Carbon Information Disclosure Quality in China’s Petroleum and Petrochemical Enterprises: An LLM Approach

Abstract

Global climate governance and corporate low-carbon transition have made carbon information disclosure important for assessing firms’ environmental governance and climate-risk responses. This study develops an industry-specific carbon information disclosure quality (CIDQ) framework for Chinese A-share listed petroleum and petrochemical firms, using 45 firm-year observations from 15 firms during 2022–2024. The framework includes 7 primary, 15 secondary, and 33 tertiary indicators. Disclosure texts were scored by the DeepSeek-V3.2 large language model (LLM) under predefined rule-based criteria, with temperature set to 0. Reliability was assessed against manual scoring of 15 reports, yielding an intraclass correlation coefficient (ICC) of 0.974. The full-sample mean score is 34.02, accounting for only 51.55% of the theoretical maximum of 66, indicating that the overall disclosure level remains relatively low. The annual mean score increased from 29.07 in 2022 to 37.60 in 2024, representing a cumulative rise of 8.53 points, or 29.34%. Substantial inter-firm differences are also observed: Sinopec recorded the highest three-year average score of 52.67, whereas Yunnan Yunwei recorded the lowest at 10.67. This study may provide a methodological reference for structured CIDQ evaluation and disclosure improvement in high-emission industries.

1. Introduction

Global climate governance has accelerated in recent years, increasing the importance of carbon information disclosure. The Emissions Gap Report 2024 issued by the United Nations Environment Programme (UNEP), which assesses the gap between current policy trajectories and the emission levels required to achieve the goals of the Paris Agreement while tracking global emissions trends, nationally determined contributions (NDCs), and sectoral mitigation potential, reports that global greenhouse gas emissions reached a record high of 57.1 GtCO2e in 2023. For firms in high-emission industries, information on carbon emissions, emission reduction targets, governance mechanisms, and climate-related risks is no longer merely a supplementary statement of environmental responsibility; rather, it has become an important basis for external stakeholders to assess firms’ environmental governance capacity, responses to climate-related risks, and progress toward a low-carbon transition. Accordingly, carbon information disclosure has gradually evolved from a predominantly voluntary corporate practice into an important information activity with implications for corporate governance, regulatory oversight, and market pricing [1,2,3,4].

As research in this field has expanded, scholarly attention has gradually shifted from whether firms disclose carbon-related information to the quality of such disclosure. Compared with merely examining whether firms disclose carbon-related information, carbon information disclosure quality (CIDQ) places greater emphasis on whether the disclosed content is complete, reliable, comparable, and useful for decision-making, thereby better reflecting firms’ actual level of climate governance and low-carbon transition capacity. However, the existing literature has yet to establish a unified framework for defining CIDQ, constructing indicator systems, and setting scoring rules, and substantial differences remain across studies in terms of evaluation dimensions and measurement outcomes [1,5,6]. This suggests that the assessment of CIDQ should not rely on the simple application of generic frameworks, but instead requires targeted adjustments that take into account the emission characteristics, transition pathways, and disclosure priorities of specific industries.

The petroleum and petrochemical industry provides an important context for research on CIDQ. This industry is characterized by high emission intensity, complex production processes, long industrial chains, and a large share of value-chain-related emissions, and therefore faces considerable pressure from climate governance and strong constraints associated with the low-carbon transition. Compared with firms in other industries, carbon information disclosure in this sector involves not only direct emissions and energy use, but also more industry-specific issues, such as Scope 3 emissions (according to the Greenhouse Gas Protocol, Scope 3 emissions refer to other indirect emissions occurring across the value chain, including both upstream and downstream emissions; see World Resources Institute (WRI) and World Business Council for Sustainable Development (WBCSD), Corporate Value Chain (Scope 3) Accounting and Reporting Standard, 2011) [7], supply chain carbon management, low-carbon technology development, responses to transition risks, and social impacts [8,9,10]. This indicates that the assessment of CIDQ in the petroleum and petrochemical industry cannot rely directly on generic evaluation frameworks, but instead requires a more targeted framework that incorporates industry-specific characteristics. Against this background, this study raises the following question: how can a CIDQ evaluation framework that captures industry characteristics be developed for China’s A-share petroleum and petrochemical industry, and how can such a framework be used to identify firms’ disclosure levels and inter-firm differences?

Recent sustainability disclosure initiatives in China have further increased the importance of standardized non-financial reporting. However, these developments do not eliminate the need for industry-specific evaluation tools. In sectors such as petroleum and petrochemicals, disclosure content is technically complex, strongly value-chain related, and closely linked to transition risk, making generic reporting guidance insufficient for fine-grained quality assessment. Accordingly, the specific gap addressed by this study is not simply the lack of a unified CIDQ framework in general, but the absence of an industry-specific, rule-based, and operationally transparent framework for evaluating carbon information disclosure quality in a high-emission sector with strong Scope 3 relevance and transition-related disclosure complexity.

Existing research on CIDQ still relies heavily on manual content analysis. Although manual coding enables researchers to identify disclosure content in light of textual context, its limitations in terms of time costs, consistency control, and repeated operations have become increasingly evident as sample sizes expand, disclosure texts become longer, and evaluation indicators grow more refined. Keyword-based text analysis is relatively efficient, but it has difficulty distinguishing substantive disclosure from symbolic statements and is less capable of capturing contextual semantics and differences in information quality. Traditional machine-learning methods offer greater flexibility in text classification and pattern recognition, but they typically depend on large-scale manually labeled samples and feature engineering. In recent years, the rapid development of large language models (LLMs) has provided a new technical pathway for structured disclosure evaluation. For tasks such as CIDQ evaluation, in which scoring criteria are relatively explicit and task boundaries are clearly defined, LLMs may be particularly suitable because they can identify and score disclosure content on the basis of predefined indicators and scoring rules, while also supporting contextual understanding and information integration across different sections of lengthy and structurally complex reports [11,12].

In this context, this study takes A-share listed companies in China’s petroleum and petrochemical industry from 2022 to 2024 as the research sample. In terms of research design, this study first constructs a CIDQ indicator system comprising 7 primary indicators, 15 secondary indicators, and 33 tertiary indicators by integrating the core dimensions of disclosure quality identified in the existing literature, the commonly adopted requirements of international climate and sustainability disclosure frameworks, and the disclosure characteristics of the petroleum and petrochemical industry. Second, based on clearly defined scoring rules for the tertiary indicators, this study employs an LLM to identify, interpret, and score carbon information disclosed in corporate social responsibility reports, ESG reports, and sustainability reports, thereby constructing a company-year-level AI-based carbon information disclosure quality index (AI-CIDQ). Finally, based on the scoring results, this study systematically analyzes the current state of CIDQ among Chinese petroleum and petrochemical enterprises in terms of overall distribution, annual changes, inter-firm differences, and performance across industry-specific disclosure dimensions, and further examines the reliability and consistency of the LLM-based scores through comparison with manual scoring. Accordingly, this study focuses on framework construction and measurement rather than on testing the determinants or causal effects of CIDQ, with the purpose of developing an industry-specific and operationally transparent framework for evaluating carbon information disclosure quality in China’s petroleum and petrochemical industry.

This study makes three main contributions. First, it situates the evaluation of carbon information disclosure quality (CIDQ) in the petroleum and petrochemical industry, a high-emission sector characterized by complex production processes, long value chains, and strong transition constraints, thereby extending CIDQ research to a context in which generic disclosure evaluation frameworks may be insufficient. Second, it develops an industry-oriented CIDQ indicator system comprising 7 primary indicators, 15 secondary indicators, and 33 tertiary indicators by integrating general disclosure quality dimensions with the commonly adopted requirements of international climate and sustainability disclosure frameworks and the industry-specific disclosure characteristics of petroleum and petrochemical firms. Third, it introduces an LLM-assisted, rule-based scoring approach into CIDQ evaluation and demonstrates, through comparison with manual scoring, that this method can provide relatively stable support for structured disclosure assessment under clearly specified evaluation rules and retained manual review.

