when tasks lack contextual relevance (D. Zhang et al., 2023). Scholars also warn of algorithmic bias—penalization of dialectal variation, accent, or speech rate—raising concerns about the fairness and credibility of automated scoring systems (Xiong et al., 2023). In vocational EFL classrooms, these limitations may be particularly detrimental: when learners perceive feedback as mechanical or inauthentic, their trust in the system diminishes, reducing both motivation and sustained engagement (Ahmed and Sidiq, 2023).

Accordingly, OA’s effectiveness depends less on its technical sophistication and more on how learners interpret and internalize feedback as credible, competence-enhancing, and relevant to vocational goals. This study therefore focuses on self-perceived English-speaking competency (ESC)—defined as learners’ subjective evaluation of their ability to communicate effectively in English during AI-based assessments. By centering perception rather than performance, the present research captures the psychological mechanisms—motivation and engagement—through which AI-augmented assessment environments shape learning outcomes in vocational education.

If OA is to enhance self-perceived ESC, learners must be motivated to act upon feedback and engage deeply with learning tasks. Self-Determination Theory (SDT) provides a robust framework for understanding how intrinsic and extrinsic motivation influence persistence and learning quality (Ryan and Deci, 2020). Motivation is sustained when autonomy, competence, and relatedness needs are supported. In digital EFL contexts, intrinsic motivation fosters persistence and active strategy use (Almayez, 2022; Alotumi, 2021), yet vocational education is inherently exam-driven and career-oriented, cultivating controlled motivation—driven by external incentives such as certification, employability, and institutional recognition (Csizér and Dörnyei, 2005).

Rather than undermining learning, these extrinsic forms of motivation can be productive when aligned with vocational identity and career relevance. In Chinese vocational contexts, learners often perceive language learning as instrumental to securing employment in service, logistics, or tourism sectors. Thus, feedback that emphasizes employability and skill progression can satisfy the need for competence even under constrained autonomy, reinforcing sustained effort (Zeng et al., 2025; Yuan and Dao, 2025).

Engagement acts as the behavioral and cognitive bridge between motivation and learning outcomes (Fredricks et al., 2004). It encompasses behavioral participation (effort, time on task), emotional involvement (interest, enjoyment, and persistence), and cognitive strategies (self-regulation and reflection). In digital classrooms, engagement is easily mistaken for activity metrics—logins, task completions, or time online—yet these indicators may mask pseudo-engagement (Consoli and Curle, 2024). For vocational learners, engagement quality is heavily contingent upon whether tasks feel authentic and vocationally relevant. When AI feedback explicitly connects speaking performance to workplace competencies—such as customer interaction or service explanation—learners exhibit higher persistence and deeper reflection (Nguyen et al., 2025; Vo, 2023). Complementary evidence from Hong Kong EFL classrooms indicates that hybrid teacher-plus-AI feedback produces higher learner motivation, perceived feedback quality, and performance than AI-only feedback (Lo et al., 2025a). This supports our assumption that it is the motivational climate surrounding AI-augmented assessment—rather than automation alone—that drives sustained engagement and perceived competence.

Thus, motivation (the psychological energy to learn) and engagement (the enactment of that energy) operate as sequential mediators in AI-mediated learning environments. Their interaction determines whether learners treat feedback as a mechanical process or as a meaningful opportunity for growth.

This study integrates LOA, SDT, and engagement frameworks—complemented by sociocultural and expectancy–value perspectives—to explain how OA shapes self-perceived ESC among vocational EFL learners. Each framework offers distinct but complementary insights.

Learning-Oriented Assessment (LOA) reconceptualizes assessment as part of the learning cycle, emphasizing iterative feedback, reflection, and self-regulation (Carless, 2014). In AI-based environments, OA operationalizes these principles through frequent, low-stakes speaking tasks, adaptive scoring, and immediate feedback (Sumaharani et al., 2023). Yet LOA tends to assume that learners automatically act on feedback, underplaying the motivational filters that determine feedback uptake (Lam, 2021).

