Towards Next-Generation Intelligent Maintenance: Collaborative Fusion of Large and Small Models

Huiming Ma; Mingqing Guo; Ming Xu; Qiming Huang; Xiao He; Xiaoyi Yuan; Zeyi Liu

REVIEW 3 major objections 5 minor 14 references

Towards Next-Generation Intelligent Maintenance: Collaborative Fusion of Large and Small Models

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 The proposed Sentosa LLM framework claims that fusing domain-specific small models with large language models materially improves industrial maintenance, reducing repair time by 41 percent and raising temperature-control accuracy to 97.8…

desk verdict A clear architecture paper with real deployments whose case studies don't isolate the framework's contribution; the gains likely come from generic LLM augmentation and correction. read the letter →

arxiv 2506.05854 v1 pith:NZRLYKBQ submitted 2025-06-06 eess.SY cs.SY

Xiaoyi Yuan , Qiming Huang , Mingqing Guo , Huiming Ma , Ming Xu , Zeyi Liu , Xiao He This is my paper · ORCID

classification eess.SYcs.SY

keywords intelligentmaintenancelargelanguagemodelssmallmodelfusiontemperaturepredictionfaultdetectionenergystorageknowledgegraph

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

Sentosa LLM is a five-layer framework for industrial maintenance that couples domain-specific small models, such as LSTM forecasters, with large language models. The paper's central claim is that this collaboration outperforms either kind of model alone: temperature-prediction error falls, air-conditioning control keeps the room in the target band more often, and the time to repair an energy-storage fault drops by 41 percent. A sympathetic reader should care because industrial maintenance has lagged in adopting LLMs, and the paper offers a concrete architecture that routes heterogeneous sensor data, logs, and manuals through a shared knowledge layer so the LLM can reason while small models keep number crunching fast and accurate.

What carries the argument

The carrier of the argument is the five-layer Sentosa LLM architecture: Device, Data, Model, Application, and Decision and Interaction layers. At its core sit the Domain-specific SMs module and the LLM Collaboration module, connected by an intermediate-representation layer built from an Industrial Knowledge Graph, Semantic Layer, and Vector Database, plus prompt engineering, retrieval-augmented generation, a Workflow Engine, and standardized tool-invocation interfaces such as APIs, function calling, and the Model Context Protocol. The small models handle precise, low-latency computation, such as forecasting temperatures or scoring battery faults, while the LLM plans tasks, retrieves knowledge, calls those small models as tools, and synthesizes corrections and maintenance instructions. The paper describes the resulting interpretability as a chain built from transparent model design, traceable interaction processes, deep integration of domain knowledge, and multi-dimensional evaluation and verification.

What would settle it

Run the two case studies again with an off-the-shelf LLM producing the same augmented data and corrected outputs, without the industrial knowledge graph, semantic layer, workflow engine, or tool-invocation interfaces, and compare mean absolute error, temperature-control accuracy, and mean time to repair; if the gains persist, the framework's distinctive architecture is not the carrier of the results.

Watch

Extended reading notes

Core claim

The central discovery is that the largest measured gains come from using the LLM in two complementary roles around a fixed small model: generating augmented training data from historical records and domain text, then correcting the small model's predictions with contextual signals from external weather and internal system states. In the equipment-room case, LLM data augmentation lowers the LSTM's mean absolute error from 0.453 to 0.383, and LLM correction lowers it further to 0.300, while temperature-control accuracy rises from 86.5 percent, under the default strategy, to 97.8 percent and weekly energy consumption falls from 15,837 kWh to 12,525 kWh. In the energy-storage case, the small model issues an initial micro-short-circuit alert on day 9, the LLM identifies the faulty cell and predicts the fault trajectory, physical disassembly confirms a 3 mm separator perforation, and mean time to repair falls by 41 percent. The paper presents these results as evidence that the five-layer architecture, not merely prompt-based assistance, is the enabling mechanism.

Load-bearing premise

The measured gains come from the framework's distinctive machinery rather than from generic LLM-based data augmentation and output correction that any LLM chat application could provide.

Editorial extensions

If this is right

If the framework is correct, combining an LLM's planning and language skills with a small model's numerical work can cut time to repair and operating cost in real industrial deployments: 41 percent and 27 percent in the energy-storage case.
Temperature prediction and control improve enough to lower energy consumption: weekly consumption dropped from 15,837 kWh to 12,525 kWh while staying inside the 24 to 28 degree Celsius band 97.8 percent of the time.
LLM-based data augmentation and output correction are direct, repeatable levers for improving an existing small model without changing the small model's architecture.
The same five-layer pattern applies across the maintenance lifecycle, including monitoring, diagnosis, prediction, and maintenance suggestions, rather than only to isolated fault-classification tasks.
The architecture is designed to absorb heterogeneous data, including sensor time series, logs, manuals, and images, and to make model outputs traceable, which should support deployment in regulated industrial settings.

