Pith. sign in

REVIEW 3 major objections 4 minor 54 references

The paper measures, at panel scale, how much corporate building-decarbonisation disclosure actually supports the per-square-metre stranding assessments that New York's Local Law 97 and the EU's revised Buildings Directive presuppose, and fi

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 06:04 UTC pith:NTNEAVNW

load-bearing objection A genuinely first panel-scale measurement of the per-m² disclosure gap, with the most careful LLM reliability battery in this literature — but the headline rate rests on extractor nulls that have not been checked against human ground truth. the 3 major comments →

arxiv 2607.22006 v1 pith:NTNEAVNW submitted 2026-07-24 cs.CY econ.GNq-fin.ECstat.AP

Unfit for stranding assessment: a panel-scale multimodal-LLM audit of building-decarbonisation disclosure (BeDA)

Jingyi Xu , Minghui Cheng , Anchen Sun This is my paper
classification cs.CY econ.GNq-fin.ECstat.AP
keywords building decarbonisationdisclosure-fitness gapper-m² carbon intensityCRREM stranding readinessmultimodal large language modelsurban climate policysustainability disclosure auditLocal Law 97 / EPBD
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that most corporate building-decarbonisation disclosure is unfit for the stranding assessments regulators are now enacting. Using a multimodal large-language-model auditor applied to a global panel of 2,246 firms, it finds that only 21.5% of built-environment firm-reports disclose operational carbon intensity per square metre per year, the denominator that CRREM-style 1.5-degree pathways and the EU's per-m² energy-performance rules require. The rate is roughly twice as high in Europe as in the United States, indicating a jurisdiction-specific reporting gap rather than a global structural limit. If this measurement is correct, enforceable building-stranding regulation is currently bottlenecked at the disclosure source: better downstream modelling cannot substitute for data that firms simply do not report. The paper also shows that a standards-compliance credibility score is stable across model families, robust to firm-identity leakage, and modestly convergent with independent external criteria, so the measurement instrument itself can be credibly used at scale.

Core claim

The central claim is that only about one built-environment firm-report in five (21.5%, Wilson 95% CI [16.4%, 27.7%]) discloses operational carbon intensity per m²/yr, the CRREM-fit denominator, with 45.5% among 519 real-estate firm-reports. Among the 215 real-estate reports for which an intensity can be constructed, 39% already exceed the CRREM 1.5°C pathway's intensity limit at their reporting year. The paper further claims that credibility does not predict stranding readiness once portfolio size is controlled, so the stranding profile is a screening tool, not a forecast. The authors argue that the main obstacle to enforceable building-stranding regulation is therefore a measurable, jurisdi

What carries the argument

The load-bearing mechanism is BeDA, a four-phase multimodal-LLM pipeline. Its credibility component is a 'standards-compliance spine' score that measures alignment with recognised disclosure frameworks; this score is shown to be stable across model families (r ≈ 0.74–0.82) and robust to identity-anonymisation leakage. Its disclosure-fitness component is a strict extract-only LLM extractor that returns null for anything not explicitly disclosed in a report PDF, paired with a deterministic rule-based classifier that assigns each report to one of seven disclosure-fitness classes—operational carbon per m²/yr (CRREM-fit), energy-use intensity per m²/yr (EPBD-fit), per-unit or embodied, revenue/ou

Load-bearing premise

The headline 21.5% rate assumes that the LLM extractor returns 'missing' exactly when a per-m² intensity is absent from a report, and that the sample of available report PDFs is representative of the target population of built-environment firms.