The rest of this paper is organized as follows. Section 2 reviews the relevant literature. Section 3 describes the sample selection, the construction of the indicator system, the scoring rules, the LLM-based scoring method, and the reliability test. Section 4 presents the overall results, annual changes, and inter-firm differences in CIDQ. Section 5 discusses the findings and the applicability of the LLM-based approach. Section 6 concludes the paper.

2. Literature Review

2.1. Corporate Carbon Information Disclosure

As climate-related risks continue to intensify, corporate carbon information disclosure has become an increasingly important topic in research on environmental governance, capital markets, and corporate sustainability. Early studies generally suggest that the disclosure of carbon emissions, energy use, emission reduction targets, and related environmental management information helps enhance the transparency of environmental governance and enables investors, regulators, and other stakeholders to identify firms’ climate-related risks and transition pressures [13,14]. As climate issues have become increasingly embedded in international governance frameworks and capital market rules, carbon information disclosure has gradually evolved from a voluntary tool for corporate environmental communication into an important information mechanism with governance, regulatory, and market implications.

The existing literature has mainly focused on the determinants, influencing factors, and economic consequences of carbon information disclosure. With regard to disclosure incentives, firms often voluntarily disclose carbon emissions and related governance information in response to environmental concerns from investors, consumers, and the broader public, thereby improving corporate image, maintaining legitimacy, and enhancing social reputation [15]. In terms of influencing factors, firms’ environmental performance, governance structures, and external institutional pressures are widely regarded as important determinants of disclosure levels. For example, some studies find that firms with better environmental performance tend to disclose more detailed climate-related information [16], while board independence, board diversity, and other governance characteristics can also significantly affect the extent of greenhouse gas emissions disclosure [17]. Cross-country comparative studies further show that differences in regulatory systems, enforcement intensity, investor attention, and social norms lead to substantial divergence in firms’ carbon information disclosure practices [18]. This suggests that corporate carbon information disclosure is not merely an act of information release, but rather the result of the joint influence of internal governance conditions and the external institutional environment.

In addition to disclosure incentives, prior studies have also examined the governance effects and economic consequences of carbon information disclosure. A higher level of carbon information disclosure helps reduce external information asymmetry and strengthens the basis on which market participants assess firms’ environmental responsibility and transition capacity [19,20]. At the same time, disclosure itself may further encourage firms to improve carbon emissions management and environmental governance practices through reputational mechanisms, regulatory pressure, and capital market feedback [10]. In this sense, carbon information disclosure not only reflects firms’ existing environmental performance but may also influence their subsequent environmental governance practices through reputational mechanisms, regulatory pressure, and capital market feedback.

In the Chinese context, with the introduction of the “dual carbon” goals and the continued advancement of green finance, ESG governance, and environmental regulation, increasing attention has been paid to corporate carbon information disclosure. Existing studies show that public information channels such as social media can exert legitimacy pressure on firms, thereby encouraging them to disclose more carbon emissions information [21]. Climate-related disclosure has also, to some extent, promoted corporate emission reduction efforts and improvements in environmental governance [10]. Overall, the existing literature has provided substantial insights into the institutional background, driving factors, and governance consequences of corporate carbon information disclosure. However, significant differences remain across firms in terms of disclosure scope, depth, and standardization. Accordingly, how to systematically evaluate the quality of corporate carbon information disclosure has become an important extension of this line of research.

More recently, sustainability disclosure regulation and green finance developments in China have further increased the importance of carbon information disclosure. In 2024, China’s stock exchanges introduced new sustainability disclosure rules and subsequently issued implementation-oriented guidance, further strengthening expectations for more standardized non-financial reporting. (China’s stock exchanges issued sustainability reporting guidelines for listed companies in 2024. See Shanghai Stock Exchange, Guidelines No. 14 of Shanghai Stock Exchange for Self-Regulation of Listed Companies—Sustainability Report (Trial), issued on 12 April 2024 and effective from 1 May 2024; Shenzhen Stock Exchange, Guidelines No. 17 of Shenzhen Stock Exchange for Self-Regulation of Listed Companies—Sustainability Report (Trial), 2024. For subsequent implementation-oriented guidance, see Shanghai Stock Exchange, Sustainability Reporting Guidance, 2025.) At the same time, the continued advancement of green finance and the growing policy relevance of carbon-related markets have increased the importance of climate-related disclosure for listed firms, especially in high-emission sectors.

These developments suggest that the Chinese market context after 2024 is not only characterized by stronger disclosure expectations, but also by a growing need for more fine-grained evaluation tools. In the petroleum and petrochemical industry, where carbon-related disclosure is closely linked to Scope 3 emissions, value-chain management, and transition-related governance issues, generic reporting guidance alone is insufficient for assessing disclosure quality in a detailed and industry-sensitive manner.

2.2. Evaluation of CIDQ

As corporate sustainability reporting and climate-related disclosure continue to expand, the focus of research on carbon information disclosure has gradually shifted from whether firms disclose such information to the quality of that disclosure. Compared with merely examining whether firms provide carbon-related information, CIDQ places greater emphasis on the completeness, reliability, comparability, and decision relevance of disclosed content, and therefore better reflects firms’ climate governance capacity, information transparency, and progress in the low-carbon transition. In response to this issue, existing studies have explored CIDQ from multiple perspectives, including evaluation dimensions, indicator systems, and measurement methods.

In terms of evaluation content, the existing literature generally defines CIDQ as the level of information disclosed by firms with respect to carbon emissions data, emission reduction targets, governance mechanisms, risk identification, strategic arrangements, and performance outcomes, while placing particular emphasis on completeness, reliability, comparability, and relevance [2]. Patten (2002) was among the first to apply content analysis to code corporate environmental information and construct an environmental disclosure index, thereby providing a foundational approach for subsequent research on the evaluation of environmental and carbon information disclosure [22]. Since then, related studies have gradually extended the object of evaluation from general environmental information to carbon information disclosure and progressively developed multidimensional indicator frameworks to identify differences across firms in both the quantity and quality of disclosure [23].

In terms of specific methods, the existing literature mainly adopts manual content analysis, keyword-based text analysis, and machine-learning-based text analysis, while some recent studies have also explored hybrid approaches that combine rule-based screening, semantic extraction, and model-assisted classification. Manual content analysis remains important for interpretive judgment, but it is time-consuming and requires careful reliability assessment to ensure coding consistency [24]. Keyword-based and text-mining approaches can improve processing efficiency, but they may have difficulty capturing contextual meaning, disclosure depth, and dispersed climate-related information in lengthy reports [25]. Machine-learning-based approaches can support classification tasks, but their performance often depends on the selection of textual analysis tools, labeled samples, and task-specific model design [26,27]. Although these approaches differ in technical form, they share a common challenge in CIDQ research: how to conduct structured and context-sensitive quality evaluation under transparent scoring rules.

In addition to measurement methods, prior studies have also examined the conditions shaping CIDQ. For example, corporate governance characteristics are widely regarded as important factors influencing firms’ disclosure behavior [28]. Carbon assurance can enhance the credibility of disclosed information and, to some extent, improve disclosure quality [29]. In the context of developing countries, disclosure quality is also significantly affected by factors such as firm size, profitability, and industry characteristics [30,31].

Overall, although recent studies have substantially advanced ESG disclosure research and the application of AI-based textual analysis, the existing literature still provides limited guidance on how to construct an industry-specific, rule-based, and operationally transparent framework for evaluating carbon information disclosure quality in high-emission sectors. This gap is particularly evident in the petroleum and petrochemical industry, where disclosure content is technically complex, strongly value-chain related, and closely linked to transition risk and industry-specific governance issues. Therefore, how to improve the efficiency, consistency, and replicability of CIDQ evaluation while preserving the transparency of evaluation rules remains an important methodological issue.