Reading between the lines

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

The two case studies compare only a bare LSTM against LLM-augmented and LLM-corrected variants, so a natural next experiment would turn off the knowledge graph, semantic layer, workflow engine, and tool-invocation interfaces one at a time to see which layer actually carries the improvement.
If the predict-then-correct pattern generalizes, the same approach could be applied to other high-frequency industrial signals such as vibration, current, or pressure, not just temperature and battery voltage.
The 41 percent reduction in mean time to repair likely includes process and workflow changes beyond model accuracy, so a follow-up deployment should separate model-driven gains from organizational ones by recording how much of the repair time was saved at each diagnostic step.
If the authors' positioning holds, maintenance systems built today around separate forecasters and rule engines could evolve into hybrid workflows where the LLM orchestrates but the small models remain the safety-critical computation path.

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

The paper is worth a read if you work on LLM-plus-small-model systems for industrial maintenance. It describes Sentosa, a five-layer framework that combines an LLM agent with domain-specific small models, a knowledge graph, RAG, and a workflow engine. What is new is the concrete assembly and the two field deployments: a year-long telecom equipment room study and a two-year energy storage station study. The reported numbers are real and sizable - 15-18% MAE reductions in temperature prediction, improved temperature control accuracy, and a 41% MTTR reduction in the fault-detection case. The paper is also honest in positioning itself as an integration of known components; it cites AutoGPT, LangGraph, RAG, and related work, and its taxonomy of collaboration paradigms is useful.

The soft spot is the evaluation. Table III compares LSTM, LSTM with LLM data augmentation, and LSTM with LLM correction. Both of those variants can be implemented with any LLM API and a plain prompt, without the knowledge graph, semantic layer, workflow engine, or tool-invocation interfaces that define Sentosa. So the experiments do not show that the framework, rather than generic LLM augmentation and correction, produces the gains. Table IV compares 'LLM SM+LLM Correct' against rule-based LSTM, but again there is no decomposition. Case Study II is a single anecdote with no baseline detection accuracy; the MTTR and cost reductions are not tied to any specific component. The paper itself lists LLM-enhanced SMs and SMs-enhanced LLM as existing paradigms in Section II.B, so the case studies may simply be reproducing those paradigms rather than the hybrid workflow.

That said, the central claim is not wildly overreaching. The paper is clearly positioned as an architecture and field-deployment report, not a rigorous method comparison. The descriptions are detailed enough for a practitioner to replicate the setup. It would not be fair to say the framework is invalid; it is more accurate to say the evidence underdetermines which components matter.

Who should read it: practitioners building maintenance systems who want a concrete architecture and deployment numbers, and researchers interested in LLM-agent evaluation pitfalls. Would I cite it? Probably not in my own work, because the evaluation does not isolate a reusable method.

For peer review: a serious editor should send this to referees. The architecture is legitimate and the deployments are real, but the paper needs major revisions: an ablation or at least a baseline against generic LLM augmentation, error bars or repeated trials, and a rewriting of the claims to say we demonstrate a working deployment rather than the framework significantly enhances maintenance efficiency. I would accept it conditionally after those changes.

Referee Report

3 major / 5 minor

Summary. The paper proposes Sentosa LLM, a five-layer architecture for intelligent industrial maintenance that combines domain-specific small models (SMs) with large language models (LLMs). The framework includes an Equipment Layer, Data Layer, Model Layer, Application Layer, and Decision & Interaction Layer, with supporting technologies such as an industrial knowledge graph, semantic layer, vector database, prompt engineering, retrieval-augmented generation, a workflow engine, and standardized tool-invocation interfaces. The authors report two case studies: one on equipment-room temperature prediction and energy saving, where an LSTM baseline is improved via LLM-based data augmentation and LLM-based output correction, and one on internal short-circuit detection in an energy storage power station, where an LLM invokes a small LSTM model and applies spatio-temporal consistency checks. Reported results include a 15.4% MAE reduction from LLM data augmentation, a further roughly 18% reduction from LLM correction, a temperature-control accuracy of 97.8% versus 86.5% for the default strategy, and a 41% reduction in mean time to repair.