What would settle it

Re-extract the 200 stratified firm-reports with two independent human coders instructed to find any operational carbon intensity per m²/yr anywhere in the PDF, including tables, figures, and footnotes; if the true proportion exceeds roughly 40%, the central claim that most disclosure is unfit collapses. A second falsifier would be an independent non-Google multimodal extractor reproducing the classification on the same 200 reports; if the CRREM-fit rate drops below the reported interval's lower bound, the instrument itself is the source of the gap.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the 21.5% rate is correct, regulators and portfolio owners cannot currently strand-test most building portfolios from corporate disclosure; a per-m² intensity mandate is the direct policy remedy.
  • The roughly 2× higher European disclosure rate implies the gap is a closable policy choice, not an inherent property of building emissions, and can be monitored automatically as mandates take effect.
  • Credibility scores alone cannot substitute for the missing denominator: even a perfectly credible report is unusable for stranding if its intensity is normalised on revenue or output rather than floor area.
  • The finding that credibility does not predict stranding readiness once portfolio size is controlled means transition-risk screens should be treated as triage tools, not forecasts of physical stranding.
  • A targeted, jurisdiction-specific disclosure mandate—for example, requiring the per-m² denominator in regulated reporting schemas—would directly increase the fraction of reports that can feed science-based building-stranding assessment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: The strict extract-only protocol may undercount intensities embedded in complex tables, charts, or infographics that the modal model reads imperfectly; a human audit of a random subset of reports flagged as 'no usable intensity' would directly test this and is a natural next step.
  • Editorial extension: The paper's framing implies that portfolio-level transition-risk screening by investors is currently far more constrained for US real-estate holdings than for European ones; this could be tested by comparing investor-facing risk assessments against the disclosure rates reported here.
  • Editorial extension: The 39% above-pathway screen, though explicitly not a forecast, could be repurposed as a monitoring indicator that tracks whether new disclosure mandates are moving the distribution of disclosed intensities toward the pathway over time.
  • Editorial extension: Because the paper counts only operational carbon fitness, a complete built-environment audit would pair this operational axis with an embodied-carbon axis for new construction, where per-home or per-unit denominators are architecturally appropriate and should not be penalised as reporting failures.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces BeDA, a multimodal-LLM pipeline that scores the credibility of corporate sustainability reports and measures whether the disclosed data are fit for per-square-metre building-stranding assessment. Applying the instrument to a global panel (2,246 firms, 9,166 firm-years, 2003–2023) and to a 519-report real-estate corpus, the paper reports that only 21.5% of a region-stratified sample of 200 built-environment firm-reports disclose operational carbon intensity per m²/yr (Wilson 95% CI [16.4%, 27.7%]), and 45.5% of real-estate firm-reports do so, with a marked US/EU gap. The paper also reports reliability and validity evidence for the credibility spine: cross-model and cross-family concordance, leakage tests, and convergence with LSEG-ESG, RepRisk incidents, and SBTi validation. The central policy claim is that enforceable building-stranding regulation is currently bottlenecked by a jurisdiction-specific disclosure gap rather than by modelling capacity.

Significance. If the headline measurement is correct, the paper provides the first panel-scale, jurisdiction-resolved quantification of a widely assumed but unmeasured disclosure-fitness gap, with direct relevance to Local Law 97, the EPBD recast, and portfolio-level transition-risk screening. The study is commendably explicit about its pipeline, reports cross-extractor agreement, distinguishes reliability from validity, and tests confound controls that most prior LLM-based disclosure work omits. The design of the reliability battery—multi-model, multi-family, with a text-only modality arm—is a useful template. However, the central estimate rests on an LLM extractor whose null outputs have not been validated against human ground truth; the evidence currently rules out random extractor noise but not systematic false negatives. That gap in validation is load-bearing and must be addressed before the 'one in five' figure can be treated as a secure measurement.