2.3. Methods for Disclosure Evaluation

Recent studies have further expanded the literature on ESG and sustainability disclosure as well as AI-assisted textual analysis. Emerging studies in accounting, finance, and sustainability research have begun to examine the use of AI and large language models (LLMs) for ESG-related tasks such as textual classification, topic identification, readability assessment, alignment analysis, and structured disclosure evaluation [32,33]. These studies suggest that AI-based methods may improve the efficiency and depth of disclosure analysis, while also raising important questions regarding interpretability, transparency, and research design.

As the volume of corporate sustainability reports and environmental disclosure continues to expand, researchers have increasingly attempted to improve the efficiency of disclosure research through automated text analysis techniques. Early studies mainly quantified textual information in corporate reports through word-frequency statistics, keyword matching, and dictionary-based methods, and on this basis identified firms’ disclosure topics, risk expressions, and management characteristics [34,35]. However, these methods are generally more suitable for detecting specific terms or textual patterns, while their ability to support information judgment in complex contexts, cross-paragraph integration, and rule-based scoring tasks remains relatively limited [36,37].

Subsequently, machine learning methods were gradually introduced into research on corporate disclosure. Machine learning can extract structured information from large volumes of corporate report texts, thereby significantly improving research efficiency [38]. Textual features have also been used to identify firms’ climate governance capacity, risk exposure, and information transmission characteristics [39,40,41]. In addition, topic modeling, sentiment analysis, and related machine-learning-based approaches have been widely applied in financial text and corporate disclosure research to examine disclosure topics, semantic tendencies, firm characteristics, and their economic consequences [42,43,44,45,46,47,48]. Overall, these methods have advanced disclosure research from manual coding toward automated text processing. However, their core tasks have still been concentrated primarily on topic identification, text classification, and semantic measurement, rather than on conducting multi-level structured scoring of disclosure content according to explicit rules.

In this context, LLMs provide a new approach to evaluating corporate disclosure. Unlike keyword-based methods, which mainly rely on term matching, and traditional machine learning methods, which typically depend on training samples and feature engineering, LLMs may be particularly suitable for rule-based disclosure evaluation tasks that require contextual understanding, cross-section information integration, and structured scoring under predefined criteria [11,12]. However, this does not imply that LLMs are free from methodological risks. In particular, LLM outputs may contain plausible but unsupported interpretations or hallucinated inferences when processing lengthy and technically complex reports. For this reason, their use in disclosure evaluation should be combined with explicit scoring rules, textual-evidence requirements, and manual verification procedures.

In the context of CIDQ evaluation, the research objects are usually sustainability reports or ESG reports that are lengthy and structurally complex, while the evaluation task requires the model to determine, on the basis of multi-level indicators, whether certain information is disclosed, to what extent it is disclosed, and whether it satisfies the corresponding scoring criteria. Such tasks require not only keyword identification but also contextual understanding, substantive judgment of disclosure content, and the mapping of information dispersed across different sections into a unified evaluation framework. Compared with earlier methods, LLMs are more likely to improve the automation and consistency of the evaluation process while preserving the transparency of evaluation rules.

Overall, the existing literature provides important insights into corporate carbon disclosure, CIDQ evaluation, and AI-assisted textual analysis, but these streams remain only partially integrated. Studies on carbon disclosure have mainly examined incentives, influencing factors, and economic consequences; studies on CIDQ have proposed multiple dimensions and measurement approaches, yet without establishing a unified or industry-sensitive evaluation structure; and studies on automated text analysis have improved processing efficiency, but have focused mainly on classification and semantic extraction rather than structured, rule-based quality assessment.

As a result, the literature still provides limited guidance on how to construct an industry-specific, rule-based, and operationally transparent framework for evaluating carbon information disclosure quality in a high-emission sector such as petroleum and petrochemicals. Accordingly, the research problem addressed by this study is how to develop and validate such a framework in a way that is methodologically transparent and practically applicable.

3. Research Design and Methodology

3.1. Sample and Data

This study takes A-share listed companies in China’s petroleum and petrochemical industry as the research sample. According to the Guidelines for Industry Classification of Listed Companies (2012 Revision) (China Securities Regulatory Commission (CSRC), Guidelines for the Industry Classification of Listed Companies (2012 Revision), issued on 26 October 2012 and effective from the date of issuance) issued by the China Securities Regulatory Commission, no separate industry category is established for the petroleum and petrochemical industry. Instead, the sector is mainly classified into two categories: B07 Extraction of Petroleum and Natural Gas and C25 Processing of Petroleum, Coking, and Nuclear Fuel. The sample scope of this study is defined accordingly. This classification standard is retained because it provides a clear and consistent basis for identifying the relevant A-share listed firms and is consistent with the industry coding used in listed-company datasets.

To ensure sample comparability and data completeness, two additional screening criteria were applied. First, firms that were subject to *ST or ST special treatment throughout the entire sample period were excluded. Second, firms must have continuously disclosed corporate social responsibility reports, ESG reports, or sustainability reports for three consecutive years from 2022 to 2024. After screening, a final sample of 15 firms was obtained. The study covers three fiscal years from 2022 to 2024, yielding a total of 45 firm-year observations. The sample firms and their basic information are presented in

Table 1. List of sample firms.

Company Name (English)	Company Name (Chinese)	Stock Code	Industry Category	City	Country
North Huajin Chemical Industries Co., Ltd.	北方华锦化学工业股份有限公司	000059	Petroleum Processing, Coking and Nuclear Fuel Processing Industry	Panjin	China
Shenyang Chemical Co., Ltd.	沈阳化工股份有限公司	000698	Petroleum Processing, Coking and Nuclear Fuel Processing Industry	Shenyang	China
Shanxi Meijin Energy Co., Ltd.	山西美锦能源股份有限公司	000723	Petroleum Processing, Coking and Nuclear Fuel Processing Industry	Taiyuan	China
Yueyang Xingchang Petro-Chemical Co., Ltd.	岳阳兴长石化股份有限公司	000819	Petroleum Processing, Coking and Nuclear Fuel Processing Industry	Yueyang	China
Shanxi Blue Flame Holding Company Limited	山西蓝焰控股股份有限公司	000968	Oil and Gas Extraction Industry	Taiyuan	China
China Petroleum & Chemical Corporation	中国石油化工股份有限公司	600028	Oil and Gas Extraction Industry	Beijing	China
Guanghui Energy Co., Ltd.	广汇能源股份有限公司	600256	Oil and Gas Extraction Industry	Urumqi	China
Sinopec Shanghai Petrochemical Company Limited	中国石化上海石油化工股份有限公司	600688	Petroleum Processing, Coking and Nuclear Fuel Processing Industry	Shanghai	China
Yunnan Yunwei Co., Ltd.	云南云维股份有限公司	600725	Petroleum Processing, Coking and Nuclear Fuel Processing Industry	Kunming	China
Shanxi Coking Co., Ltd.	山西焦化股份有限公司	600740	Petroleum Processing, Coking and Nuclear Fuel Processing Industry	Hongtong (Linfen)	China
Geo-Jade Petroleum Corporation	洲际油气股份有限公司	600759	Oil and Gas Extraction Industry	Haikou	China
Yunnan Coal & Energy Co., Ltd.	云南煤业能源股份有限公司	600792	Petroleum Processing, Coking and Nuclear Fuel Processing Industry	Anning (Kunming)	China
CNOOC Limited	中国海洋石油有限公司	600938	Oil and Gas Extraction Industry	Hong Kong	China
Kailuan Energy Chemical Co., Ltd.	开滦能源化工股份有限公司	600997	Petroleum Processing, Coking and Nuclear Fuel Processing Industry	Tangshan	China
PetroChina Company Limited	中国石油天然气股份有限公司	601857	Oil and Gas Extraction Industry	Beijing	China

.

