This paper examines whether the adoption of Supervisory Technology (SupTech) is associated with a reduction in greenwashing in the European banking sector over the period 2014–2025. Using a progressive empirical framework combining the Synthetic Control Method (SCM), a split-sample SCM, an OLS regression of the SCM-estimated treatment effect, and a Causal Random Forest (CRL) via T-Learner applied to a panel of European banks, we provide evidence consistent with a meaningful reduction in greenwashing associated with SupTech adoption, which is robust across multiple identification and validation strategies. The split-sample SCM and OLS analyses reveal that this effect is amplified by higher capital adequacy, genuine ESG engagement, and stricter regulatory environments, while larger banks exhibit a systematically attenuated response. Contrary to the complementarity hypothesis, RegTech does not reinforce SupTech’s disciplining effect; instead, the evidence points to a substitution mechanism whereby banks with developed internal compliance infrastructure derive limited marginal benefit from external supervisory technology. The Causal Random Forest analysis provides evidence of a statistically significant and stable average treatment effect and indicates that bank digital maturity and FinTech adoption are the most consistent drivers of SupTech’s effectiveness. Policy simulations show that improving digital maturity, rather than RegTech endowment, yields the largest additional greenwashing-reduction gains. These findings suggest that SupTech acts as a credibilization mechanism whose effectiveness depends on the stringency of the external regulatory architecture and the digital absorptive capacity of supervised institutions. External validity to less harmonized regulatory environments remains an open empirical question.
Environmental, Social, and Governance (ESG) ratings increasingly shape capital allocation, corporate strategy, and regulatory oversight, yet their credibility is constrained by methodological opacity, rating divergence, and greenwashing risk. Prior reviews treat machine learning (ML) in ESG as a prediction problem. We identify an emerging research trajectory in which ML is increasingly used not only to consume ESG signals but also to verify their construction and credibility. Drawing on signaling theory, we conduct a PRISMA-guided systematic review of 127 peer-reviewed studies from Scopus and Web of Science to examine how machine learning (ML), deep learning (DL), Natural Language Processing (NLP), and Explainable AI (XAI) are transforming ESG rating analysis. We develop a tripartite framework classifying studies by the functional role of the ESG score: predicted (n = 29), used (n = 57), or assessed (n = 41). Our central contribution is the first synthesis of the methodological-assessment stream, organized into four clusters: XAI reverse-engineering of proprietary scoring functions, divergence reconciliation, greenwashing detection, and unsupervised industry-materiality clustering. The evidence assembled in this stream indicates that ESG ratings weight low-cost aspirational disclosure heavily relative to costly performance evidence, suggesting that greater reliance on aspirational disclosure relative to performance evidence may increase greenwashing risk, consistent with signaling-theory concerns. A study-level validation appraisal further shows that the most extreme fit statistics often arise in target-proximal reconstruction or non-temporal validation settings, cautioning against interpreting high R2 as evidence of transferable out-of-time forecasting.
Korean firms record the cost of entertaining customers and counterparties in a separate account and report it on its own where they judge it material. The COVID-19 pandemic interrupted the activity that account pays for. This study asks whether the stock market’s valuation of that spending changed around the interruption. Using 19,334 firm-years on 2331 Korean listed firms from 2016 to 2025, Tobin’s Q is regressed on reported entertainment expenditure scaled by sales and interacted with indicators for the pandemic years 2020 to 2021 and the years that followed, with firm and year fixed effects and standard errors clustered by firm. Entertainment intensity was positively associated with firm value before the pandemic, although that association is carried by the heaviest spenders and does not survive their removal. The pandemic interaction is imprecise, running from −20.81 to 11.66 against a benchmark of 21.916. After the pandemic, the association was eliminated: the interaction is −29.655, and it holds across nine measurement and deflator variants. The association changes sign around 2022; a test that does not impose the date locates the change there, but a discrete break and a decline that steepens cannot be separated. Advertising and research intensity attenuate at least as much, so the change belongs to discretionary expenditure as a class rather than to entertainment alone.