major comments (3)
  1. [Section 4.4, Appendix E, Section 5.5] The headline 21.5% CRREM-fit rate depends on the P4 extractor returning null exactly when a metric is absent. The validation is inter-extractor agreement between gemini-3.1-pro-preview and gemini-3.5-flash (κ=0.95 on 50 reports; κ=0.97 on 79), and Section 5.5 states that 'a human audit of the extracted numbers is a stated extension.' Two same-vendor models can share a systematic blind spot for tables or figures, so their agreement does not establish sensitivity. Moreover, Cohen's κ on the 7-class taxonomy is inflated by the skewed class distribution; a model biased toward 'no usable intensity' would agree highly with another such model. Please report class-level sensitivity and specificity against a human-coded gold set, especially for the CRREM-fit and no-usable-intensity classes, and clarify how the classifier distinguishes 'operational carbon per m²' from other GHG-intensity disclosur
  2. [Section 4.4, Appendix E] The n=200 stratified sample is drawn from 'the pool of available built-environment report PDFs, most-recent-reporting-year first' (Appendix E). The paper does not characterize the availability mechanism: firms with stronger disclosure may be more likely to publish or to have a PDF in the collection, which would bias the 21.5% rate upward. The frame-weighted estimate of ~19.6% adjusts for the sampling design but not for non-availability. Please provide a comparison of the available-PDF pool with the target population (e.g., by region, firm size, and reporting year), state inclusion/exclusion criteria for the pool, and report sensitivity bounds under plausible availability assumptions. The current limitation paragraph acknowledges that rates are conditional on reporting listed firms, but not the within-listed-firm availability selection.
  3. [Section 4.5] The 45.5% real-estate rate and the US/EU/UK regional contrasts are computed on a 'purpose-assembled supplementary collection' of 1,235 reports from 266 firms, only five of which overlap the 2,246-firm panel. No sampling frame or inclusion criteria for this corpus are provided. If the collection was assembled to include firms with accessible sustainability reports, the rate could overstate disclosure fitness for the real-estate sector as a whole. The paper treats this as a scale-up, but it is a convenience corpus unless its selection process is described and compared with a defined target population. Please state the inclusion rules, the universe from which the 266 firms were drawn, and how the 519 reports were selected within firms.
minor comments (4)
  1. [Section 3.2 / Table 3] The relationship between the built-environment lead stratum (116 firms, expanded to 127) and the n=200 report sample is not fully explicit. Clarify how the 200 reports relate to the firm-level stratum and how the 21.6% crosswalk-unresolved rate interacts with report-level sampling.
  2. [Figure 8] The caption says 'US-listed vs non-US,' but the text reports US/EU/ROW strata. Make the caption consistent with the three-stratum design.
  3. [Section 4.4] The statement that the 4% per-unit/per-home disclosures are 'legitimate as an embodied-carbon metric for homebuilders' is helpful; consider moving this nuance into the abstract or executive summary, since the 'one in five' headline may otherwise be read as discounting all alternative but valid denominators.
  4. [Appendix B] The human-leniency interpretation of the negative LLM–human correlation (r≈-0.24 to -0.29) is plausible but not directly tested; Section 5.5 already acknowledges the competing surface-feature explanation. Consider reporting the planned stratified-extreme adjudication protocol in the main text rather than only in the limitations, as this is the test that would settle the interpretation.

Circularity Check

0 steps flagged

No significant circularity: the headline gap is a measured classification share, not a fitted or self-referential prediction.

full rationale

The central disclosure-fitness claim (21.5% CRREM-fit in n=200, 45.5% in real estate, Sections 4.4 and 4.5) is an empirical classification result from a strict extract-only LLM protocol (Appendix E) plus a deterministic rule-based classifier. No parameter is fitted to the outcome and no equation maps the inputs into the headline by construction; the main uncertainty is measurement error, which the paper explicitly acknowledges in Section 5.5: 'a human audit of the extracted numbers is a stated extension.' That is a validity/accuracy limitation, not circularity. The credibility spine is reused from the co-authored SSSR corpus (Section 3.1), which is self-referential as a data source, but the paper does not rely on the citation for its validity: it re-establishes reliability across model families and anchors validity on three external criteria (LSEG, RepRisk, SBTi), and reports the one LLM-versus-human divergence in Appendix B. The same-vendor limitation for reliability validators is stated in Section 4.1 ('a vendor-level artefact cannot be excluded') and is again a validity caveat rather than a circular derivation. Section 3.5 even rejects a fine-tuned comparator because 'distilling from B's own scores would be circular,' showing explicit awareness of the distinction between reuse and circular argument. The CRREM 39% above-pathway screen uses external CRREM pathways and disclosed intensities, not a fitted target. No load-bearing step in the derivation chain reduces to its own input by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 7 axioms · 0 invented entities

The paper does not introduce a new physical entity or force. Its central claim rests on domain assumptions about the regulatory benchmark (CRREM/EPBD), the per-m² operational-carbon criterion, the accuracy of LLM extraction, the representativeness of the sample, and the interpretation of the negative human-correlation finding. No fitted free parameters drive the headline gap; the listed thresholds are hand-chosen analysis choices that affect secondary or robustness results.