The final sample should be understood as the full set of A-share listed petroleum and petrochemical firms that met the study’s disclosure-based inclusion criteria during 2022–2024, rather than as an arbitrary subset. Although the sample size is limited, it reflects the set of observable eligible firms within this industry context that provided the necessary consecutive disclosure texts for structured evaluation. The disclosure texts used in this study are collected mainly from ESG reports, sustainability reports, and social responsibility reports publicly available on CNINFO (Juchao Information Network, an information disclosure website for Chinese listed companies) and the official websites of the sample firms.

3.2. Indicator System and Scoring Rules

Existing research has not yet established a unified framework for defining and evaluating corporate CIDQ, and substantial differences remain across studies in terms of indicator selection, evaluation priorities, and scoring methods [1]. Based on the preceding literature review and in light of the characteristics of the petroleum and petrochemical industry, this study develops an indicator system for evaluating corporate CIDQ.

The main reference bases used for constructing the indicator system are summarized in

Table 2. Basis and corresponding content for the CIDQ evaluation indicators.

Basis/Reference Framework	Corresponding Indicator Content	Role in the Evaluation Framework
The Greenhouse Gas Protocol: A Corporate Accounting and Reporting Standard	Greenhouse gas emission boundaries, Scope 1 and Scope 2 emissions, accounting methods, and reporting boundaries (corresponding to D3 Data Disclosure)	Provides a standardized basis for the accounting and reporting of direct emissions and indirect emissions associated with purchased energy, and clarifies the disclosure scope and norms for basic emissions information.
Corporate Value Chain (Scope 3) Accounting and Reporting Standard	Scope 3 emissions, supply chain carbon management, and value chain coverage (corresponding to D3 Data Disclosure and G2 Supply Chain Management)	Provides a specific reference for the disclosure of indirect emissions across upstream and downstream value chains and related management information, and defines the scope and requirements for value-chain carbon disclosure.
TCFD recommendations and IFRS S2 Climate-related Disclosures	Governance, strategy, risk management, and metrics and targets (corresponding to D1 Strategy and Targets and B2 Disclosure of Negative Information)	Provides a systematic disclosure framework for corporate climate governance arrangements, strategic responses, risk identification, and performance targets, and standardizes the dimensions of climate-related management disclosure.
Characteristics of the petroleum and petrochemical industry	Low-carbon technology development, logistics emission optimization, supply chain emission reduction, and just transition (corresponding to G1 Technological Innovation, G2 Supply Chain Management, and G3 Social Impact)	Reflects the industry-specific characteristics of high emissions, long value chains, and strong transition pressure in the petroleum and petrochemical industry, enhances the industry applicability of the indicator system and its ability to capture differences in disclosure, and addresses the limitations of generic frameworks in industry adaptation.

.

In designing the indicator system, this study draws primarily on the GHG Protocol, the recommendations of the Task Force on Climate-related Financial Disclosures (TCFD), and IFRS S2 Climate-related Disclosures, while also incorporating discussions in the existing literature on disclosure quality dimensions. (TCFD: Available online: https://www.fsb.org/2017/06/recommendations-of-the-task-force-on-climate-related-financial-disclosures-2 (accessed on 1 April 2026). The TCFD provides a climate-related disclosure framework covering governance, strategy, risk management, and metrics and targets. This study draws on this framework in designing related indicators. IFRS S2: Available online: https://www.ifrs.org/issued-standards/ifrs-sustainability-standards-navigator/ifrs-s2-climate-related-disclosures (accessed on 1 April 2026). IFRS S2, issued by the ISSB, sets out climate-related disclosure requirements across governance, strategy, risk management, and metrics and targets. The relevant indicators in this study are developed with reference to its core requirements. GHG Protocol: Available online: https://ghgprotocol.org/standards (accessed on 1 April 2026). Developed by WRI and WBCSD, the GHG Protocol provides a widely used framework for greenhouse gas accounting and reporting. This study draws on this standard in setting emission boundaries, accounting scope, and related indicators.) On this basis, seven first-level dimensions are identified: understandability, reliability, comparability, relevance, completeness, timeliness, and industry specificity.

Ultimately, this study constructs an evaluation system comprising 7 primary indicators, 15 secondary indicators, and 33 tertiary indicators. The primary indicators reflect the core dimensions of CIDQ, the secondary indicators specify the main components of each dimension, and the tertiary indicators correspond to disclosure items that can be directly identified and assessed. In this way, the indicator system is designed to balance theoretical rigor, industry applicability, and operational feasibility. The full indicator system is reported in

Table 3. CIDQ evaluation indicator system.

Primary Indicators	Secondary Indicators	Tertiary Indicators
A. Understandability	A1. Diversity of Expression	A11. Combination of Data and Charts
		A12. Explanation of Professional Terminology
	A2. Structural Clarity	A21. Index/Table of Contents
B. Reliability	B1. Data Accuracy	B11. Disclosure of Calculation Methods
		B12. Third-Party Verification
	B2. Disclosure of Negative Information	B21. Emission Increases or Violation Incidents
		B22. Risk Disclosure
C. Comparability	C1. Consistency in Calculation Methods	C11. Cross-Period Consistency
		C12. Alignment with Industry Standards
	C2. Format Standardization	C21. Stability of Report Format
		C22. Data Standardization
D. Relevance	D1. Strategy and Targets	D11. Emission Reduction Strategy
		D12. Emission Reduction Targets
		D13. Climate-Related Governance
	D2. Actions and Performance	D21. Emission Reduction Measures
		D22. Operation of Emission Reduction Facilities
		D23. Carbon Trading and Offsetting
	D3. Data Disclosure	D31. Scope 1 Emissions
		D32. Scope 2 Emissions
		D33. Scope 3 Emissions
		D34. Product Carbon Footprint
		D35. Carbon Intensity Indicators
E. Completeness	E1. Coverage Scope	E11. Full Value Chain Coverage
		E12. All Emission Sources
	E2. Level of Detail	E21. Data by Department/Region
		E22. Historical Data Comparison
F. Timeliness	F1. Disclosure Time	F11. Timeliness of Report Disclosure
		F12. Regular Updates
G. Industry Specificity	G1. Technological Innovation	G11. Low-Carbon Technology R&D
	G2. Supply Chain Management	G21. Supplier Carbon Management
		G22. Logistics Emission Optimization
	G3. Social Impact	G31. Community Communication
		G32. Just Transition

.

This study adopts a content-analysis-based scoring method and applies a three-level scoring scheme to each tertiary indicator [24]. Specifically, a score of 0 is assigned when the relevant information is not disclosed; a score of 1 is assigned when the firm provides only qualitative descriptions without verifiable quantitative information; a score of 2 is assigned when the disclosure provides quantitative data, measurable targets, or sufficiently specific and verifiable information that satisfies the indicator-specific criterion. In the aggregation process, all tertiary indicators are assigned equal weight [49,50], and the resulting scores are summed to construct AI-CIDQ. Under the current indicator system, the theoretical maximum score of AI-CIDQ is 66.

All 33 tertiary indicators were assigned equal weight because the present study aimed to establish a transparent and replicable rule-based evaluation framework, and no widely accepted industry-specific weighting benchmark is currently available for CIDQ. In the present framework, dimensions such as Relevance, Timeliness, and Completeness are treated as conceptually distinct components of disclosure quality rather than as empirical effect variables requiring differential weights on the basis of observed statistical dominance.

Although PCA-based weighting is a possible alternative, it was not adopted in this study for three reasons. First, PCA-derived weights mainly reflect the covariance structure of the data rather than the theoretical importance of individual indicators. Second, PCA is more suitable when indicators are sufficiently correlated and may be less appropriate when conceptually important indicators are weakly correlated or represent distinct dimensions. Third, prior studies have noted that weighting and aggregation choices in composite indicators may affect interpretability, indicator importance, and ranking robustness [51,52,53]. Therefore, at the initial stage of framework construction, equal weighting is more consistent with the objectives of transparency, interpretability, and replicability adopted in this study. Alternative weighting schemes, including PCA-based or AHP-based approaches, may be explored in future extensions of the framework.