Self-Determination Theory (SDT) addresses this gap by specifying the conditions under which feedback enhances motivation—namely, when it satisfies autonomy (sense of control), competence (sense of effectiveness), and relatedness (sense of belonging or connection) (Ryan and Deci, 2020). In vocational EFL settings, these psychological needs are reframed around career readiness: autonomy manifests as flexibility in pacing and task selection; competence reflects measurable improvement in employable communication; and relatedness is reinterpreted as workplace communication identity, or the learner’s perceived capacity to interact confidently in professional contexts (Almayez, 2022; Csizér and Dörnyei, 2005).

Engagement theory completes this triad by explaining how motivated learners convert psychological intent into observable behaviors. Engagement encompasses behavioral persistence, cognitive investment, and emotional commitment to learning tasks (Fredricks et al., 2004; Zhang, 2024). In AI-mediated assessment, engagement can be observed in how learners respond to automated feedback—revising responses, reflecting on progress, and sustaining interaction with the platform. When engagement is authentic, learners internalize feedback and demonstrate incremental growth; when superficial, engagement becomes digital compliance (Dao, 2024).

Taken together (Figure 1), these frameworks depict OA as a multi-level process: an instructional scaffold (LOA) that activates psychological drivers (SDT), which in turn manifest through behavioral pathways (engagement). Each theory compensates for the others’ limitations—LOA explains how assessment structures learning opportunities, SDT clarifies why learners are motivated to act on them, and engagement theory reveals when and to what extent these actions translate into perceived competence. This integrated model avoids “theory stacking” by aligning the frameworks around a shared purpose: to explain how AI-mediated assessment environments foster self-perceived speaking development under the vocational education pressures of exam accountability and employability imperatives.

For learners facing fragile proficiency, limited authentic practice, and strong institutional pressures, this framework captures both the promise and vulnerability of AI-augmented assessment. It recognizes that sustained improvement in self-perceived communicative competence depends not only on technical precision but also on motivational resonance, engagement depth, and the perceived authenticity of AI feedback. Hence, the study hypothesizes that OA is associated with self-perceived ESC primarily through motivational activation and engagement enactment, forming sequential and configurational pathways that reveal how AI-mediated feedback environments operate in China’s vocational EFL landscape.

The study involved 463 non-English major students enrolled in compulsory English courses across three public vocational colleges in Guangdong Province, China. The sampling frame was provided by the English departments, which maintain enrolment rosters for all mandatory English classes. To ensure representativeness across disciplines, a stratified cluster sampling strategy was adopted: strata were defined by academic major (business vs. technology) and year of study, and intact class sections were randomly selected within each stratum.

All students in the selected sections were invited to participate through announcements posted on the Chaoxiang learning platform and shared via official WeChat course groups. Participation was voluntary, informed consent was obtained, and no course credit or grades were affected by participation. The final valid response rate was approximately 89%, indicating strong engagement and reliability of the data source.

The sample size was deemed statistically adequate according to multiple benchmarks. A priori power analysis (G*Power 3.1; Faul et al., 2009) for multiple regression with three predictors (OA, MODT, and ENDT), assuming a medium effect size (f² = 0.15), α = 0.05, and desired power of 0.95, yielded a required minimum of 176 participants. The final N = 463 thus exceeded this threshold and satisfied the 10-times rule for complex mediation and higher-order models (Hair et al., 2022). Nonetheless, the study’s findings should be interpreted with caution regarding generalizability: the sample was limited to vocational EFL learners in Guangdong, which may not represent all regions or academic tiers within China’s broader vocational education landscape.

The structured questionnaire comprised two parts. Part I collected demographic and background information, including gender, age, major, year of study, CET band, and years of English learning. Part II measured the four focal constructs—Online Assessment (OA), Motivation in Digital Teaching (MODT), Engagement in Digital Teaching (ENDT), and English-Speaking Competency (ESC)—using previously validated multi-item scales, rated on a five-point Likert scale (1 = strongly disagree to 5 = strongly agree).

A rigorous translation and back-translation procedure ensured linguistic and conceptual equivalence (Brislin, 1986). Two bilingual lecturers translated the instrument into Chinese, and an independent translator back-translated it into English; discrepancies were reconciled through expert review to maintain semantic accuracy and cultural appropriateness. A pilot test with 30 vocational students confirmed clarity, comprehensibility, and contextual relevance to Chaoxiang’s AI-enabled learning environment, leading to minor wording refinements (Alam et al., 2025; van de Vijver and Leung, 1997).