free parameters (3)
  • Physical-plausibility cap for extracted GHG intensity = 2000 kgCO2e·m−2·yr−1
    Hand-chosen threshold in the disclosure-fitness classifier (Appendix E); values above it are flagged as extraction errors, affecting the 4% 'implausible/extraction error' class.
  • CRREM constructibility cutoff (reporting year ≥ 2015) = 2015
    Analyst-chosen cutoff in the CRREM sub-study (Appendix D); 21 disclosing reports before 2015 are excluded, reducing the constructible set from 236 to 215. Does not affect the headline 21.5% gap.
  • Pre-2020 reporting-year comparison to pathway start = 2020 pathway starting value
    Chosen convention in Appendix D; reports dated 2015–2019 are compared to the 2020 pathway start, the least stringent point of the declining curve. The authors state this is conservative.
axioms (7)
  • domain assumption CRREM 1.5°C pathways are the appropriate science-based benchmark for building stranding assessment.
    Invoked in Sections 4.5 and Appendix D when mapping disclosures to CRREM pathways; the paper's stranding-readiness profile depends on this benchmark being the right one.
  • domain assumption Operational carbon intensity per m²/yr is the fitness criterion presupposed by the cited stranding regulations.
    Stated in Sections 1 and 3.3; the entire disclosure-fitness classification is built around this operational per-m² criterion, while embodied-carbon metrics are explicitly set aside.
  • domain assumption The LLM strict-extract protocol returns null exactly when a metric is not explicitly disclosed; systematic under-extraction is absent.
    Core to the 21.5% estimate. Cross-extractor κ bounds model variance, but the paper states in Section 5.5 that a human audit of extracted numbers is a stated extension.
  • domain assumption Unresolved crosswalk firms contain built-environment firms at the same rate as resolved firms.
    Section 3.4: 21.6% of the panel is crosswalk-unresolved; the built-environment rate among resolved firms is applied to the unresolved share. The authors state this as an assumption.
  • domain assumption LSEG ESG ratings, RepRisk incident counts, and SBTi validation status are valid external criteria for disclosure credibility.
    Section 4.2 and Appendix C: these are the three convergent criteria used to support the credibility spine's construct validity.
  • ad hoc to paper The human-coder compression to the upper third of the scale reflects human leniency rather than LLM over-discrimination on surface features.
    Appendix B and Section 5.5: the paper interprets the negative LLM–human correlation as human leniency, but explicitly states it cannot exclude the competing hypothesis. This interpretation supports anchoring validity on external criteria rather than on the human gold set.
  • domain assumption The sampled available report PDFs are representative of the target population of built-environment disclosures.
    Section 4.4 and Appendix E: reports are drawn from the pool of available PDFs, most-recent-reporting-year first. The paper does not quantify PDF-availability bias, which could skew the gap statistic.

pith-pipeline@v1.3.0-alltime-deepseek · 25307 in / 11748 out tokens · 123309 ms · 2026-08-01T06:04:22.655075+00:00 · methodology

0 comments
read the original abstract

Buildings account for roughly 34% of global final energy use and 37% of energy- and process-related CO$_2$ emissions. Stranding regulation now being enacted (New York City Local Law 97, the EU Energy Performance of Buildings Directive recast) presupposes that a building portfolio's carbon intensity can be measured per square metre and compared against a science-based pathway. Whether corporate disclosure is actually fit for that comparison has not, to our knowledge, been measured at scale. We introduce BeDA (the Built-environment Decarbonisation-disclosure Auditor), a multimodal large-language-model instrument, and apply it to a global firm panel (2,246 firms, 2003-2023). Its standards-compliance score is reliable across models and model families and convergent with three independent external criteria. Most disclosure is unfit: only about one built-environment firm-report in five discloses operational carbon intensity per $m^2$ (21.5% in a region-stratified sample of 200 firm-reports, Wilson 95% CI [16.4%, 27.7%], inter-extractor $\kappa$=0.95; 45.5% across 519 real-estate firm-reports, $\kappa$=0.97). The rate is roughly twice as high in Europe as in the United States (64-74% versus 37% for listed real estate). Among the 215 real-estate firm-reports for which an intensity can be constructed, 39% already exceed the Carbon Risk Real Estate Monitor (CRREM) 1.5 {\deg}C pathway's intensity limit. Credibility does not predict stranding readiness once portfolio size is controlled; this is a screening tool, not a forecast. The main obstacle to enforceable building-stranding regulation is therefore a measurable, jurisdiction-specific reporting gap, one that a targeted disclosure mandate can close and that BeDA can monitor.

Figures

Figures reproduced from arXiv: 2607.22006 by Anchen Sun, Jingyi Xu, Minghui Cheng.