Although cryptocurrency markets trade continuously, the intraday distribution of Bitcoin’s volatility has migrated toward United States trading hours as the asset has institutionalized. Using ten years of hourly Kraken XBT/USD data (2016–2025; 87,672 observations), I document this migration and tie its timing to the U.S. trading calendar with two natural experiments. When U.S. clocks change, the intraday volatility peak shifts by one hour in UTC, tracking the displaced equity open; the shift appears only in the institutionalized period. On weekday NYSE holidays, when the U.S. cash market is closed while most other markets trade, the U.S.-hours share of realized variance falls by 13.9 percentage points relative to matched weekdays, close to the uniform benchmark of 0.375 (the share expected if variance were distributed evenly across the 24 h day), and this effect is also absent before 2019. Neither result is consistent with an explanation fixed in UTC. A window-free circular index of intraday concentration rises by more than 40% over the decade, with a structural break in November 2021, and no local break at the 2017 futures launch or the 2024 spot-ETF approval. A placebo simulation shows that naive whole-sample event contrasts on this trending series are significant for 100% of random pseudo-event dates. The microstructure of a nominally 24/7 market increasingly bears the imprint of the U.S. equity calendar.
Expected stock returns reflect cash distributions, fundamental growth, and market valuation, yet these sources are often examined separately. This study develops a Theoretical Rate of Return (TRR) framework that organizes them within a common multiplicative representation. The theoretical foundation is a one-period ex post decomposition of realized total shareholder return into dividend yield, internal growth, and valuation adjustment. Its empirical implementation instead uses Growth and Value characteristics observable at portfolio formation as predictive proxies; these variables are neither individually novel nor treated as exact equivalents of subsequently realized components. Using quarterly Chinese A-share data, the study evaluates the joint Growth–Value return relation through within-quarter sequential double sorting, log-linear ordinary least squares, interaction and centered quadratic specifications, and threshold regression. The full-sample results indicate that Growth generally provides the stronger first-order relation, whereas the standalone Value coefficient is less consistent across horizons. The interaction term is not robust to multiple-testing adjustment, while evidence of curvature is confined to selected specifications and horizons; threshold evidence is likewise limited and requires cautious interpretation. Quarter-by-quarter cross-sectional regressions and a purged four-period out-of-sample design further show that the quadratic model is not systematically more stable or predictively superior to the parsimonious linear benchmark. Supplementary evidence from Hong Kong and the United States indicates that the relative importance of Growth and Value varies across market environments. The contribution is therefore not a new pricing factor or econometric method. Rather, TRR provides a common economic organization for established characteristics and a disciplined framework for distinguishing full-sample functional-form evidence from temporal stability and out-of-sample predictability.
Accurate conditional volatility forecasting is essential for risk management, asset pricing, and portfolio optimization. Despite the widespread use of GARCH family models and the proliferation of deep learning architectures, a methodological gap persists: neural networks are rarely trained under criteria statistically coherent with the distributional properties of financial returns. This paper makes two contributions. First, we establish formal structural equivalences between classical heteroscedastic models and neural architectures, showing that ARCH(p) is equivalent to a single-layer linear MLP and GARCH(1,1) to a constrained LSTM, with an explicit parameter correspondence. Second, we propose LSTM-SSE-t-Student, a parsimonious LSTM trained with a hybrid loss that combines the sum of squared errors with the Student-t negative log-likelihood, penalizing errors in the tails of the return distribution. The model is evaluated on six daily series spanning three asset classes—Bitcoin, Ethereum, Gold, Oil, the DJIA, and the S&P 500—across diverse regimes, including the COVID-19 period, against a broad set of econometric and deep learning benchmarks. Relative to GARCH(1,1), it significantly improves point forecast accuracy and probabilistic calibration over naive, short-memory, and regime-switching specifications, while matching the strongest GARCH family and deep learning competitors; a sensitivity analysis shows that the likelihood term lowers the QLIKE loss for most series, with a market-dependent optimal weighting. Statistical significance is assessed via Diebold–Mariano tests with a correction for multiple comparisons, and Value-at-Risk and Expected Shortfall backtests confirm adequate tail calibration for the equity and cryptocurrency series. Interpretability is preserved through the GARCH-consistent structure, whose learned gate dynamics are stable across random seeds.