The indicator system is designed as an operational evaluation framework composed of conceptually distinct dimensions of disclosure quality, rather than as a reflective psychometric scale measuring a single latent construct. Therefore, CFA and internal-consistency statistics such as Cronbach’s alpha were not applied in the present study.

In addition to the general 0–1–2 scoring logic, the present study operationalized the rubric through indicator-specific decision rules. For each tertiary indicator, a score of 0 was assigned when no relevant disclosure was identified; a score of 1 was assigned when the firm provided qualitative or general descriptive information without verifiable quantitative support; and a score of 2 was assigned when the disclosure included quantitative data, measurable targets, or other verifiable evidence directly relevant to the indicator. To improve consistency and reproducibility, a detailed operational scoring guide for all 33 tertiary indicators is provided in Appendix B. The appendix reports the indicator descriptions, the 0–1–2 scoring criteria, and brief boundary notes that clarify how each criterion was applied.

In addition, several tertiary indicators with relatively broad conceptual scope were further operationalized to improve scoring transparency. Specifically, “Risk Disclosure” refers to the disclosure of climate-related transition risks, regulatory risks, market risks, or physical risks associated with carbon management, rather than general statements of environmental commitment. “Full Value Chain Coverage” refers to whether the report extends beyond direct operational emissions to cover major upstream and downstream carbon-related activities, such as purchased materials, transportation, product use, or logistics arrangements. “Just Transition” refers to disclosures concerning the social implications of low-carbon transition, including employee reskilling, workforce adjustment, community impact, or stakeholder communication in transition processes.

3.3. LLM-Based Evaluation Process

To improve the efficiency of text processing and maintain consistency in scoring standards, this study employed an LLM to assist in the evaluation of corporate carbon disclosure content. Because ESG reports, social responsibility reports, and sustainability reports are often lengthy, contain substantial professional terminology, and differ across firms in terms of expression, information distribution, and disclosure scope, a fully manual item-by-item assessment would be costly and difficult to standardize consistently. Considering the model’s Chinese-language capability and accessibility, this study uses the DeepSeek-V3.2 LLM. (In this study, the DeepSeek-V3.2 model, accessed through the DeepSeek–chat interface, was used only as an auxiliary tool for rule-based scoring of corporate disclosure texts under predefined criteria. It was not used to draft, revise, or polish the manuscript. All research design, indicator construction, result interpretation, and manuscript writing were completed by the authors.) Prior studies have also shown the potential of LLMs in sustainability reporting and structured disclosure evaluation [11,12]. The temperature parameter was set to 0 in order to reduce output randomness and improve consistency in rule-based scoring tasks. This setting was intended to enhance result stability in practice, but it should not be interpreted as guaranteeing complete determinism across repeated runs.

To improve methodological transparency, the scoring prompts were designed indicator by indicator on the basis of the predefined indicator description and the corresponding 0–1–2 scoring rule. Each prompt asked the model to determine whether the relevant information was disclosed, whether the disclosure was qualitative only or also supported by verifiable quantitative information, and to return the score together with a brief rationale, textual evidence, and an evidence-location field. Because the reports were often lengthy and structurally complex, the full reports were first converted into machine-readable text while preserving the original section order and major headings as far as possible. For each tertiary indicator, the evaluation was conducted at the full-document level rather than on isolated text fragments, so that the model could identify relevant information dispersed across different sections of the same report. The present study did not employ an external retrieval-augmented generation module or a separate grounding system; instead, contextual consistency was maintained by combining the full report text with indicator-specific prompts that instructed the model to locate, interpret, and score disclosure content according to predefined rules. A representative prompt used in the evaluation process is provided in Appendix A.

The evaluation process consists of three steps. First, the ESG reports, social responsibility reports, and sustainability reports of the sample firms were collected and organized, and the full reports were converted into machine-readable text. Second, based on the predefined indicator system, indicator descriptions, and scoring rules, scoring prompts were constructed separately for each tertiary indicator and entered into the LLM together with the full report text. The model then identified information relevant to the current indicator across the entire document and returned the corresponding score, scoring rationale, textual evidence, and evidence-location information. Third, the model outputs were aggregated and reviewed. Under the equal-weight assumption, the scores of all tertiary indicators were then summed to construct a company-year AI-CIDQ score.

Figure 1 shows the automated LLM-based scoring process for CIDQ. It presents the main steps of the evaluation procedure, including text input, rule-based prompt design, model scoring, structured output generation, and score aggregation.

Table 4. AI-CIDQ scores of sample firms (2022–2024).

Company Name (English)	Company Abbreviation	City	Country	2022	2023	2024
North Huajin Chemical Industries Co., Ltd.	North Huajin	Panjin	China	31	40	32
Shenyang Chemical Co., Ltd.	Shenyang Chemical	Shenyang	China	32	38	41
Shanxi Meijin Energy Co., Ltd.	Meijin Energy	Taiyuan	China	48	49	47
Yueyang Xingchang Petro-Chemical Co., Ltd.	Yueyang Xingchang	Yueyang	China	9	23	39
Shanxi Blue Flame Holding Company Limited	Blue Flame Holdings	Taiyuan	China	7	11	24
China Petroleum & Chemical Corporation	Sinopec	Beijing	China	47	55	56
Guanghui Energy Co., Ltd.	Guanghui Energy	Urumqi	China	27	37	33
Sinopec Shanghai Petrochemical Company Limited	Sinopec Shanghai Petrochemical	Shanghai	China	51	51	52
Yunnan Yunwei Co., Ltd.	Yunnan Yunwei	Kunming	China	11	10	11
Shanxi Coking Co., Ltd.	Shanxi Coking	Hongtong (Linfen)	China	29	29	41
Geo-Jade Petroleum Corporation	Geo-Jade Petroleum	Haikou	China	12	26	27
Yunnan Coal & Energy Co., Ltd.	Yunnan Coal & Energy	Anning (Kunming)	China	17	26	15
CNOOC Limited	CNOOC	Hong Kong	China	46	48	49
Kailuan Energy Chemical Co., Ltd.	Kailuan Energy Chemical	Tangshan	China	18	35	44
PetroChina Company Limited	PetroChina	Beijing	China	51	53	53

reports the AI-CIDQ scores of the sample firms from 2022 to 2024. It presents the firm-level scoring results and provides the basis for the subsequent analysis of annual trends and inter-firm differences.

After model scoring, manual review was conducted only for cases involving ambiguous semantics, negative statements, or insufficient evidence. This review was applied as a targeted verification procedure rather than as a full manual recoding of all reports or a blinded multi-coder coding exercise. The purpose was to examine whether the score returned by the model was adequately supported by the identified textual evidence under the predefined scoring rules, and to correct possible misjudgments where necessary.

3.4. Reliability and Consistency Test

To examine the reliability and consistency of the LLM-based scores, this study randomly selected 15 reports from the 45 sample reports and conducted manual scoring independently of the LLM-based scores using the unified tertiary-indicator system and the 0–1–2 scoring rules. The reliability assessment concerns the consistency between LLM-based scores and manually assigned benchmark scores, rather than interrater reliability among multiple independent human coders.

The main evaluation process involved 45 firm-year reports and 33 tertiary indicators, resulting in 1485 indicator-level LLM scoring tasks. In the reliability test, the 15 sampled reports were manually scored at the tertiary-indicator level, yielding 495 paired observations. These figures indicate the operational scale of the computational and annotation process in the present study. It should be noted that

Table 5. Comparison of Manual-CIDQ and AI-CIDQ scores for sampled reports.