Figure 1
Figure 1. Figure 1: The BeDA instrument. P1 produces the credibility scores; P2–P3 establish that those scores are [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The SSSR corpus. (a) Temporal coverage, full panel vs the built-environment subset, 2003–2023. (b) Built [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Credibility-score distribution: the built-environment stratum vs the full panel. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cross-model agreement of the credibility spine between [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cross-family concordance vs the primary gemini-3.1-pro-preview. The standards-compliance spine stays high across the Gemma family and Flash-tier Gemini variants; topic-breadth and the multimodal scores collapse. Short labels identify the specific cross models; full identifiers in [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: SSSR dimension reliability by model. Only the standards-compliance spine retains high concordance across model [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Partial-regression (added-variable) plot of the credibility spine B against LSEG-ESG after partialling out disclosure [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Disclosure-fitness composition, US-listed versus non-US (operational carbon per m [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 9 canonical work pages

  1. [1]

    rep., UNEP and GlobalABC (2024)

    United Nations Environment Programme, Global status report for buildings and construction 2024: Beyond Foundations — mainstreaming sustainable solutions to cut emissions from the buildings sector, Tech. rep., UNEP and GlobalABC (2024)

    2024

  2. [2]

    Berrill, E

    P. Berrill, E. J. H. Wilson, J. L. Reyna, A. D. Fontanini, E. G. Hertwich, Decarbonization pathways for the residential sector in the United States, Nature Climate Change 12 (2022) 712–718.doi:10.1038/ s41558-022-01429-y

    2022

  3. [3]

    Y. Q. Ang, Z. M. Berzolla, S. Letellier-Duchesne, C. F. Reinhart, Carbon reduction technology path- ways for existing buildings in eight cities, Nature Communications 14 (2023) 1689.doi:10.1038/ s41467-023-37131-6

    2023

  4. [4]

    97 of 2019: Building emissions limits, Local Laws of the City of New York, No

    New York City Council, Local law no. 97 of 2019: Building emissions limits, Local Laws of the City of New York, No. 97 (2019)

    2019

  5. [5]

    European Union, Directive (EU) 2024/1275 of the European Parliament and of the Council of 24 april 2024 on the energy performance of buildings (recast), Official Journal of the European Union, L 2024/1275 (2024)

    2024

  6. [6]

    C. E. Kontokosta, D. Spiegel-Feld, S. Papadopoulos, The impact of mandatory energy audits on building energy use, Nature Energy 5 (2020) 309–316.doi:10.1038/s41560-020-0589-6

    work page doi:10.1038/s41560-020-0589-6 2020

  7. [7]

    J. Xu, M. Cheng, A. Sun, Assessing sustainable practices in architecture: A data-driven analysis of LEED certification adoption and impact in top firms from 2000 to 2023, Frontiers of Architectural Research 14 (2025).doi:10.1016/j.foar.2024.10.002

    work page doi:10.1016/j.foar.2024.10.002 2000

  8. [8]

    Ruiz Manuel, K

    I. Ruiz Manuel, K. Blok, Quantitative evaluation of large corporate climate action initiatives shows mixed progress in their first half-decade, Nature Communications 14 (2023) 3487.doi:10.1038/ s41467-023-38989-2

    2023

  9. [9]

    Webersinke, M

    N. Webersinke, M. Kraus, J. A. Bingler, M. Leippold, ClimateBert: A pretrained language model for climate-related text (2022).arXiv:2110.12010

    Pith/arXiv arXiv 2022

  10. [10]

    J. A. Bingler, M. Kraus, M. Leippold, N. Webersinke, Cheap talk and cherry-picking: What Climate- BERT has to say on corporate climate risk disclosures, Finance Research Letters 47 (2022) 102776. doi:10.1016/j.frl.2022.102776

    arXiv 2022

  11. [11]

    Loughran, B

    T. Loughran, B. McDonald, When is a liability not a liability? textual analysis, dictionaries, and 10-Ks, The Journal of Finance 66 (1) (2011) 35–65.doi:10.1111/j.1540-6261.2010.01625.x

    arXiv 2011

  12. [12]

    rep., CRREM (2023)

    Carbon Risk Real Estate Monitor, CRREM risk assessment reference guide V2, Tech. rep., CRREM (2023)

    2023

  13. [13]

    rep., PCAF/CRREM/GRESB, version 1.0 (2023)

