AIMSCheck: Leveraging LLMs for AI-Assisted Review of Modern Slavery Statements Across Jurisdictions

Adriana Eufrosina Bora; Akshatha Arodi; Arsene Fansi Tchango; Duoyi Zhang; Jordan Bannister; Kerrie Mengersen; Md Abul Bashar; Mirko Bronzi; Richi Nayak

REVIEW 3 major objections 5 minor 14 references

AIMSCheck: Leveraging LLMs for AI-Assisted Review of Modern Slavery Statements Across Jurisdictions

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A sentence classifier fine-tuned on Australian corporate statements keeps most of its accuracy on UK and Canadian modern-slavery reports, and the authors package this into a three-level compliance review framework called AIMSCheck.

desk verdict The real contribution is the two new benchmark datasets and the cross-jurisdiction evaluation; the generalization claim is plausible but slightly overstated because the criteria mapping has a known legal conflation for remediation. read the letter →

arxiv 2506.01671 v1 pith:EJHZ4V7O submitted 2025-06-02 cs.CY cs.CL

Adriana Eufrosina Bora , Akshatha Arodi , Duoyi Zhang , Jordan Bannister , Mirko Bronzi , Arsene Fansi Tchango , Md Abul Bashar , Richi Nayak

show 1 more author

Kerrie Mengersen

This is my paper

classification cs.CYcs.CL

keywords modernslaverystatementscross-jurisdictionalgeneralizationsentenceclassificationlargelanguagemodelscompliancemonitoringevidencetrackingexplainableAIannotatedbenchmarks

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a sentence-classification model trained only on Australian corporate statements can be applied to UK and Canadian modern-slavery statements with a modest drop in accuracy, and that this makes cross-jurisdictional compliance checking practical. To test the claim, the authors built two new expert-annotated datasets, AIMS.uk and AIMS.ca, aligned with the existing Australian AIMS.au data around nine reporting criteria common to all three laws, and packaged everything in AIMSCheck, a framework that gives reviewers sentence-level relevance flags, token-level justifications, and evidence status (already done, promised, or denied). The stakes are concrete: roughly 21,000 statements are filed annually across the three registries, and manual review of even a few hundred statements is slow, so a single system that assists human reviewers across jurisdictions would change how compliance monitoring can be scaled.

What carries the argument

The load-bearing object is the nine-common-criteria mapping, an expert-curated translation of the Australian MSA's mandatory criteria, the UK Act's reporting suggestions, and the Canadian Act's obligations into one label set: Approval, Signature, Structure, Operations, Supply Chains, Risk Description, Risk Mitigation, Remediation, and Effectiveness. The sentence-level classifier makes nine binary relevance decisions per sentence, using a 100-word context window around the target sentence, and the framework layers SHAP token attributions and two evidence trackers on top (a tense classifier for future commitments and a zero-shot BART-MNLI model for denials). The mapping is what makes the three datasets commensurable, and the context window plus the trackers are what turn raw classifications into output a human reviewer can audit.

What would settle it

Have an independent legal expert who was not involved in the annotation guidelines re-annotate a random sample of AIMS.uk and AIMS.ca statements directly from the statutes and compare with the released labels; if Cohen's kappa falls below about 0.6, the nine-criteria mapping is not a stable ground truth and the cross-jurisdiction scores would rest on one expert's interpretation. A second check: train on AIMS.uk or AIMS.ca and test on AIMS.au; if transfer works only in the Australian-to-others direction, the reported generalization is an artifact of which dataset was labeled, not evidence of legal equivalence.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that cross-jurisdictional generalization works for sentence-level compliance classification: a Llama3.2 3B model fine-tuned on AIMS.au reaches 0.738 average F1 on Australian test sentences, 0.719 on the new Canadian dataset, and 0.686 on the new UK dataset, while a fine-tuned BERT trails slightly and zero-shot and few-shot GPT-4o and quantized DeepSeek-R1 land lower on every dataset. The authors also find that fine-tuned models consistently beat the prompt-based models, that calibration is good enough for reviewers to use probabilities as confidence signals, and that the evidence-status trackers can find future commitments and explicit denials in the statements. From this they conclude that, with the right compliance mapping, domain adaptation across modern-slavery jurisdictions is achievable and that the released benchmark datasets and framework support it.