The OA was modeled as a formative–reflective higher-order construct using the two-stage approach. It comprised four reflective dimensions—assessment value, feedback immediacy, technological integration, and task relevance/design—adapted from Sumaharani et al. (2023) and refined to capture Chaoxiang’s AI feedback functions. MODT was conceptualized as a reflective–reflective higher-order construct encompassing expectancy beliefs/self-efficacy, task value/goal orientation, and affective regulation, with nine items adapted from the Motivated Strategies for Learning Questionnaire (Pintrich and de Groot, 1990). ENDT followed Maroco et al. (2016) as a reflective–reflective construct including behavioral, emotional, and cognitive dimensions (10 items). ESC was modeled as a formative–reflective construct comprising six reflective dimensions—fluency, accuracy, vocabulary range, coherence, rhetorical organization, and expressive delivery—adapted from Zhang et al. (2019) to capture learners’ self-perceived speaking competence rather than performance-based proficiency, a distinction acknowledged in the study’s limitations.

To assess potential common method bias (CMB), a theoretically unrelated construct—“Perception of Blue”—was included as a marker variable (Podsakoff et al., 2003; Liang et al., 2007). This variable captures general aesthetic color preference and is conceptually independent from OA, MODT, ENDT, and ESC, ensuring that it serves purely as a statistical control for shared method variance. Empirically, the marker variable exhibited low, non-significant cross-loadings with focal constructs, supporting discriminant distinctiveness, while demonstrating acceptable reliability (α = 0.79) and average variance extracted (AVE = 0.56).

Reliability and validity of all constructs were established through Cronbach’s α > 0.70, Composite Reliability > 0.70, Average Variance Extracted > 0.50, and HTMT < 0.85 (Hair and Alamer, 2022). Multicollinearity among formative indicators was assessed using VIF < 3.0, confirming indicator independence. All final measurement items, including the marker-variable items used for the CMB analysis, are provided in Appendix A along with descriptive statistics, reliability coefficients, and factor loadings to ensure transparency and replicability.

In the studied colleges, AI-based OA was implemented primarily through the Chaoxiang Learning Platform, an institutional digital ecosystem used for blended English instruction. The platform’s AI modules employ automatic speech recognition (ASR) and machine scoring engines similar to ETS SpeechRater™. Students typically complete three to five AI-assisted speaking tasks weekly, including read-alouds, short monologic responses, and situational dialogs relevant to vocational contexts such as customer service, hotel booking, or technical explanation. Each task provides immediate, machine-generated feedback on pronunciation, fluency, grammar, and prosody, accompanied by visual progress dashboards.

These tasks serve predominantly as formative practice to encourage repeated performance and self-monitoring. However, instructors occasionally integrate AI feedback scores into continuous assessment grades, increasing their perceived authenticity and accountability. The integration of Chaoxiang’s AI modules thus represents a typical instance of AI-augmented Learning-Oriented Assessment (LOA), where automated analytics support—but do not replace—teacher evaluation.

Data analysis proceeded in two major stages.

Stage 1: PLS-SEM (SmartPLS 4.1)- Measurement models were validated through internal consistency (Cronbach’s α, CR), convergent validity (AVE), and discriminant validity (HTMT). Bootstrapping with 10,000 resamples estimated path significance, indirect effects, and sequential mediation (OA → MODT → ENDT → ESC). Predictive validity was evaluated using R², Q², and f² indices. Given the cross-sectional nature of the data, all interpretations are restricted to associative relationships rather than causal inferences.

Stage 2: fsQCA (fsQCA 4.0)- To complement the linear SEM findings, fsQCA was employed to examine equifinal pathways leading to high self-perceived ESC. Likert-scale responses were calibrated into fuzzy scores using percentile-based thresholds (full membership = 0.95 quantile, crossover = 0.50, full non-membership = 0.05). The analysis first tested necessity (conditions consistently present in high-ESC cases, consistency ≥ 0.90) and then sufficiency (configurations of OA, MODT, ENDT jointly associated with high ESC). Solution consistency and coverage were reported, and core vs. peripheral conditions were identified based on raw and unique coverage values.