    PCAF, CRREM and GRESB, Accounting and reporting of GHG emissions from real estate operations: Technical guidance for the financial industry, Tech. rep., PCAF/CRREM/GRESB, version 1.0 (2023)

    2023

  14. [14]

    L. Zhou, W. Schellaert, F. Martínez-Plumed, Y. Moros-Daval, C. Ferri, J. Hernández-Orallo, Larger and more instructable language models become less reliable, Nature 634 (2024) 61–68.doi:10.1038/ s41586-024-07930-y

    2024

  15. [15]

    Pataranutaporn, N

    P. Pataranutaporn, N. Powdthavee, C. Archiwaranguprok, P. Maes, Simulating human well-being with large language models: systematic validation and misestimation across 64,000 individuals from 64 countries, Proceedings of the National Academy of Sciences 122 (2025) e2519394122.doi:10.1073/ pnas.2519394122. 28

    2025

  16. [16]

    Perera, D

    K. Perera, D. Kuruppuarachchi, S. Kumarasinghe, M. T. Suleman, The impact of carbon disclosure and carbon emissions intensity on firms’ idiosyncratic volatility, Energy Economics 128 (2023) 107053. doi:10.1016/j.eneco.2023.107053

    arXiv 2023

  17. [17]

    L.-Y. He, L. Wang, Can artificial intelligence curb greenwashing? firm-level evidence based on large language model, Energy Economics 152 (2025) 108954.doi:10.1016/j.eneco.2025.108954

    arXiv 2025

  18. [18]

    Zhang, The pathway to curb greenwashing in sustainable growth: The role of artificial intelligence, Energy Economics 133 (2024) 107562.doi:10.1016/j.eneco.2024.107562

    D. Zhang, The pathway to curb greenwashing in sustainable growth: The role of artificial intelligence, Energy Economics 133 (2024) 107562.doi:10.1016/j.eneco.2024.107562

    arXiv 2024

  19. [19]

    Zhang, Green financial system regulation shock and greenwashing behaviors: Evidence from Chinese firms, Energy Economics 111 (2022) 106064.doi:10.1016/j.eneco.2022.106064

    D. Zhang, Green financial system regulation shock and greenwashing behaviors: Evidence from Chinese firms, Energy Economics 111 (2022) 106064.doi:10.1016/j.eneco.2022.106064

    arXiv 2022

  20. [20]

    X. Yuan, Z. Li, J. Xu, L. Shang, ESG disclosure and corporate financial irregularities – evidence from Chinese listed firms, Journal of Cleaner Production 332 (2022) 129992.doi:10.1016/j.jclepro.2021. 129992

    work page doi:10.1016/j.jclepro.2021 2022

  21. [21]

    F. Berg, J. F. Kölbel, R. Rigobon, Aggregate confusion: The divergence of ESG ratings, Review of Finance 26 (6) (2022) 1315–1344.doi:10.1093/rof/rfac033

    work page doi:10.1093/rof/rfac033 2022

  22. [22]

    Zhang, J

    H. Zhang, J. Lai, Z. Guo, When greenwashing meets ESG: Exploring the role of third-party ESG ratings in corporate greenwashing behavior, Journal of Business Research 200 (2025) 115611.doi: 10.1016/j.jbusres.2025.115611

    arXiv 2025

  23. [23]

    Serapio-García, M

    G. Serapio-García, M. Safdari, C. Crepy, L. Sun, et al., A psychometric framework for evaluating and shaping personality traits in large language models, Nature Machine Intelligence 7 (2025) 1954–1968. doi:10.1038/s42256-025-01115-6

    doi:10.1038/s42256-025-01115-6 2025

  24. [24]

    Demszky, D

    D. Demszky, D. Yang, D. S. Yeager, et al., Using large language models in psychology, Nature Reviews Psychology 2 (2023) 688–701.doi:10.1038/s44159-023-00241-5

    doi:10.1038/s44159-023-00241-5 2023

  25. [25]

    Gilardi, M

    F. Gilardi, M. Alizadeh, M. Kubli, ChatGPT outperforms crowd workers for text-annotation tasks, Proceedings of the National Academy of Sciences 120 (30) (2023) e2305016120.doi:10.1073/pnas. 2305016120

    doi:10.1073/pnas 2023

  26. [26]