Two arguments in the literature bear on how accounting information should be priced in emerging markets. The first holds that where enforcement of financial reporting is weak, investors favour book value over earnings because book value depends less on managerial judgement. The second holds that value relevance falls when macroeconomic uncertainty rises. Vietnam offers an informative setting for both because reporting enforcement is uneven and measurable, individual investors supply more than eighty per cent of trading value, and the sample period contains unusually wide macroeconomic variation. Using 631 non-financial listed firms over 2009 to 2024 and 7982 firm-year observations, we find no supporting evidence for either prediction. Earnings carry more than twice the economic weight of book value, and the earnings result withstands selection on unobservables more than four times that on observables, against approximately one for book value. Value relevance is no weaker among small firms, where full disclosure compliance is roughly 50 per cent, than among large firms, where it is roughly 85 per cent. Testing state dependence item by item through interaction terms rather than through aggregate explanatory power, the valuation weights on earnings and book value are statistically indistinguishable across favourable and adverse states under two independent definitions, while the weights on operating cash flow and firm size shift significantly. The results hold under lagged, dynamic and logarithmic specifications.
This paper examines the association between fiscal policy-based loans (PBLs) and macroeconomic performance in five Latin American and Caribbean countries that received fiscal PBLs in response to the 2008 global downturn. We combine three alternative counterfactual estimators for GDP per capita—Synthetic Control Method (SCM), Matrix Completion (MC), and a Cointegration-Based Counterfactual (CBC)—with an Event Study with Single Treated Units (ES-STU) for a broader set of macroeconomic outcomes. Across Costa Rica, the Dominican Republic, El Salvador, and Guatemala, SCM, MC, and CBC show substantial convergence in the direction of the longer-run GDP-per-capita gap, although the timing and magnitude of the estimates differ. Jamaica is the main exception, displaying greater dispersion across counterfactual estimators, consistent with the greater uncertainty revealed by the sensitivity analyses and with the identification challenges posed by its unusually long intervention period, which increases the scope for intervening shocks and domestic reforms to confound the estimated effects. For the broader macroeconomic outcomes, the ES-STU evidence is strongest for inflation, with favorable post-intervention differences and sufficiently comparable pre-treatment trends in Costa Rica, the Dominican Republic, El Salvador, and Guatemala; selected fiscal-balance outcomes, particularly in Costa Rica and El Salvador, also provide relatively strong evidence. Our findings are therefore best interpreted as case-specific evidence on PBL-supported reform episodes as a whole—combining financing, conditionality, institutional support, and domestic reforms—rather than as treatment effects attributable to any single multilateral institution.
Drawing on agency theory, resource dependence theory, and upper echelons theory, this study examines how board characteristics and ownership structure relate to digital transformation in 29 Vietnamese commercial banks from 2012–2024. Using 377 bank-year observations, the study develops a multidimensional Digital Transformation Index (DTI) comprising four dimensions. The baseline models use Prais–Winsten panel-corrected standard errors (PCSEs) and are complemented by nonlinear, post-2020, ownership-moderation, robustness, endogeneity, and alternative-measurement analyses. The results indicate that board size is negatively associated with digital transformation. In contrast, board independence is consistently and positively associated with it, and directors’ educational attainment is positively associated with selected dimensions. The nonlinear analysis further suggests that the negative association of board size becomes more pronounced beyond a moderate board size, without implying a universally optimal threshold. State ownership is negatively associated with the aggregate DTI and several of its dimensions, whereas foreign ownership shows no consistently positive association. We find no statistically significant evidence that state ownership systematically strengthens or weakens the relationships between board gender diversity or board independence and digital transformation. The post-2020 analysis further reveals temporal heterogeneity, particularly in the relationships involving board size, chief executive officer (CEO) board membership, and board independence. The principal findings remain broadly robust across alternative specifications and endogeneity analyses. Based on these findings, the study offers practical implications for bank managers and policymakers regarding board composition and expertise, ownership-related governance, and regulatory support for effective and sustainable bank digital transformation.
This study asks whether chief executive compensation at private nonprofit four-year institutions reflects financial performance or financial scale. The sample is 569 institutions in the 1000 to 4999 enrollment band, matched to IPEDS finance data for 2018–19 through 2023–24 and IRS Form 990 compensation data. Financial scale dominates: when entered jointly, the revenue coefficient is 0.295, enrollment is insignificant, and a Wald test rejects the equality of coefficients, although compensation remains positively associated with performance (0.088, p = 0.001). As a test of agency theory, institutions whose revenue and net tuition both declined are compensated about 14 percent below prediction; in an exploratory severity analysis, the discount reaches about 16 percent when each fell more than 20 percent in constant dollars. These are cross-sectional associations, not causal effects of board policy. Panel estimates are consistent with the discount developing over the window (the growth differential is significant at the 10 percent level). Adjustment is incomplete: about half of institutions in real decline are paid above prediction; the breakaway core label for this group is descriptive, not a finding of excess, and its larger enrollment is not distinguishable from the sector-wide pattern. Because the sector’s recovery is nominal rather than real, above-benchmark compensation occurs amid real contraction.