Combining PLS-SEM and fsQCA provided a methodological triangulation: PLS-SEM clarified average associations among constructs, while fsQCA revealed multiple sufficient condition sets leading to similar high self-perceived speaking competency outcomes. This dual analysis enriched understanding of how AI-augmented assessment environments support learner motivation, engagement, and perceived communicative development in vocational EFL settings.

The measurement model was evaluated for internal consistency, convergent validity, and discriminant validity, following established PLS-SEM guidelines (Dijkstra and Henseler, 2015; Hair and Alamer, 2022). Prior to structural analysis, confirmatory factor analysis verified that the adapted scales exhibited the intended dimensionality, providing a sound basis for the hierarchical component model adopted in this study (Figure 2).

All indicator loadings (Table 2) exceeded 0.70, demonstrating that the observed variables contributed strongly to their respective latent constructs (Fornell and Larcker, 1981). Multicollinearity was not a concern, as all indicator-level variance-inflation-factor (VIF) values were below 3.0 (Kock, 2015), including those for the dimensions composing the higher-order constructs. Composite reliability (CR) values ranged from 0.84 to 0.94, exceeding the 0.70 benchmark for internal consistency, and average variance extracted (AVE) values ranged between 0.53 and 0.78, satisfying the 0.50 threshold for convergent validity.

The OA and ESC were modeled as formative–reflective higher-order constructs (Table 3), consistent with the conceptual logic that their multidimensional facets (e.g., assessment value, feedback immediacy, and technological integration for OA; fluency, accuracy, coherence, and expressive delivery for ESC) jointly form the overall construct and are not interchangeable (Jarvis et al., 2003).

Motivation (MODT) and Engagement (ENDT) were specified as reflective–reflective higher-order constructs, because their dimensions (e.g., expectancy beliefs, task value, affective regulation; behavioral, emotional, and cognitive engagement) are viewed as manifestations of the same latent trait.

A two-stage estimation approach was employed: first extracting latent scores for lower-order components, then using them as indicators in the higher-order model. This approach minimizes indicator redundancy and is recommended for complex educational models (Becker et al., 2012).

Discriminant validity was confirmed using the Heterotrait–Monotrait (HTMT) ratio (Dijkstra and Henseler, 2015); all values were below 0.85 (Table 4), confirming satisfactory discriminant validity among constructs. The theoretically related yet distinct dimensions—such as expectancy beliefs (MODT) and behavioral engagement (ENDT)—remained statistically separable, mitigating concerns about conceptual redundancy in digital-learning contexts (Reeve and Tseng, 2011; Skinner et al., 2009).

The measurement model demonstrates strong psychometric properties. Reflective constructs show high reliability and validity, while the formative higher-order constructs are conceptually justified and empirically supported by significant indicator weights and acceptable VIF values. These results provide a robust foundation for testing the hypothesized structural relationships among OA, MODT, ENDT, and ESC in the subsequent analyses.

Given that all data were self-reported and collected using a single instrument, potential common method variance (CMV) was statistically assessed. Following the guidelines of Podsakoff et al. (2003) and Liang et al. (2007), a marker-variable approach was employed, using Perception of Blue—a theoretically unrelated construct—to capture and partial out possible method variance. This approach is widely recognized in PLS-SEM for evaluating method effects in educational and behavioral research (Hair et al., 2022).

As shown in Table 5, introducing the method factor slightly reduced the magnitude and t-values of several substantive paths (e.g., OA → MODT: t = 15.05 → 8.72; OA → ENDT: t = 9.62 → 5.44), suggesting that a portion of shared variance among constructs stems from common-method influences. The explained variance for ENDT increased from R² = 0.456 to 0.502, and for MODT from R² = 0.264 to 0.336, indicating modest method-related covariance absorbed by the marker factor.