Load-bearing premise

The load-bearing premise is that the nine common criteria, curated by domain experts, capture legally equivalent reporting requirements across the Australian, UK, and Canadian statutes; if that mapping gives way, the cross-jurisdictional F1 scores would measure an artificial common-denominator task rather than generalization of real compliance checking.

Editorial extensions

If this is right

A compliance tool can be trained once on AIMS.au and applied in the UK and Canada without new annotation, since fine-tuned models keep most of their accuracy across the three jurisdictions.
The release of AIMS.uk and AIMS.ca gives researchers and supervisory agencies public benchmarks for compliance sentence classification outside Australia, where only the Australian dataset existed before.
Evidence-status tracking separates implemented measures from future promises and explicit denials, so reviewers can distinguish companies that have acted from companies that only say they will act.
The per-criterion F1 pattern (high on approval and signature, low on risk mitigation and effectiveness) identifies where human reviewers should focus their attention when using the framework.
Compliance-trend analysis of the UK dataset suggests that industry and infrastructure firms disclose risks less transparently than public-facing sectors, a pattern the pipeline can be used to probe at scale.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The appendix's mapping of the nine criteria onto French, German, and Norwegian due-diligence laws is only a feasibility sketch; a direct test would be to annotate a few statements from one of those regimes and see whether an Australian-trained model survives the shift from 'modern slavery' to broader human-rights language.
The consistent edge of a fine-tuned open model over much larger closed models implies that domain-labeled data, not model scale, is the binding constraint for this task; agencies with limited compute would likely get more value from fine-tuning than from API prompting, a cost conclusion the paper does not draw.
The paper's Jensen-Shannon vocabulary-divergence measure could be reused as a domain-shift alarm: monitors could compute divergence on each new statement batch and re-annotate or retrain only when drift from the training distribution grows.
A longitudinal design the paper gestures at but does not run would chain each company's statements across publication years, turning evidence-status labels into a per-company progress scoreboard of the kind regulators would need for enforcement and for measuring whether the Acts improve reporting over time.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Worth a look for anyone in legal NLP or compliance monitoring. The genuinely new pieces are the AIMS.uk and AIMS.ca datasets, each with 50 manually annotated statements from national registries, and the cross-jurisdictional benchmark built on them. The AIMSCheck framework itself is an assembly of standard parts — sentence-level fine-tuning, zero/few-shot prompting, SHAP, and a rule-based evidence-status module — but the evidence-tracking distinction between implemented actions, future promises, and explicit inaction is a useful idea, and the error analysis is more careful than usual. The fine-tuned models do retain most of their F1 when trained on AIMS.au and tested on UK and Canadian data, and the per-criterion scores in Table 3 show where that holds and where it does not. The paper is transparent about its limitations and the materials are promised for release, which counts for something. The soft spots are real but mostly not fatal. There are no confidence intervals or significance tests, so the gap between AIMS.au (0.738) and AIMS.uk (0.686) could be noise; the new datasets were annotated by a single expert, with agreement measured on only 140 samples; and the negative-evidence threshold was moved from 0.5 to 0.35 without any appearance of a validation split, which is likely tuning on the test set. The largest concern, and the one the stress-test note correctly identifies, is the nine-criteria mapping in Figure 3. The paper itself says that Canada requires remediation of income loss caused by anti-slavery measures, a duty absent in Australia and the UK, and that the Canadian remediation rate is consequently much higher. For that one criterion, the label does not mean the same legal thing across jurisdictions, so the cross-jurisdictional F1 for C4 is partly measuring an artificial common-denominator task. That does not sink the whole paper — most criteria map cleanly, and the aggregate result is consistent with the per-criterion numbers — but it does mean the phrase “generalize well across jurisdictions” should be softened, or the remediation analysis should be presented separately with a legal caveat. The citation pattern is fine; the reliance on the authors' own AIMS.au work is expected and not circular, since the new datasets are held out. I would take this paper seriously in review, with the request that the authors validate the legal mapping against an external source or restrict the generalization claim to the criteria where the mapping is unambiguous. It is a solid, honest piece of applied work, and the datasets alone justify referee time.