Company Name (English)	City	Country	Year	AI-CIDQ	Manual-CIDQ
China Petroleum & Chemical Corporation	Beijing	China	2022	47	48
Sinopec Shanghai Petrochemical Company Limited	Shanghai	China	2023	51	52
CNOOC Limited	Hong Kong	China	2024	49	50
Shanxi Meijin Energy Co., Ltd.	Taiyuan	China	2022	48	49
PetroChina Company Limited	Beijing	China	2023	53	54
North Huajin Chemical Industries Co., Ltd.	Panjin	China	2024	32	33
Shenyang Chemical Co., Ltd.	Shenyang	China	2022	32	33
Guanghui Energy Co., Ltd.	Urumqi	China	2023	37	38
Shanxi Coking Co., Ltd.	Hongtong (Linfen)	China	2024	41	42
Kailuan Energy Chemical Co., Ltd.	Tangshan	China	2022	18	19
Yueyang Xingchang Petro-Chemical Co., Ltd.	Yueyang	China	2023	23	24
Shanxi Blue Flame Holding Company Limited	Taiyuan	China	2024	24	25
Yunnan Yunwei Co., Ltd.	Kunming	China	2022	11	11
Geo-Jade Petroleum Corporation	Haikou	China	2023	26	27
Yunnan Coal & Energy Co., Ltd.	Anning (Kunming)	China	2024	15	15

reports only the total-score comparison for the 15 sampled reports, whereas the consistency statistics in

Table 6. Consistency test results based on 495 paired tertiary-indicator scores ¹.

Test Indicator	Test Method/Type	Statistical Value	p-Value	Criterion	Result Explanation
Overall Score Correlation	Pearson	0.975	p < 0.001	r > 0.7 indicates high correlation	LLM-based scores are highly correlated with manual scores at the tertiary-indicator level
Overall Score Correlation	Spearman	0.975	p < 0.001	ρ > 0.7 indicates high correlation	High consistency in ranking at the tertiary-indicator level
Systematic Bias	Paired t-test	t = −3.007	0.003	p < 0.05 indicates significant difference	LLM-based tertiary-indicator scores are slightly lower than manual scores overall
Overall Consistency	ICC (Single Measurement, Absolute Agreement)	0.974	p < 0.001	ICC > 0.75 indicates good consistency	LLM-based scoring consistency is good
Mean Absolute Error	MAE	0.0384	-	Smaller MAE indicates higher consistency	The average error between LLM-based and manual is approximately 0.0384

¹ Note: The consistency statistics reported in Table 6 are based on 495 paired observations at the tertiary-indicator level (15 sampled reports × 33 indicators), rather than on the 15 total-score observations reported in Table 5.

are based on the 495 paired tertiary-indicator observations. On this basis, this study uses the Pearson correlation coefficient, the Spearman rank correlation coefficient, the paired-samples t-test, the intraclass correlation coefficient (ICC) [54], and the mean absolute error (MAE) [55] to assess the consistency between the LLM-based scores and the manual scores. The analysis is then further extended to the seven primary indicators in order to examine the stability of the LLM’s scoring performance across different dimensions. The detailed results are reported in Table 5, Table 6 and

Table 7. Consistency test results by primary indicator.

Primary Indicator	Pearson	Spearman	p-Value	Correlation Criterion	Test Result
A. Understandability	0.977	0.984	<0.001	Yes	Pass
B. Reliability	1.000	1.000	<0.001	Yes	Pass
C. Comparability	0.989	0.986	<0.001	Yes	Pass
D. Relevance	0.998	0.992	<0.001	Yes	Pass
E. Completeness	0.989	0.986	<0.001	Yes	Pass
F. Timeliness	1.000	1.000	<0.001	Yes	Pass
G. Industry Specificity	0.976	0.978	<0.001	Yes	Pass

.

From the comparison of total scores, the LLM-based scores show a high degree of consistency with the manual scores. As shown in Table 5, with the exception of two sampled reports—Yunnan Yunwei (2022) and Yunnan Coal & Energy (2024)—the LLM-based score is 1 point lower than the manual score for all other sampled reports, with no substantial deviation observed. For example, the scores for Sinopec (2022) are 47 and 48, respectively; those for PetroChina (2023) are 53 and 54, respectively; and those for Sinopec Shanghai Petrochemical (2023) are 51 and 52, respectively.

Table 6 further reports the consistency test results based on 495 paired observations at the tertiary-indicator level. The Pearson and Spearman correlation coefficients are both 0.975 and are statistically significant at the 1% level, indicating that the LLM-based scores and the manual scores are highly consistent in terms of both score variation and sample ranking. The paired-samples t-test yields t = −3.007 (p = 0.003), suggesting that the LLM-based scores are marginally lower than the manual scores overall. The ICC of 0.974 indicates a high degree of absolute agreement between the two scoring approaches. The MAE of 0.0384 further suggests that the average absolute deviation between the LLM-based scores and the manual scores is small at the tertiary-indicator level.

Table 7 shows that the results at the primary-indicator level are broadly consistent with those at the total-score level. Across all seven dimensions—understandability, reliability, comparability, relevance, completeness, timeliness, and industry specificity—the Pearson correlation coefficients and Spearman rank correlation coefficients range from 0.976 to 1.000 and are all statistically significant at the 1% level. This indicates that the LLM-based scores remain highly consistent with the manual scores across all dimensions, demonstrating overall stability in scoring performance.

Overall, the differences between the LLM-based scores and the manual scores are small, and the degree of consistency between them is high. At the same time, the paired-samples t-test indicates that the LLM-based scores are, on average, slightly lower than the manual scores. These results suggest that the scoring procedure developed in this study has good reliability and can maintain a high level of agreement with manual scoring under the predefined evaluation criteria, while also exhibiting a modest downward scoring tendency that should be taken into account in interpretation.

Although all seven primary-indicator dimensions show high consistency with manual scoring, the “Industry Specificity” dimension yields the lowest agreement coefficients among them. This does not indicate weak performance in an absolute sense, since its Pearson and Spearman correlations remain above 0.97, but it may suggest that industry-specific disclosure items are comparatively more difficult to evaluate consistently. This is likely because such disclosures often involve technically complex and dispersed information, including Scope 3 emissions, supply chain carbon management, low-carbon technology, and transition-related social arrangements.

4. Results

This section reports measurement-oriented descriptive evidence on CIDQ under the proposed evaluation framework. Because the present study is designed as a framework-construction and measurement study rather than as a hypothesis-testing study, the results focus on the distribution, annual changes, firm-level heterogeneity, and dimension-level characteristics of the constructed CIDQ index.

4.1. Overall Analysis

Table 8. Overall descriptive statistics of AI-CIDQ scores.

Statistic	Value
Sample Size (N)	45
Mean	34.02
Median	35.00
Standard Deviation	14.98
Minimum	7.00
Maximum	56.00
25th Percentile (P25)	24.00
75th Percentile (P75)	48.00
Skewness	−0.301
Kurtosis	−1.161

reports the descriptive statistics of the AI-CIDQ scores for the full sample. The statistics reported in Table 8 are intended to characterize the observed distribution of CIDQ in the study sample, including its central tendency, dispersion, and shape, rather than to provide inferential estimates of population parameters.

The mean score is 34.02, the median is 35.00, the standard deviation is 14.98, the minimum is 7.00, and the maximum is 56.00. The 25th and 75th percentiles are 24.00 and 48.00, respectively. Given that the theoretical maximum of the index is 66, the sample mean accounts for only 51.55% of the upper limit. This result indicates that the overall level of CIDQ in the petroleum and petrochemical industry remains relatively limited. At the same time, the score range is wide, and the interquartile range reaches 24 points, indicating substantial differences in disclosure quality across firms.

4.2. Annual Analysis

To examine changes in CIDQ during the study period, this study reports annual descriptive statistics of the AI-CIDQ scores in

Table 9. Annual descriptive statistics of AI-CIDQ scores (2022–2024).