Although several coefficients weakened slightly, the overall structural pattern remained stable. The direct path OA → ESC, significant in the baseline model (β = 0.174, p < 0.001), became smaller yet remained significant in the method-adjusted model (t = 3.142, p = 0.002), implying that while some inflation was present, the relationship persists at a modest level. More importantly, the indirect chain OA → MODT → ENDT → ESC remained strong and statistically robust, confirming that online assessment primarily enhances self-perceived English-speaking competency through motivational and engagement mechanisms rather than through a direct instructional pathway.

To assess the out-of-sample predictive capability of the structural model, the PLSpredict procedure was implemented following the guidelines of Shmueli et al. (2019) and Hair et al. (2022). This approach evaluates the model’s performance in predicting new data by comparing its prediction errors with those of naïve statistical benchmarks.

Table 6 presents the Q² predict values alongside the Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) results for the PLS-SEM model, a linear-regression benchmark (LM), and an indicator-average benchmark (IA).

All Q² predict values exceeded zero (range = 0.201–0.267), confirming predictive relevance for all indicators of self-perceived English-speaking competency (ESC). Among sub-skills, coherence/structure (CS = 0.267) and fluency (FL = 0.256) showed the highest predictive accuracy, suggesting that the model most effectively forecasts learners’ organizational control and fluency performance. This outcome aligns with Learning-Oriented Assessment (LOA) principles, which emphasize that iterative feedback and low-stakes practice cycles enhance fluency and structural competence (Carless, 2015; Lee and Lee, 2023).

Lower yet positive Q² values for rhetorical expression (RHE = 0.201) and vocabulary range (VR = 0.204) indicate that these qualitative language features are more difficult for AI-driven assessment systems to predict. This pattern supports recent findings that automated scoring models exhibit lower sensitivity to lexical diversity and pragmatic nuance in EFL speaking contexts (Donnelly et al., 2025; Seih et al., 2020). Similar asymmetries appear in AI-supported writing research: Lo et al. (2025b) found that generative-AI assistance during essay revision primarily improved surface-realization and structural coherence, with weaker effects on lexical and rhetorical development. This parallel suggests that AI-augmented assessments better capture fluency and structure than higher-order discourse features, clarifying why the present model predicts coherence and fluency most strongly.

The RMSE and MAE comparisons further corroborate the model’s robustness. Across all indicators, PLS-SEM exhibited consistently smaller prediction errors than both the linear-model and indicator-average benchmarks (e.g., FL: 0.865 vs. 0.867 vs. 1.003). Although absolute differences were modest, this consistent advantage suggests that the PLS-SEM model captures meaningful variance in motivational and engagement factors—variables ignored in simpler benchmarks. These results reinforce the Self-Determination Theory (SDT) premise that competence-supportive feedback and intrinsic motivation jointly enhance perceived mastery (Ryan and Deci, 2020).

The hypothesized structural model was tested using bootstrapping with 10,000 resamples. Figure 3 illustrates the structural paths, and Table 7 presents the model-fit indices, while Table 8 reports the standardized coefficients, confidence intervals, effect sizes, and explained variance.

OA displayed statistically significant associations with MODT (β = 0.514, p < 0.001) and ENDT (β = 0.382, p < 0.001), while its direct association with ESC (β = 0.174, p < 0.001) was comparatively weaker. When common-method bias was accounted for, this direct path became statistically non-significant, suggesting that OA’s contribution to self-perceived competency operates mainly through motivational and engagement mechanisms rather than directly.

The largest association (OA → MODT) is consistent with Self-Determination Theory (SDT), where assessment structures that provide competence cues and autonomy support are strongly linked to enhanced motivation (Ryan and Deci, 2020). MODT showed substantial positive associations with both ESC (β = 0.420, p < 0.001) and ENDT (β = 0.394, p < 0.001), indicating that motivated learners are more likely to invest effort and perceive higher proficiency gains. ENDT, in turn, was positively associated with ESC (β = 0.319, p < 0.001), reinforcing Engagement Theory’s emphasis on behavioral, emotional, and cognitive participation as a precondition for learning advancement (Fredricks et al., 2004; Skinner et al., 2009).