Referee Report

3 major / 5 minor

Summary. This paper introduces AIMS.uk and AIMS.ca, two new sentence-level annotated datasets of 50 modern slavery statements each, and proposes AIMSCheck, a three-level framework for compliance review: sentence-level multi-label classification, token-level SHAP explanations, and evidence-status tracking (future actions vs. negative evidence). The authors fine-tune BERT and Llama3.2 on the Australian AIMS.au dataset and evaluate on held-out Australian, UK, and Canadian test sets, comparing with zero-shot and few-shot GPT-4o and DeepSeek-R1. They report that fine-tuned models outperform prompting-based methods and generalize across jurisdictions, with only modest F1 degradation on UK and Canadian data. The paper also includes per-criterion results, calibration curves, error analysis, vocabulary-similarity analysis, and illustrative compliance-trend analyses.

Significance. If the main claim holds, the contribution is practically valuable: it would demonstrate that a single fine-tuned sentence classifier can transfer across three legal regimes, and the released datasets would support further work in a data-scarce and socially important domain. Strengths include genuinely held-out evaluation (training only on AIMS.au), public release of datasets, prompts, and weights under CC BY 4.0, per-criterion F1 reporting, calibration checks, and detailed qualitative error analyses. The main uncertainty is whether the nine-criteria legal mapping used for cross-jurisdictional evaluation is substantively valid; the paper itself notes that the Canadian Act contains a remediation obligation absent in Australia and the UK, which weakens the assumption that the merged C4 criterion is equivalent across all three jurisdictions.

major comments (3)

[§2, Figure 3 / Table 4] The central claim that models trained on AIMS.au "generalize well" to the UK and Canada presupposes that the nine common criteria denote legally equivalent reporting obligations. The manuscript itself reports that the Canadian Act requires remediation of income loss caused by anti-slavery measures, a duty absent in Australia and the UK, and that a 1-to-1 mapping required merging granular criteria into broader categories. For C4 (remediation), AIMS.ca labels therefore include sentences addressing a different statutory obligation than those in AIMS.au and AIMS.uk. A model trained on Australian remediation is scored against Canadian sentences that are relevant to a distinct duty, so a high or low F1 on that criterion does not directly measure generalization of compliance checking. The only validation offered is agreement between two experts who were both involved in developing the annotation guidelines (Cohen's Kappa 0.776), which does not independently validate the mapping against external legal authority. Please add an external legal validation of the mapping, report per-criterion results that separate shared-core from jurisdiction-specific content, or otherwise demonstrate that the merged categories are substantively equivalent.
[§2 Annotation Validation and §7 Limitations] The two new datasets were annotated by a single domain expert, and the reported agreement (Cohen's Kappa 0.776, Jaccard 0.813 on 140 samples) involves a second expert who was "deeply involved in developing the annotation guidelines." This is a useful check but not an independent validation of label quality, and the agreement is not broken down by criterion. Because the cross-jurisdictional F1 scores in Tables 2 and 3 depend on the reliability of AIMS.uk and AIMS.ca labels, the paper should provide per-criterion agreement, confidence intervals, and ideally an additional annotation pass by an expert not involved in designing the guidelines. At minimum, the single-annotator limitation should be reflected in the strength of the generalization claim.
[§4, Table 2] The "generalize well" conclusion is based on F1 differences that are small in absolute terms (e.g., Llama3.2-100: 0.738 on AIMS.au, 0.719 on AIMS.ca, 0.686 on AIMS.uk), with no confidence intervals, significance tests, or statement-level variance. With only 50 statements per jurisdiction and many criteria having low positive-class rates, these differences may lie within sampling noise. Report bootstrap confidence intervals, per-statement standard deviations, or paired statistical tests across the nine criteria, and state a pre-specified threshold for what degree of degradation is compatible with "generalize well." Without this, the headline claim is not quantitatively supported.

minor comments (5)