Significance. If the evaluation adequately isolated the framework's contributions, the paper would provide a useful architectural template for integrating LLMs and task-specific predictive models in industrial maintenance. The paper gives a clear taxonomy of collaboration paradigms and a detailed system design, and the two field studies address relevant practical problems in telecom equipment rooms and energy storage plants. However, the empirical validation as presented does not distinguish the proposed five-layer architecture from generic LLM-based data augmentation and output correction, and the second case study is largely anecdotal. The central claim that the Sentosa framework 'significantly enhances maintenance efficiency' is therefore plausible but not yet substantiated. The paper's main value at this stage is as an architecture proposal and application description rather than a controlled demonstration of the framework's distinctive benefits.

major comments (3)

[Section IV.C, Table III] The evaluation in Section IV.C and Table III compares an LSTM baseline with (i) LSTM trained on LLM-augmented data and (ii) LSTM outputs refined by LLM correction. Both variants can be implemented with any general-purpose LLM API and plain prompts; they do not use the proposed framework's distinctive components described in Section III (Industrial Knowledge Graph, Semantic Layer, Workflow Engine, or the standardized tool-invocation interfaces). Without an ablation that turns these components on and off, or a control that applies generic LLM augmentation/correction outside the Sentosa framework, the reported 15.4%/17.5% and subsequent ~18% error reductions cannot be attributed to the five-layer architecture. This is load-bearing because the paper's central claim is that the Sentosa framework itself enhances maintenance efficiency.
[Section V.C, Table V] Case Study II is presented as a single anecdotal fault (Module #B23) with no quantitative baseline for detection accuracy, detection lead time, or false-alarm rate. The claims of a 41% MTTR reduction and 27% O&M cost reduction in Table V are not accompanied by a comparison protocol, a control strategy, or an analysis of how the framework components contribute; the same spatio-temporal consistency checks could, for example, be implemented as a rule-based post-processor. As it stands, this case study does not substantiate the statement that 'the proposed framework significantly improves early warning' for internal short circuits.
[Section II.B and III] The taxonomy in Section II.B already lists 'LLM-Enhanced SMs' and 'SMs-Enhanced LLM' as existing collaboration paradigms, and the two experiments in Section IV.C are exactly these two paradigms (LLM data augmentation and LLM output correction). The paper does not identify which of the Sentosa components go beyond these paradigms or provide any test of the 'Hybrid Collaboration' workflow that the framework is designed to enable. Consequently, the paper's novelty claim is under-specified: a clear statement of the expected incremental benefit of the workflow engine and intermediate-representation layers, together with an experiment that exercises them, is needed before the framework can be considered validated.

minor comments (5)

[Section II.A and Table I] The terminology for small models is confusing: Section II.A introduces 'SLM (Small Large Model)' for distilled/quantized LLMs, but Table I labels the same class 'LLM (Small)' and lists DeepSeek-R1-1.5B as an example; please unify the naming and clarify the relationship between SLM and the 'LLM (Small)' column.
[Section IV.C and Table III] The text states that the refined predictions yield 'additional reductions ... of approximately 18%' compared with the augmented-data-only model, but the tabulated MAE reduction is 21.7% (0.383 to 0.300) and RMSE reduction is 19.9% (0.503 to 0.403), and MAPE is not reported in Table III; please provide the exact MAPE values and specify the reference condition for the percentage reductions.
[Table IV] The row labeled 'LLM SM+LLM Correct' uses abbreviations that are not defined; please spell out what 'SM' and 'Correct' denote and describe the control strategy used in that configuration, including which components of the Sentosa framework were active.
[Figure 2 and Section IV] Figure 2 is very dense, and the case-study text in Section IV.C does not indicate which of the displayed modules (knowledge graph, semantic layer, workflow engine, MCP interfaces) were actually enabled for the reported results; adding a mapping from each experimental configuration to the active framework components would make the evaluation interpretable.
[Throughout] There are subject-verb agreement issues in sentences such as 'The SMs is responsible...', 'The LMs provides...', and 'SMs is utilized' in Sections II.B and III; please correct these and ensure consistent capitalization of 'LLM' and 'SMs' throughout the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the framework is descriptive and the reported gains are externally measured, not derived from the framework's own parameters.