This study examines whether investor attention and the market information environment transmit environmental, social, and governance (ESG) information into stock returns. The analysis employs a balanced panel of 31 mining companies listed on the Indonesia Stock Exchange from 2019 to 2023, comprising 155 firm-year observations. ESG performance is measured using an external ESG score, investor attention is proxied by the Google Search Volume Index, and the information environment is captured inversely by the relative bid–ask spread. Firm fixed-effects models with heteroskedasticity-robust standard errors clustered at the firm level are estimated for the investor-attention, spread, and stock-return equations. Indirect effects are assessed using 5000 firm-level cluster-bootstrap replications. The results show that ESG performance is not significantly associated with contemporaneous stock returns. ESG is positively but only marginally associated with investor attention and significantly associated with a narrower relative bid–ask spread, indicating a more favourable information environment. Investor attention and the relative spread are significantly associated with stock returns. However, neither the attention-mediated effect nor the spread-mediated effect is statistically significant. These findings distinguish ESG signal recognition from signal pricing, showing that ESG information can influence investor attention and the market information environment without forming a statistically significant transmission mechanism to contemporaneous stock returns.
A key question for investors is how to diversify their clean energy stock holdings amid adverse conditions. Gold has a long history of being an effective diversifier for financial assets. Energy tokens exhibit a low correlation with clean energy stocks and may also serve as a good diversifier. This paper examines whether energy tokens or gold is the most effective diversifier for a clean energy stock portfolio. The analysis uses R2 and TVP-VAR measures of return connectedness on a dataset that includes ETFs for wind, solar, nuclear, grid connectivity, electric vehicles, gold, and two energy tokens (POWR and SNC). Network connectedness was highest at the start of the COVID-19 pandemic and during the escalation of the Russia–Ukraine war. R2 and TVP-VAR total network connectedness correlate highly (0.93). Grid connectivity is a dominant net transmitter of shocks. Gold is a dominant net receiver of shocks. POWR and SNC have low net connectedness with the other assets. Portfolio analysis reveals that, on a risk-adjusted basis, gold is a more effective diversifier than energy tokens. These results are robust across several portfolio choices (minimum variance, minimum correlation, and minimum connectedness) and representative transaction costs. For three of the four portfolios studied, a portfolio of clean energy stocks and energy tokens has lower risk-adjusted returns than one that invests only in clean energy. Minimum variance portfolios have the highest Sharpe ratios.
Low correlation alone does not establish an implementable diversification benefit. This study tests whether Bitcoin and Ethereum improved a fixed, retrospectively selected sample of 13 Thai REITs—not a point-in-time investable universe—for a Thai baht-based investor, using dependence-aware bootstrap inference (Bonferroni, Benjamini–Hochberg, and Romano–Wolf correction) across 1562 matched daily observations (16 January 2020–30 June 2026); modelled costs cover quarterly top-level reallocations only, not daily REIT-basket weight maintenance. The most robust finding is a cost, not a benefit: daily 95% conditional value-at-risk deteriorates significantly and consistently across all three correction methods for five of eight crypto-inclusive portfolios. By contrast, the seemingly compelling point-estimate pattern—all eight portfolios show higher Sharpe ratios and smaller maximum drawdowns—does not survive the same scrutiny: none of the eight ΔSharpe improvements is significant under Bonferroni or Benjamini–Hochberg, only the rolling strategy (P9) is significant under Romano–Wolf, and all unadjusted evidence disappears under sample-window sensitivity checks. Annual inclusion margins were driven largely by cryptocurrency performance, while counterfactual re-centering descriptively illustrated declining ΔSharpe as benchmark strength increased; none of the underlying comparisons survived multiplicity adjustment. The findings illustrate how seemingly robust point-estimate gains can fail to survive multiplicity-corrected inference, while a downside-risk cost remains statistically significant across all three multiplicity procedures in the principal specification—a cautionary result for practitioners and future diversification studies alike.