[§4, Table 2] The DeepSeek-R1 row reports values in the AIMS.ca and AIMS.uk columns, but the note below the table says DeepSeek-R1 is evaluated on AIMS.au; please clarify which datasets were actually scored.
[§3.3 and §E.3] The negative-evidence detector's threshold is lowered from 0.5 to 0.35 with no sensitivity analysis; please report performance across thresholds or justify the chosen value.
[Appendix C] The materials release section mentions GitHub, Hugging Face, and Figshare but gives no persistent identifiers; include URLs or DOIs for the datasets, prompts, and code.
[Throughout] There are several typos and grammatical slips, including "mititation" (Section 3), "where as" (Section 2), "accessment" (Appendix A), "overal" (Appendix E.1), "Lllama" (Appendix H.2), and "Sentances" (Appendix H.2 title).
[Figure 1 and Figure 3] The figures use similar but not identical criterion labels (e.g., "Approval" and "Approval from the board of directors"); consider harmonizing the labels and adding a caption note clarifying which criteria are common versus jurisdiction-specific, especially for C4 remediation.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: held-out cross-jurisdictional evaluation supports the transfer claim; expert mapping and self-citations are assumptions and inputs, not reductions of the claimed result.

full rationale

The paper's central claim is that models fine-tuned on the Australian AIMS.au dataset generalize to the newly introduced AIMS.uk and AIMS.ca datasets. This claim is supported by genuinely held-out evaluation: the UK and Canadian statements are newly collected from government registries, annotated separately, and no sentence-level classifier parameter is fitted to the UK or Canadian labels. The negative-evidence threshold adjustment from 0.5 to 0.35 (Appendix E.3) concerns the auxiliary evidence-status component, is disclosed, and does not affect the headline sentence-level generalization result. Self-citations to Bora et al. (2025) provide the prior AIMS.au dataset, annotation guidelines, and prompt templates; these are inputs to the evaluation design rather than the result being established, and the new datasets are independently sourced. The expert-curated nine-criteria jurisdictional mapping (Section 2, Figures 1 and 3) is a substantive validity assumption, particularly for the Canadian remediation criterion where Section 2 notes a distinct statutory income-loss remediation requirement; however, this is a correctness and external-validity risk, not circularity, because the mapping is not derived from model outputs and is not fitted to the test labels. No load-bearing step reduces, by the paper's own equations or by self-citation, to the result it claims to predict.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The framework rests on the validity of the cross-jurisdictional criteria mapping, on the transferability of AIMS.au annotation guidelines, and on sentence-level labels being a sufficient proxy for compliance. The only numeric parameter adjusted after observing model behavior is the negative-evidence threshold (0.35). No new entities are postulated.

free parameters (2)

BART-MNLI negative evidence threshold = 0.35
Adjusted from the default 0.5 to improve sensitivity in the zero-shot negative evidence detector (Appendix E.3). No separate validation split is described, so the threshold may have been tuned on test data.
Context window length = 100 words
The paper reports experiments with context lengths 0, 100, 200, and 500 and selects 100 as the best tradeoff (Section 3.1). This is a hyperparameter chosen by the authors, and all main results use it.

assumptions (3)

domain assumption The nine-criteria cross-jurisdictional mapping accurately reflects equivalent reporting obligations under the Australian, UK, and Canadian Modern Slavery Acts.
The mapping is curated by the authors' domain experts (Section 2, Figure 3). If the criteria are not truly equivalent across jurisdictions, the cross-jurisdiction evaluation is not measuring the same construct.
domain assumption The AIMS.au annotation guidelines apply without modification to UK and Canadian statements (Section 2).
Statements are annotated by a single expert following prior guidelines from Bora et al. (2025); legal differences may require different annotation decisions.
domain assumption Sentence-level binary relevance labels are a sufficient representation of statutory compliance for the nine selected criteria (Section 3.1).
Compliance often depends on document-level or entity-level facts; reducing it to sentence-level labels may miss cross-sentence dependencies.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AIMSCheck: Leveraging LLMs for AI-Assisted Review of Modern Slavery Statements Across Jurisdictions." pith.science (2026). https://pith.science/paper/EJHZ4V7O