Year	Sample Size (N)	Mean	Median	Standard Deviation	Minimum	Maximum
2022	15	29.07	29.00	16.30	7.00	51.00
2023	15	35.40	37.00	14.45	10.00	55.00
2024	15	37.60	41.00	13.72	11.00	56.00

. These statistics are intended to characterize temporal variation within the observed sample under the proposed evaluation framework. Accordingly, the year-to-year changes are interpreted as measurement-oriented descriptive evidence of how the constructed index evolved over time, rather than as the outcome of formal repeated-measures hypothesis testing.

The annual statistics show that the CIDQ of the sample firms exhibited a continuous upward trend from 2022 to 2024. The mean score increased from 29.07 in 2022 to 35.40 in 2023 and further to 37.60 in 2024, representing a cumulative increase of 8.53 points, or 29.34%, over the three-year period. The median also rose from 29.00 to 41.00, indicating that the improvement in disclosure quality was not driven solely by a small number of high-scoring firms, but reflected an overall improvement across the sample. In terms of dispersion, the differences in scores across firms showed some degree of convergence. The standard deviation declined from 16.30 in 2022 to 13.72 in 2024, while the minimum score increased from 7.00 to 11.00, suggesting that firms with relatively low levels of disclosure also improved. The maximum score rose from 51.00 to 56.00, indicating that higher-scoring firms continued to improve on the basis of their already relatively strong performance. Overall, during the study period, the disclosure quality of the sample firms was characterized by an overall upward trend, narrowing disparities, and persistent heterogeneity.

Figure 2 illustrates the annual trend in mean CIDQ scores. The largest increase occurred from 2022 to 2023, suggesting that this was the period of fastest improvement among the sample firms. From 2023 to 2024, the mean score continued to rise, but at a slower pace.

Figure 3 provides further evidence of this trend. The boxplots show an upward shift in the score distribution over the three years, although the differences across firms did not disappear. Overall, the CIDQ of China’s A-share petroleum and petrochemical firms improved continuously from 2022 to 2024. However, by 2024, the mean score had reached only 56.97% of the theoretical maximum, indicating that the industry as a whole remained in a transitional stage from basic disclosure toward higher-quality disclosure.

The upward trend in annual mean scores may reflect several overlapping factors. These include the continued strengthening of climate-related regulatory expectations in China, firms’ accumulated experience in carbon-related reporting, and the gradual standardization of ESG and sustainability disclosure practices. Taken together, these factors may have contributed to the steady improvement in CIDQ observed during the sample period.

4.3. Firm-Level Heterogeneity in CIDQ

To further examine firm-level differences in CIDQ, this study reports the annual AI-CIDQ scores of the 15 sample firms for 2022–2024, together with their three-year average scores, in

Table 10. AI-CIDQ scores and three-year averages by firm.

Company Name (English)	City	Country	2022 Score	2023 Score	2024 Score	Average Score
China Petroleum & Chemical Corporation	Beijing	China	47	55	56	52.67
PetroChina Company Limited	Beijing	China	51	53	53	52.33
Sinopec Shanghai Petrochemical Company Limited	Shanghai	China	51	51	52	51.33
Shanxi Meijin Energy Co., Ltd.	Taiyuan	China	48	49	47	48.00
CNOOC Limited	Hong Kong	China	46	48	49	47.67
Shenyang Chemical Co., Ltd.	Shenyang	China	32	38	41	37.00
North Huajin Chemical Industries Co., Ltd.	Panjin	China	31	40	32	34.33
Shanxi Coking Co., Ltd.	Hongtong (Linfen)	China	29	29	41	33.00
Guanghui Energy Co., Ltd.	Urumqi	China	27	37	33	32.33
Kailuan Energy Chemical Co., Ltd.	Tangshan	China	18	35	44	32.33
Yueyang Xingchang Petro-Chemical Co., Ltd.	Yueyang	China	9	23	39	23.67
Geo-Jade Petroleum Corporation	Haikou	China	12	26	27	21.67
Yunnan Coal & Energy Co., Ltd.	Anning (Kunming)	China	17	26	15	19.33
Shanxi Blue Flame Holding Company Limited	Taiyuan	China	7	11	24	14.00
Yunnan Yunwei Co., Ltd.	Kunming	China	11	10	11	10.67

. The firm-level comparison is intended to describe cross-firm variation in the constructed CIDQ index within the observed sample. The descriptive labels used in this subsection are based on observed score levels and rankings, rather than on statistically validated clusters derived from formal group-comparison tests.

At the firm level, the results indicate noticeable heterogeneity in CIDQ. Sinopec, PetroChina, Sinopec Shanghai Petrochemical, Meijin Energy, and CNOOC rank at the top, with three-year average scores of 52.67, 52.33, 51.33, 48.00, and 47.67, respectively. Relative to the full-sample mean of 34.02, these five firms score 18.65, 18.31, 17.31, 13.98, and 13.65 points higher, respectively, indicating that these higher-scoring firms performed above the sample average in descriptive terms. Among them, the score ranges of PetroChina, Sinopec Shanghai Petrochemical, Meijin Energy, and CNOOC over the three years are only 2, 1, 2, and 3 points, respectively, suggesting that their disclosure performance is not only relatively strong but also comparatively stable.

Table 10 reports the detailed firm-level scores and three-year averages.

By contrast, Yunnan Yunwei, Blue Flame Holdings, and Yunnan Coal & Energy record three-year average scores of only 10.67, 14.00, and 19.33, respectively, which are 23.35, 20.02, and 14.69 points below the sample mean. Yunnan Yunwei records scores of 11, 10, and 11 over the three years, with a maximum fluctuation of only 1 point, indicating that it remained at a relatively low disclosure level throughout the sample period, with little sign of improvement. Blue Flame Holdings increases from 7 to 24, representing a cumulative increase of 17 points, but its 2024 score still reaches only 63.83% of the 2024 sample mean of 37.60. Yunnan Coal & Energy records scores of 17, 26, and 15, respectively; with its 2024 score 11 points lower than that of 2023, indicating relatively large fluctuations. These results suggest that some low-scoring firms still have a weak foundation for carbon information disclosure, and that their improvement lacks continuity.

The middle tier also shows considerable variation. Shenyang Chemical, North Huajin, Shanxi Coking, Guanghui Energy, and Kailuan Energy Chemical record three-year average scores of 37.00, 34.33, 33.00, 32.33, and 32.33, respectively, which are close to or slightly below the sample mean. Among them, Kailuan Energy Chemical increases from 18 to 44, representing a cumulative increase of 26 points. Yueyang Xingchang increased from 9 to 39, with a cumulative increase of 30 points over the three years, representing the largest increase in the full sample. These results indicate that some medium- and low-scoring firms began to strengthen carbon information disclosure markedly during the study period, but had not yet developed a stable pattern of high-quality disclosure.

In terms of inter-firm gaps, the top-ranked Sinopec records a three-year average score of 52.67, whereas the bottom-ranked Yunnan Yunwei records only 10.67, producing a difference of 42.00 points; the former is approximately 4.94 times the latter. Looking only at 2024, the highest score is 56 and the lowest is 11, producing a difference of 45 points; the former is approximately 5.09 times the latter. Overall, the results reveal substantial firm-level heterogeneity in CIDQ among China’s petroleum and petrochemical enterprises. Higher-quality disclosure is concentrated among a relatively small number of higher-scoring firms, whereas medium- and low-scoring firms, despite some improvement, differ considerably in both the extent and stability of that improvement. A balanced and broadly consistent disclosure pattern has not yet emerged within the industry. The persistently high scores of higher-scoring firms such as China Petroleum & Chemical Corporation and PetroChina Company Limited may reflect their relatively stronger governance foundations, more mature data collection and management systems, and greater experience in continuous environmental and climate-related disclosure. By contrast, firms with relatively low scores may still face constraints in disclosure capability, internal reporting infrastructure, and the continuity of carbon-related information management, which may limit both the level and stability of their disclosure performance.