Tax fraud constitutes a significant challenge to fiscal sustainability, as it reduces public revenues, distorts market competition, and undermines confidence in tax systems. The digitalization of tax administration and the growing availability of large-scale financial data have created opportunities for applying Artificial Intelligence (AI) and Machine Learning (ML) to tax fraud detection and risk assessment. Building on recent research on AI/ML-based tax fraud detection, the present study systematically reviews 82 empirical studies published between 1996 and 2025 covering both individual and corporate taxpayers, with particular attention to the predictive relevance of financial and non-financial indicators. The findings show that AI and ML techniques enhance the detection of complex fraud patterns by analyzing large-scale datasets. The study highlights the importance of financial and non-financial indicators as input variables in ML models. Importantly, frequently used indicators do not necessarily demonstrate greater predictive relevance for tax fraud detection, and their relevance varies across different AI/ML modelling approaches. Nevertheless, significant challenges remain regarding data quality, privacy protection, prediction reliability, and model transparency. The study provides an evidence-based foundation for future research and the development of more effective, reliable, and adaptive tax audit systems.
Artificial intelligence (AI) is reshaping financial markets—accelerating execution, automating allocation, and concentrating analytical capacity within a shrinking set of foundational models. Prior work has documented AI’s contribution to instability through algorithmic herding and high-frequency volatility, but the literature lacks a coherent classification of the mechanisms through which AI triggers catastrophic, self-reinforcing market dislocations, or “flash crashes.” This prospective review proposes a three-category taxonomy: (1) endogenous algorithmic herding crashes, driven by correlated model behavior; (2) exogenous model error cascade crashes, in which AI system failures propagate across interconnected venues; and (3) adversarial generative AI (GenAI) disinformation crashes, in which fabricated narratives trigger automated trading responses. The taxonomy is further motivated by the structural parallel between contemporary AI model homogeneity and the homogenization of Value-at-Risk (VaR) models before the 2008 crisis—a link recently formalized in the modeling literature. The analysis is extended to the emerging frontier of agentic AI—autonomous systems capable of multi-step planning and inter-agent interaction—which introduces qualitatively new systemic risks that existing regulatory frameworks are unprepared to address. The October 2025 cryptocurrency liquidation cascade, which liquidated over $19 billion within 24 h, serves as the primary empirical case study. The article concludes with policy recommendations on model-diversity mandates, real-time AI trading surveillance, adaptive circuit-breaker design, and cross-regulatory coordination on GenAI financial disinformation.
This study develops a two-period regulatory model in which precautionary capital, information-producing reporting, provider participation, and supervisory architecture are chosen jointly. Reporting may produce a verified signal before continuation capital is set, but it also entails direct, participation, and fixed setup costs. Under a known prior, an ignorable diagnostic has weakly nonnegative gross decision value, yet reporting is activated only if its optimized net surplus exceeds the no-reporting option. Under recursive maxmin, an admissible adverse-state-certainty model can eliminate learning and shift policy toward precaution; uniformly interior priors can preserve learning. Recursive smooth ambiguity converges only to the maxmin problem defined on the same finite model support. For arbitrary noisy histories, information dominance under ambiguity additionally requires projective consistency of experiments, conditional prior sets, and second-order weights; rectangularity alone is not sufficient. The resulting regime theorem separates information value, reporting intensity, activation, and architecture choice. Analytical proofs establish the claims, while deterministic code, independent grid calculations, and fresh-process reruns verify the registered examples.
Rising geopolitical tensions have made international investment flows increasingly sensitive to differences in risk conditions across countries. This study examines the relationship between geopolitical risk and foreign direct investment (FDI) inflows using panel data for 36 major investing countries over the period 2002–2022. Complementing conventional country-specific measures of geopolitical risk, the study adopts a relational perspective by constructing geopolitical risk relative to Vietnam as the common host economy. Fixed-effects models are combined with moderation, geopolitical shock, robustness, and heterogeneity analyses. The results show that higher geopolitical risk in source countries relative to Vietnam is significantly associated with lower bilateral FDI inflows. Political stability significantly moderates this relationship, although the marginal effect of geopolitical risk becomes statistically insignificant at relatively high levels of political stability. The shock analysis further shows that large geopolitical disturbances are associated with lower FDI inflows, while lagged geopolitical risk remains negatively associated with subsequent investment. The main findings remain robust to an alternative ratio-based measure of relative geopolitical risk, the inclusion of exchange-rate conditions, and a dynamic System GMM specification. The negative association between geopolitical risk and FDI is also significantly stronger for developing than for developed source economies. Overall, the findings demonstrate that geopolitical risk, considered in relation to host-country conditions, is relevant to understanding patterns of international capital allocation under heightened geopolitical uncertainty.