    Ziems, W

    C. Ziems, W. Held, O. Shaikh, J. Chen, Z. Zhang, D. Yang, Can large language models transform computational social science?, Computational Linguistics 50 (1) (2024) 237–291.doi:10.1162/coli_ a_00502

    doi:10.1162/coli_ 2024

  27. [27]

    Y. Xie, L. Liang, S. Li, Y. Lu, Z. Xiao, M. Shi, J. Huang, M. Wang, Y. Xie, Evaluating the statistical realism of LLM-generated social science data, Proceedings of the National Academy of Sciences 123 (2026) e2538145123.doi:10.1073/pnas.2538145123

    doi:10.1073/pnas.2538145123 2026

  28. [28]

    Schramowski, C

    P. Schramowski, C. Turan, N. Andersen, C. A. Rothkopf, K. Kersting, Large pre-trained language models contain human-like biases of what is right and wrong to do, Nature Machine Intelligence 4 (2022) 258–268.doi:10.1038/s42256-022-00458-8

    doi:10.1038/s42256-022-00458-8 2022

  29. [29]

    Gemini Team, Google, Gemini: A family of highly capable multimodal models (2023).arXiv:2312. 11805

    2023

  30. [30]

    Gemma Team, Google DeepMind, Gemma 3 technical report (2025).arXiv:2503.19786

    Pith/arXiv arXiv 2025

  31. [31]

    Carbon Risk Real Estate Monitor, CRREM global pathways V2.04, Dataset, CRREM, released 28 August 2025 (2025). 29

    2025

  32. [32]

    Semieniuk, P

    G. Semieniuk, P. B. Holden, J.-F. Mercure, P. Salas, H. Pollitt, K. Jobson, P. Vercoulen, U. Chew- preecha, N. R. Edwards, J. E. Viñuales, Stranded fossil-fuel assets translate to major losses for investors in advanced economies, Nature Climate Change 12 (2022) 532–538.doi:10.1038/ s41558-022-01356-y

    2022

  33. [33]

    A. von Dulong, Concentration of asset owners exposed to power sector stranded assets may trigger climatepolicyresistance, NatureCommunications14(2023)6442.doi:10.1038/s41467-023-42031-w

    work page doi:10.1038/s41467-023-42031-w 2023

  34. [34]

    Wetzer, R

    T. Wetzer, R. Stuart-Smith, A. Dibley, Climate risk assessments must engage with the law, Science 383 (2024) 152–154.doi:10.1126/science.adj0598

    work page doi:10.1126/science.adj0598 2024

  35. [35]

    Zhang, F

    G. Zhang, F. Chang, J. Liu, Carbon emission prediction of 275 cities in China considering artificial intel- ligence effects and feature interaction: A heterogeneous deep learning modeling framework, Sustainable Cities and Society 114 (2024) 105776.doi:10.1016/j.scs.2024.105776

    arXiv 2024

  36. [36]

    Zhang, S

    G. Zhang, S. Ma, M. Zheng, C. Li, F. Chang, F. Zhang, Impact of digitization and artificial intelligence on carbon emissions considering variable interaction and heterogeneity: An interpretable deep learning modeling framework, Sustainable Cities and Society 125 (2025) 106333.doi:10.1016/j.scs.2025. 106333

    work page doi:10.1016/j.scs.2025 2025

  37. [37]

    S. Xia, H. Zhu, N. Zhang, Y. Li, C. Zhou, A dual-feature channel deep network with adaptive variable weight reconstruction for urban water demand prediction, Sustainable Cities and Society 119 (2025) 106118.doi:10.1016/j.scs.2024.106118

    arXiv 2025

  38. [38]

    H.Nan, J.Lin, Z.Xia, S.Zeng, Governanceofsmartservicesocialresponsibility: Asystematicliterature review, IEEE Transactions on Engineering Management 72 (2025) 3365–3382.doi:10.1109/TEM.2025. 3597053

    work page doi:10.1109/tem.2025 2025

  39. [39]

    Z. Xia, S. Zeng, A. Sun, H. Sun, X. Cai, Global evolution of social responsibility in smart-service industries: Insights from a cross-sector hybrid large language models approach, Corporate Social Re- sponsibility and Environmental Management 33 (2026) 1815–1831.doi:10.1002/csr.70257

    work page doi:10.1002/csr.70257 2026

  40. [40]