5. Discussion

Taken together, the results show that the carbon information disclosure quality of China’s A-share petroleum and petrochemical firms improved from 2022 to 2024, but the overall disclosure level remained limited and firm-level differences persisted. This pattern suggests that the industry is still moving from basic carbon-related disclosure toward more complete, comparable, and verifiable disclosure. In the petroleum and petrochemical industry, carbon-related reporting often involves technically complex emissions accounting, value-chain-related information, transition-risk disclosure, and industry-specific governance arrangements. These features increase the difficulty of producing high-quality disclosure, especially for firms with weaker internal data systems or less experience in sustainability reporting.

The improvement observed during the sample period may be related to several overlapping factors, including stronger climate-related disclosure expectations in China, firms’ accumulated experience in ESG and sustainability reporting, and the gradual standardization of non-financial reporting practices. At the same time, the persistent differences across firms suggest that disclosure quality is not determined only by report length or the amount of disclosed information. Rather, it may also depend on firms’ governance foundations, internal data collection and management capacity, and the continuity of carbon-related disclosure practices. Higher-scoring firms such as China Petroleum & Chemical Corporation and PetroChina Company Limited may be better able to organize carbon-related data and maintain consistent reporting practices, whereas firms with relatively low scores may still face constraints in data availability, internal reporting infrastructure, and cross-departmental coordination. These findings are broadly consistent with prior studies showing that carbon disclosure is influenced by firms’ environmental performance, governance structures, external institutional pressures, and stakeholder expectations, and that climate-related disclosure may be associated with improvements in environmental governance and disclosure practices over time [10,16,17,18,19,20,21]. The present study adds industry-specific evidence by showing that, in a high-emission sector with technically complex and value-chain-related disclosure requirements, firm-level differences in disclosure quality may remain substantial even as reporting practices become more common.

The results can also be interpreted in light of legitimacy theory and stakeholder theory. From a legitimacy perspective, the improvement in CIDQ may reflect firms’ efforts to maintain social and regulatory acceptance as climate governance and sustainability disclosure expectations become stronger. From a stakeholder perspective, carbon disclosure provides a channel through which firms communicate information on emissions, reduction actions, climate-related risks, and transition progress to investors, regulators, communities, and other stakeholders. The persistence of firm-level differences suggests that firms differ not only in their willingness to disclose, but also in their capacity to provide disclosure that is specific, consistent, and supported by verifiable information [15,21].

Viewed against existing climate-disclosure frameworks, the weaker areas identified in this study are closely related to the core dimensions emphasized by TCFD and IFRS S2, including governance, strategy, risk management, and metrics and targets. The proposed framework should therefore be understood not as a substitute for these broader frameworks, but as an industry-specific operationalization of their key disclosure expectations. In the Chinese context, where sustainability reporting requirements have recently become more standardized, such a framework may provide a more detailed diagnostic tool for high-emission industries. More specifically, regulators and standard setters may use indicator-level results to identify disclosure items that require more targeted guidance, such as emission boundaries, Scope 3 emissions, calculation methods, transition-risk disclosure, value-chain coverage, and just transition arrangements. The framework may also help translate broad sustainability reporting requirements into more concrete industry checklists, for example by encouraging firms to disclose calculation bases, historical comparability, quantitative targets, and evidence for emission-reduction measures. For stock exchanges and industry associations, the framework may provide a reference for reviewing whether high-emission firms disclose not only general commitments but also verifiable data, implementation details, and value-chain-related information. For firms, the results suggest that improving CIDQ requires not only increasing the amount of report content, but also strengthening the completeness, comparability, and verifiability of disclosed information.

Methodologically, the reliability test suggests that LLM-assisted scoring can provide relatively stable support for structured CIDQ evaluation when explicit scoring rules, structured output requirements, textual evidence, and manual review are retained. At the same time, the LLM should be understood as an assistive tool rather than as an autonomous evaluator. This boundary is important because LLM outputs may contain unsupported interpretations or hallucinated inferences when processing lengthy and technically complex ESG reports. The current validation also focuses on consistency with manual benchmark scores and should not be interpreted as proving superiority over keyword retrieval, conventional text-classification methods, or other LLMs. Future applications of the framework should therefore continue to combine automated scoring with evidence-location requirements and researcher verification.

Overall, the findings provide measurement-oriented and industry-specific evidence on the current state of CIDQ in China’s petroleum and petrochemical industry. They show that disclosure quality improved during the sample period, but that the industry still faces challenges in moving from basic disclosure toward more complete, comparable, and verifiable disclosure. The findings should therefore be read as industry-specific and temporally comparative evidence rather than as causal or broadly generalizable conclusions.

6. Conclusions

This study developed an industry-specific framework for evaluating carbon information disclosure quality in China’s petroleum and petrochemical industry and applied an LLM-assisted scoring procedure to 45 firm-year observations from 2022 to 2024. The results show that CIDQ improved during the sample period, but the overall disclosure level remained relatively limited and substantial differences persisted across firms. These findings indicate that the industry has made progress in carbon-related reporting, while still facing challenges in providing complete, comparable, and verifiable disclosure, particularly for technically complex and value-chain-related items.

The study contributes to CIDQ research in two main respects. First, it develops a structured 7/15/33 evaluation framework that combines general disclosure-quality dimensions with the characteristics of the petroleum and petrochemical industry, including Scope 3 emissions, value-chain coverage, supply chain carbon management, low-carbon technology development, logistics emission optimization, and just transition. Second, it demonstrates that LLM-assisted scoring can support rule-based disclosure evaluation when explicit scoring criteria, structured output requirements, textual evidence, and manual review are retained. In this sense, the study provides a methodological reference for structured CIDQ evaluation in high-emission industries.

The study also has practical implications. For regulators and standard setters, the framework may serve as a diagnostic tool for identifying weak areas in carbon disclosure among petroleum and petrochemical firms, especially in value-chain-related and transition-related disclosure areas. It may also provide a reference for developing more detailed industry-specific disclosure guidance within the broader movement toward standardized sustainability reporting. For firms, the findings point to the need to improve internal carbon-data management and to provide more specific, consistent, and evidence-supported disclosure, rather than merely increasing the length or generality of report content.

Several limitations should be noted. First, the sample is limited to 15 A-share listed petroleum and petrochemical firms in China. Although the sample represents the firms that met the study’s disclosure-based inclusion criteria during 2022–2024, the external validity of the findings still requires further examination. Therefore, the observed annual changes and firm-level differences should be interpreted as measurement results within the defined sample, rather than as population-level or causal evidence for the entire petroleum and petrochemical industry. Future research may apply the framework to larger samples, longer time periods, other high-emission industries, or different country contexts to examine its broader applicability. Second, the construction of the index involves researcher-defined indicator selection, scoring boundaries, and equal-weight aggregation. Although these choices were made to preserve transparency, interpretability, and replicability, alternative weighting schemes or different indicator groupings may affect aggregate scores and firm rankings. Future research may therefore conduct sensitivity analyses and compare equal weighting with expert-based, PCA-based, AHP-based, or other weighting approaches where broader data and different research objectives are available. Third, the validation of the scoring results can be further strengthened. The present study compares LLM-based scores with manual benchmark scores, but it does not use a multi-coder human validation design or repeated-run stability tests under identical prompts. It also does not provide systematic comparisons with keyword retrieval, conventional text-classification methods, or other LLMs. In addition, LLM-assisted scoring may be affected by prompt sensitivity, model-version changes, context loss in long reports, hallucinated or unsupported interpretations, and a slight downward scoring tendency relative to manual scores. For this reason, the LLM-based procedure should be used as an assistive tool under explicit rules and evidence verification, rather than as a replacement for expert judgment. Future research may extend the validation design by involving multiple independent coders, testing model stability under different settings, comparing alternative automated methods, and examining the sensitivity of aggregated CIDQ scores to different weighting schemes. Finally, because this study is designed as a framework-construction and measurement study, the findings should be interpreted as measurement-oriented and industry-specific evidence rather than as causal conclusions about the determinants or consequences of CIDQ.