This study examines whether internal audit competency, structural independence, and technology adoption form an interrelated governance capability structure. A cross-sectional survey of 397 governance, risk, and financial-control professionals in Saudi Arabia was analysed using PLS-SEM. Competency is strongly associated with independence (β = 0.654), while competency (β = 0.366) and independence (β = 0.388) are associated with technology adoption. The model explains 43% and 47% of the variance in independence and technology adoption, respectively. Independence carries a significant indirect association between competency and technology adoption (β = 0.254). The findings support systematic interdependence among the three capabilities. A post hoc model-implied sensitivity analysis shows that alternative directional orderings are covariance-equivalent; accordingly, the evidence supports relational structure but not a unique causal ordering.
This paper intends to predict the prices of commodities in metals, energy, agriculture, and other industrial products sectors. This study covers a large span of about ten and a half years with 3484 daily observations starting from 7 April 2016 to 7 April 2026. The commodities taken into account for this study are: gold, silver, copper, Brent oil, natural gas, coal, corn, wheat, soybean, aluminum, nickel, and lead. On the methodological front, this paper is a little different as it conducts a comparative analysis of out-of-sample forecasting capability of four models, namely, a naïve one-day-lag (random walk) benchmark, a linear autoregressive model, a bagged decision-tree ensemble model, and a gradient-boosted decision-tree ensemble model. Forecast accuracy is measured in terms of error-based and directional forecasting metrics, while Diebold-Mariano tests, sector-level comparisons, volatility-regime analysis, and robustness checks serve as evaluation tools for qualitative assessments. The results highlight that forecasting performance differs not only with the time horizon but also with the specific sectors. A linear autoregressive model ranks as the most reliable and regularly successful among the four models under comparison. A bagged decision-tree ensemble model and a gradient-boosted decision-tree ensemble model, on the other hand, as machine learning tree-based models, may only offer fleeting and unstable gains. The further the forecast extends in time (from 1 day to 5 and 10 days), the less accurate it gets. Energy commodities are the hardest to predict, while industrial commodities seem to be relatively more predictable.
Accrual accounting reform can be formally adopted before the internal audit routines needed to assure it are operationally embedded. This exploratory diagnostic study examines perceived implementation risk among 95 internal audit professionals working in Saudi governmental entities. A content audit of an existing questionnaire showed that its original accrual-basis accounting application (ABAA), senior management support (SMS), and internal audit effectiveness (IAE) blocks combined conceptually adjacent benefit, support, resource, and outcome statements. The analysis therefore uses theory-guided, content-separated diagnostic domains and an item-level timeliness model rather than treating the original blocks as discriminant latent constructs or testing mediation. Within-respondent contrasts were assessed using paired t-tests and Wilcoxon signed-rank checks. Because timely completion is a five-category ordinal outcome, ordinal regression is the scale-respecting approach. The full five-predictor ordered-logit model was retained as a theory-complete diagnostic because E4 explicitly includes staffing, but its global proportional-odds restriction was rejected due to staffing. Accordingly, substantive ordinal inference and model-implied probabilities are based on the assumption-compatible reduced model excluding staffing, while the full model is reported transparently as a diagnostic sensitivity specification. Formal reform orientation exceeded applied reform capacity by 0.437 points (p < 0.001; dz = 0.441); visible governance and monitoring exceeded operational audit infrastructure by 0.674 points (p < 0.001; dz = 0.830); and perceived staffing sufficiency exceeded timely task completion by 0.800 points (p < 0.001; dz = 0.615). In the assumption-compatible reduced ordered-logit model, annual planning (b = 0.581, p = 0.016) and the accrual-aligned internal audit guide (b = 1.570, p < 0.001) were positively associated with timeliness; the corresponding full-model estimates were nearly identical. The OLS-HC3 robustness model accounted for 61.2% of respondent-level variation, although same-source response tendencies may contribute to this fit. The observed pattern is compatible with uneven layered institutionalization and a risk of decoupling between visible or formal elements and operational routines; it does not establish temporal sequencing, sector-wide prevalence, intentional symbolic compliance, or causal effects. Conclusions are bound by the cross-sectional, same-source design and the post hoc diagnostic use of an existing instrument that requires prospective validation.