@misc{pith2026250601671,
  author       = {Pith},
  title        = {Pith review of: AIMSCheck: Leveraging LLMs for AI-Assisted Review of Modern Slavery Statements Across Jurisdictions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EJHZ4V7O}},
  note         = {Machine review of arXiv:2506.01671}
}

read the original abstract

Modern Slavery Acts mandate that corporations disclose their efforts to combat modern slavery, aiming to enhance transparency and strengthen practices for its eradication. However, verifying these statements remains challenging due to their complex, diversified language and the sheer number of statements that must be reviewed. The development of NLP tools to assist in this task is also difficult due to a scarcity of annotated data. Furthermore, as modern slavery transparency legislation has been introduced in several countries, the generalizability of such tools across legal jurisdictions must be studied. To address these challenges, we work with domain experts to make two key contributions. First, we present AIMS.uk and AIMS.ca, newly annotated datasets from the UK and Canada to enable cross-jurisdictional evaluation. Second, we introduce AIMSCheck, an end-to-end framework for compliance validation. AIMSCheck decomposes the compliance assessment task into three levels, enhancing interpretability and practical applicability. Our experiments show that models trained on an Australian dataset generalize well across UK and Canadian jurisdictions, demonstrating the potential for broader application in compliance monitoring. We release the benchmark datasets and AIMSCheck to the public to advance AI-adoption in compliance assessment and drive further research in this field.

Figures

Figures reproduced from arXiv: 2506.01671 by the authors.

**Figure 2.** The AIMSCheck pipeline is designed to process modern slavery statements mandated by the Modern [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Mapping the AU MSA mandatory criteria, UK MSA reporting suggestions, and Canadian Act reporting [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (15 more)

**Figure 4.** Figure 4: The distributions of word counts per sentence, [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 6.** Figure 6: The SHAP plot illustrates the influence of tokens on the model’s output. Here, we show the text (top) and [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: This example tracks the evidence status of a statement. The three underlined sentences are classified under [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 8.** Figure 8: Proportion of Compliance for C3 (Risk De [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

**Figure 9.** Figure 9: Jensen-Shannon (JS) divergences between AIMS.au (training) and testing vocabularies of all other datasets: Analysing overall, negative, and positive labelconditioned distributions across nine labels. Lower values indicate greater similarity. Vocabulary Similarity To …

**Figure 10.** Figure 10: Example UK statement with nine criteria and evidence status annotations. [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

**Figure 11.** Figure 11: The processing workflow of a real statement example using the AIMSCheck framework. At statement [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

**Figure 12.** Figure 12: Zero-shot prompt template used in model experiments under the [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]

**Figure 13.** Figure 13: Chain-of-Thoughts zero-shot prompt template used for zero-shot model experiments under the [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]

**Figure 14.** Figure 14: Chain-of-Thoughts few-shot prompt template used for zero-shot model experiments under the [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]

**Figure 15.** Figure 15: Calibration curves for the Llama3.2 3B model with 100 context words. The dashed black lines correspond [PITH_FULL_IMAGE:figures/full_fig_p021_15.png]

**Figure 16.** Figure 16: Visualization of SHAP values for the target sentence (encompassing the special header tags) with [PITH_FULL_IMAGE:figures/full_fig_p022_16.png]

**Figure 17.** Figure 17: Compliance Proportion by Turnover (£ Millions). for Each Criterion [PITH_FULL_IMAGE:figures/full_fig_p027_17.png]

**Figure 18.** Figure 18: Compliance Proportion by Sector for Each Criterion [PITH_FULL_IMAGE:figures/full_fig_p027_18.png]

**Figure 19.** Figure 19: Compliance Proportion by Publication Year for Each Criterion [PITH_FULL_IMAGE:figures/full_fig_p027_19.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 12 canonical work pages

[1]

We continue due diligence across all our direct and indirect suppliers

Target Sentence: "We continue due diligence across all our direct and indirect suppliers.”

work page
[2]

Question: Is the target sentence relevant? (YES/NO)

work page
[3]

It lacks the detailed information required to describe the supply chain

Reasoning: This sentence mentions due diligence efforts but does not provide specific descriptions of suppliers, their products or services, locations, or other attributes. It lacks the detailed information required to describe the supply chain

work page
[4]