    L. F. Cabeza, Q. Bai, P. Bertoldi, J. M. Kihila, O. Lucon, É. Mata, S. Mirasgedis, A. Novikova, Y. Saheb, et al., Buildings, in: Climate Change 2022: Mitigation of Climate Change. Contribution of Working Group III to the Sixth Assessment Report of the IPCC, Cambridge University Press, 2022, Ch. 9

    2022

  41. [41]

    International Energy Agency, Buildings, Energy System Tracking Report, IEA, Paris (2024)

    2024

  42. [42]

    Carlini, F

    N. Carlini, F. Tramèr, E. Wallace, M. Jagielski, A. Herbert-Voss, K. Lee, A. Roberts, T. Brown, D. Song, Ú. Erlingsson, A. Oprea, C. Raffel, Extracting training data from large language models, in: 30th USENIX Security Symposium (USENIX Security 21), 2021, pp. 2633–2650

    2021

  43. [43]

    N.Carlini, D.Ippolito, M.Jagielski, K.Lee, F.Tramèr, C.Zhang, Quantifyingmemorizationacrossneu- ral language models, in: The Eleventh International Conference on Learning Representations (ICLR), 2023

    2023

  44. [44]

    Gemini Team, Google, Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities (2025).arXiv:2507.06261

    Pith/arXiv arXiv 2025

  45. [45]

    Gemma Team, Google DeepMind, Gemma 4 technical report (2026).arXiv:2607.02770

    Pith/arXiv arXiv 2026

  46. [46]

    J. A. Bingler, M. Kraus, M. Leippold, N. Webersinke, How cheap talk in climate disclosures relates to climate initiatives, corporate emissions, and reputation risk, Journal of Banking & Finance 164 (2024) 107191.doi:10.1016/j.jbankfin.2024.107191. 30

    arXiv 2024

  47. [47]

    Cohen, A coefficient of agreement for nominal scales, Educational and Psychological Measurement 20 (1) (1960) 37–46.doi:10.1177/001316446002000104

    J. Cohen, A coefficient of agreement for nominal scales, Educational and Psychological Measurement 20 (1) (1960) 37–46.doi:10.1177/001316446002000104

    doi:10.1177/001316446002000104 1960

  48. [48]

    Cohen, Weighted kappa: Nominal scale agreement with provision for scaled disagreement or partial credit, Psychological Bulletin 70 (4) (1968) 213–220.doi:10.1037/h0026256

    J. Cohen, Weighted kappa: Nominal scale agreement with provision for scaled disagreement or partial credit, Psychological Bulletin 70 (4) (1968) 213–220.doi:10.1037/h0026256

    doi:10.1037/h0026256 1968

  49. [49]

    J. R. Landis, G. G. Koch, The measurement of observer agreement for categorical data, Biometrics 33 (1) (1977) 159–174.doi:10.2307/2529310

    doi:10.2307/2529310 1977

  50. [50]

    Cohen, Statistical Power Analysis for the Behavioral Sciences, 2nd Edition, Lawrence Erlbaum Associates, Hillsdale, NJ, 1988

    J. Cohen, Statistical Power Analysis for the Behavioral Sciences, 2nd Edition, Lawrence Erlbaum Associates, Hillsdale, NJ, 1988

    1988

  51. [51]

    Bjørn, S

    A. Bjørn, S. M. Lloyd, M. Brander, H. D. Matthews, Renewable energy certificates threaten the in- tegrity of corporate science-based targets, Nature Climate Change 12 (2022) 539–546.doi:10.1038/ s41558-022-01379-5

    2022

  52. [52]

    rep., EPRA, Brussels, third version (2017)

    European Public Real Estate Association, Sustainability best practices recommendations guidelines, Tech. rep., EPRA, Brussels, third version (2017)

    2017

  53. [53]

    European Commission, Commission delegated regulation (EU) 2023/2772 of 31 july 2023 supplementing directive 2013/34/EU as regards sustainability reporting standards, Official Journal of the European Union, L 2023/2772 (2023)

    2023

  54. [54]

    Zhong, M

    X. Zhong, M. Hu, S. Deetman, B. Steubing, H. X. Lin, G. A. Hernandez, C. Harpprecht, C. Zhang, A. Tukker, P. Behrens, Global greenhouse gas emissions from residential and commercial building materials and mitigation strategies to 2060, Nature Communications 12 (2021) 6126.doi:10.1038/ s41467-021-26212-z. 31

    2060