full rationale

This paper does not present a derivation chain that reduces to its own inputs. The proposed Sentosa LLM architecture is a five-layer conceptual framework with the Model Layer containing domain-specific small models and an LLM collaboration module. The supporting evidence consists of two empirical case studies with external performance metrics: temperature-prediction MAE/RMSE/MAPE reductions, temperature-control accuracy percentages, total energy consumption in kWh, and mean time to repair / O&M cost reductions. These metrics are measured on real deployments and are not defined in terms of fitted parameters of the proposed framework. The comparisons in Section IV.C are between an LSTM baseline, an LSTM trained on LLM-augmented data, and an LSTM with LLM correction; each is an external predictive or control outcome, not a quantity reconstructed from the framework's definitions. No equation in the paper makes a reported result equal to an input by construction, and no fitted parameter is renamed as a prediction. The paper does invoke prior work by some of the same authors (e.g., references [1] and [10]), but these citations are contextual reviews and not used to force the empirical outcomes or to forbid alternative explanations. The skeptic's concern that the case-study gains might be attributable to generic LLM augmentation and correction rather than to the distinctive Sentosa components is an experimental-validity criticism, not a circularity: the claimed contribution could be unsupported by the ablation design, but that is a question of attribution and external validity, not of the derivation being equivalent to its inputs. Therefore, under the stated criteria, the paper shows no circularity and receives a score of 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claims rest on unverified operational assumptions: the datasets are proprietary, measurement protocols are absent, and the evaluation does not isolate the framework's components. No fitted constants are stated, so the free-parameter ledger is empty; the main burden falls on domain assumptions about data quality, measurement consistency, information leakage, and representativeness of the single fault case.

assumptions (4)

domain assumption LSTM, Kalman filtering, wavelet transform, and EMD are valid and reliable for the described forecasting and fault-detection tasks.
Invoked in Sections IV.C and V.C without task-specific validation or comparison against alternative methods.
domain assumption Operational metrics across compared configurations are measured under identical conditions and are accurately recorded.
Sections IV.C and V.C report energy, accuracy, and MTTR numbers with no measurement protocol, sample sizes, or variance.
domain assumption LLM-generated augmented data and output corrections do not leak test-period information.
Section IV.C does not describe how the LLM augmentation and correction steps were constrained, so information leakage from test intervals cannot be ruled out.
domain assumption The single reported fault case in Case Study II is representative of overall fault-detection performance.
Section V.C and Table V report one micro-short-circuit case with no detection rate, false-alarm rate, or aggregate statistics.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Next-Generation Intelligent Maintenance: Collaborative Fusion of Large and Small Models." pith.science (2026). https://pith.science/paper/NZRLYKBQ

@misc{pith2026250605854,
  author       = {Pith},
  title        = {Pith review of: Towards Next-Generation Intelligent Maintenance: Collaborative Fusion of Large and Small Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NZRLYKBQ}},
  note         = {Machine review of arXiv:2506.05854}
}

read the original abstract

With the rapid advancement of intelligent technologies, collaborative frameworks integrating large and small models have emerged as a promising approach for enhancing industrial maintenance. However, several challenges persist, including limited domain adaptability, insufficient real-time performance and reliability, high integration complexity, and difficulties in knowledge representation and fusion. To address these issues, an intelligent maintenance framework for industrial scenarios is proposed. This framework adopts a five-layer architecture and integrates the precise computational capabilities of domain-specific small models with the cognitive reasoning, knowledge integration, and interactive functionalities of large language models. The objective is to achieve more accurate, intelligent, and efficient maintenance in industrial applications. Two realistic implementations, involving the maintenance of telecommunication equipment rooms and the intelligent servicing of energy storage power stations, demonstrate that the framework significantly enhances maintenance efficiency.

Figures

Figures reproduced from arXiv: 2506.05854 by the authors.

**Figure 2.** Architecture of the Sentosa LLM framework for collaboration between large and small models. 4) Workflow Engine, responsible for orchestrating and scheduling the collaborative workflow between the LLM and SMs based on predefined templates or dynamically, and automatically executing a series of complex task flows that include data processing, model invocation, information integration, and result generation. 5) Tool In… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 13 canonical work pages

[1]

Multi-condition fault diagnosis of dynamic systems: A survey, insights, and prospects,

P. Han, Z. Liu, X. He, S. X. Ding, and D. Zhou, “Multi-condition fault diagnosis of dynamic systems: A survey, insights, and prospects,” IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 15728–15744, 2025

2025
[2]

Fault diagnosis in power grids with large language model,

J. Liu and A. Rahman, “Fault diagnosis in power grids with large language model,”arXiv preprint, 2024

work page 2024
[3]

Faultexplainer: Leveraging large language models for interpretable fault detection and diagnosis,