Approval

Final Answer: NO Now classify the following target sentence: The target sentence to classify is the following: ———— TARGET_SENTENCE ———— The same target sentence inside its original block of text: ———— SENTENCE_IN_CONTEXT ———— **Question**: Is the target sentence relevant? (YES/NO) **Answer**: Lets think step-by-step. In order to provide the correct answe...

work page 2017
[8]

Better sentence segmentation techniques and improve list extraction to reduce errors caused by formatting inconsistencies

work page
[9]

Enhanced prompt engineering to specifically address false positives

work page
[10]

Additional dataset augmentation to reduce confusion in overlapping criteria and the false negatives in fine-tuning models. I Compliance Trend in the AIMS.uk Dataset In this appendix, we present a compliance trend analysis based on the prediction outcomes of the Llama3 context-100 model on the 50 statements of AIMS.uk. For each criterion where a statement ...

work page 2025
[11]

The Industry & Infrastructure sector includes industries such as automotive, construction, durable consumer goods, mining, utilities, waste management, transportation, defense, and security services

work page

Show all 14 references

[12]

The Commerce & Services sector covers con- sumer services, including hospitality, tourism, fashion, cosmetics, food and agriculture, fi- nance, professional services, IT, and media
[13]

The Public & Healthcare sector consists of public sector organizations, non-profits, edu- cation, healthcare, and forestry-related indus- tries
[14]

Figure 18 shows the compliance proportion by sector for each criterion

The Other category includes the statements that companies that had "other" in their met- data. Figure 18 shows the compliance proportion by sector for each criterion. 25 Key Observations • Companies within the Public and Healthcare sectors demonstrate the highest levels of com...

2022
[2019]

plan to" and

and Llama3.2 3B (Llama Team, 2024). The models were trained, starting from checkpoints available on HuggingFace repositories (Wolf et al., 2020), using the annotated training data from the AIMS.au database (Bora et al., 2025). For BERT, the full model weights were trained, whi...

2025
[2020]

In Findings of the Association for Com- putational Linguistics: EMNLP 2020 , pages 2898– 2904, Online

LEGAL-BERT: The muppets straight out of 10 law school. In Findings of the Association for Com- putational Linguistics: EMNLP 2020 , pages 2898– 2904, Online. Association for Computational Lin- guistics. R. Chambers and A.Y . Vastardis. 2020. Human rights disclosure and due dil...

2020 arXiv
[2022]

Tech- nical Report

Broken promises: Two years of corporate reporting under australia’s modern slavery act. Tech- nical Report. Federal Republic of Germany. 2021. Act on corporate due diligence obligations in supply chains (lieferket- tensorgfaltspflichtengesetz). Accessed: 2025-05-29. Neel Guha,...

2021 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

We continue due diligence across all our direct and indirect suppliers

Target Sentence: "We continue due diligence across all our direct and indirect suppliers.”

work page

[2] [2]

Question: Is the target sentence relevant? (YES/NO)

work page

[3] [3]

It lacks the detailed information required to describe the supply chain

Reasoning: This sentence mentions due diligence efforts but does not provide specific descriptions of suppliers, their products or services, locations, or other attributes. It lacks the detailed information required to describe the supply chain

work page

[4] [4]

Approval

Final Answer: NO Now classify the following target sentence: The target sentence to classify is the following: ———— TARGET_SENTENCE ———— The same target sentence inside its original block of text: ———— SENTENCE_IN_CONTEXT ———— **Question**: Is the target sentence relevant? (YES/NO) **Answer**: Lets think step-by-step. In order to provide the correct answe...

work page 2017

[5] [8]

Better sentence segmentation techniques and improve list extraction to reduce errors caused by formatting inconsistencies

work page

[6] [9]

Enhanced prompt engineering to specifically address false positives

work page

[7] [10]

Additional dataset augmentation to reduce confusion in overlapping criteria and the false negatives in fine-tuning models. I Compliance Trend in the AIMS.uk Dataset In this appendix, we present a compliance trend analysis based on the prediction outcomes of the Llama3 context-100 model on the 50 statements of AIMS.uk. For each criterion where a statement ...

work page 2025

[8] [11]

The Industry & Infrastructure sector includes industries such as automotive, construction, durable consumer goods, mining, utilities, waste management, transportation, defense, and security services

work page