Multiple indirect associations were confirmed. The paths OA → MODT → ESC (β = 0.216, p < 0.001) and OA → MODT → ENDT (β = 0.202, p < 0.001) illustrate that motivation functions as a central psychological mechanism linking assessment design to both outcomes and engagement. The sequential association OA → MODT → ENDT → ESC (β = 0.065, p < 0.001) further depicts a layered process: assessment practices relate to increased motivation, which is associated with higher engagement, which in turn aligns with improved self-perceived competency.

In terms of explanatory power, the model accounted for 60.7% of the variance in ESC, 45.6% in ENDT, and 26.4% in MODT, exceeding benchmarks for acceptable predictive adequacy in educational-technology SEM studies (Hair et al., 2022; Shmueli et al., 2019). Effect sizes (f²) ranged from small (OA → ESC = 0.047) to large (OA → MODT = 0.358). Within a vocational-EFL pedagogical context, f² ≈ 0.05 is considered modest yet meaningful because even small increases in motivational engagement can translate into noticeable self-confidence gains, while f² > 0.30 indicates substantial practical relevance.

Overall, these findings are consistent with both LOA and SDT frameworks rather than extending them: AI-supported online assessment is associated with stronger motivational and engagement states, and these in turn are linked to higher levels of self-perceived speaking competence. The structural results therefore highlight that motivation serves as the core conduit between assessment design and learners’ perceived oral development, reaffirming that effective digital assessment depends less on automation itself and more on the motivational climate it creates.

To complement the PLS-SEM analysis, fsQCA was conducted to uncover multiple associative pathways leading to high self-perceived English-speaking competency (ESC). Unlike regression-based approaches that estimate net effects, fsQCA identifies combinations of conditions sufficient to explain outcomes, thereby accommodating equifinality and causal complexity (Ragin, 2008).

All indicators were transformed into fuzzy scores (0–1) using percentile-based thresholds: full membership (95th percentile), crossover (50th percentile), and full non-membership (5th percentile). These calibration anchors were selected to reflect the empirical distribution of the Likert-scale data and to ensure comparability across constructs (OA, MODT, ENDT, and ESC). Sensitivity checks using alternative anchors (90/50/10) produced consistent results, confirming the robustness of the calibration procedure.

Table 9 presents the necessity analysis results. All three conditions—OA, MODT, and ENDT—displayed high consistency values (≥0.80) but did not exceed the conventional threshold of 0.90. Hence, these variables are more appropriately described as frequently present or highly typical among cases of high ESC rather than strictly necessary conditions. MODT emerged as the most consistently associated factor (consistency = 0.88, coverage = 0.72), emphasizing the pivotal role of motivational processes in facilitating speaking competency development within digital assessment contexts.

Table 10 displays the main configurations of conditions sufficient for high self-perceived English-speaking competency (ESC). Three dominant combinations were identified.

The overall solution exhibited a consistency of 0.89 and coverage of 0.58, demonstrating strong explanatory adequacy. Motivation (MODT) appeared as a core condition across all configurations, reaffirming its pivotal role in linking assessment, engagement, and learning outcomes.

The fsQCA results complement the PLS-SEM findings by revealing that while online assessment (OA) contributes to ESC primarily through motivational and engagement pathways, high ESC can also emerge through configurations where OA is absent but learners’ motivation and engagement remain strong. This pattern aligns with Self-Determination Theory (Ryan and Deci, 2020), which emphasizes the autonomous and self-regulatory nature of motivation in sustaining learning performance even when external structures such as AI-driven assessment are limited.

To enhance interpretability, Table 11 summarizes the three dominant causal configurations leading to high ESC. These simplified “causal recipes” illustrate the principle of equifinality—where multiple alternative condition sets can produce the same outcome (Ragin, 2008).

Post hoc analysis of alternative practice behaviors. To further interpret the configuration where high ESC emerged without OA (C2: OA○ MODT● ENDT●), we conducted chi-square comparisons across learner groups. Results indicated that students in the MODT● ENDT● OA○ group reported significantly higher frequencies of (a) peer speaking practice inside/outside class and (b) use of self-study mobile speaking apps compared with both OA-inclusive and low-MODT/ENDT groups [χ²(4, N = 436) = 9.87, p = 0.042]. These findings suggest that highly motivated and engaged learners compensate for the absence of AI-based assessment by constructing alternative practice ecosystems that sustain performance gains.