A. Khan, R. Nahar, H. Chen, G. E. C. Flores, and C. Li, “Faultexplainer: Leveraging large language models for interpretable fault detection and diagnosis,”Computers & Chemical Engineering, p. 109152, 2025

work page 2025
[4]

Fd-llm: Large language model for fault diagnosis of machines,

H. A. A. M. Qaid, B. Zhang, D. Li, S.-K. Ng, and W. Li, “Fd-llm: Large language model for fault diagnosis of machines,”arXiv preprint, 2024

work page 2024
[5]

Integrating llms for explainable fault diagnosis in complex systems,

A. J. Dave, T. N. Nguyen, and R. B. Vilim, “Integrating llms for explainable fault diagnosis in complex systems,”arXiv preprint, 2024

work page 2024
[6]

Llm-based framework for bearing fault diagnosis,

L. Tao, H. Liu, G. Ning, W. Cao, B. Huang, and C. Lu, “Llm-based framework for bearing fault diagnosis,”Mechanical Systems and Signal Processing, vol. 224, p. 112127, 2025

work page 2025
[7]

Maintagt: Sim2real-guided multimodal large model for in- telligent maintenance with chain-of-thought reasoning,

H. He, J. Huang, Q. Li, X. Wang, F. Zhang, K. Yang, L. Meng, and F. Chu, “Maintagt: Sim2real-guided multimodal large model for in- telligent maintenance with chain-of-thought reasoning,”arXiv preprint, 2024

work page 2024
[8]

Artificial intelligence for it operations in era of large language model,

D. Pei, S. Zhang, Y . Sun,et al., “Artificial intelligence for it operations in era of large language model,”ZTE Technology Journal, vol. 30, no. 2, pp. 56–62, 2024

work page 2024

Show all 14 references

[9]

Application of artificial intelligence large language model in power equipment operation and maintenance,

X. Chen, W. Fu, C. Liu, Z. Liu, J. Li, Z. Hu, and D. Hu, “Application of artificial intelligence large language model in power equipment operation and maintenance,”Strategic Study of CAE, vol. 25, no. 1, pp. 180–192, 2025

2025
[10]

An outline of prognostics and health management large model: Concepts, paradigms, and challenges,

L. Tao, S. Li, H. Liu, Q. Huang, L. Ma, G. Ning, Y . Chen, Y . Wu, B. Li, W. Zhang,et al., “An outline of prognostics and health management large model: Concepts, paradigms, and challenges,”Mechanical Systems and Signal Processing, vol. 232, p. 112683, 2025

2025
[11]

Synergetic event understanding: A collaborative approach to cross-document event coreference resolution with large language models,

Q. Min, Q. Guo, X. Hu, S. Huang, Z. Zhang, and Y . Zhang, “Synergetic event understanding: A collaborative approach to cross-document event coreference resolution with large language models,” inProceedings of the 62nd Annual Meeting of the Association for Computational Linguis...

2024
[12]

CoGenesis: A framework collaborating large and small language models for secure context-aware instruction following,

K. Zhang, J. Wang, E. Hua, B. Qi, N. Ding, and B. Zhou, “CoGenesis: A framework collaborating large and small language models for secure context-aware instruction following,” inProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: L...

2024
[13]

Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face,

Y . Shen, K. Song, X. Tan, D. Li, W. Lu, and Y . Zhuang, “Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face,”Advances in Neural Information Processing Systems, vol. 36, pp. 38154–38180, 2023

2023
[14]

Autogen: Enabling next-gen llm applications via multi-agent conversa- tion,

Q. Wu, G. Bansal, J. Zhang, Y . Wu, B. Li, E. Zhu, L. Jiang, X. Zhang, S. Zhang, J. Liu, A. Awadallah, R. W. White, D. Burger, and C. Wang, “Autogen: Enabling next-gen llm applications via multi-agent conversa- tion,”arXiv preprint, 2023

2023

Pith tools

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

[1] [1]

Multi-condition fault diagnosis of dynamic systems: A survey, insights, and prospects,

P. Han, Z. Liu, X. He, S. X. Ding, and D. Zhou, “Multi-condition fault diagnosis of dynamic systems: A survey, insights, and prospects,” IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 15728–15744, 2025

2025

[2] [2]

Fault diagnosis in power grids with large language model,

J. Liu and A. Rahman, “Fault diagnosis in power grids with large language model,”arXiv preprint, 2024

work page 2024

[3] [3]

Faultexplainer: Leveraging large language models for interpretable fault detection and diagnosis,

A. Khan, R. Nahar, H. Chen, G. E. C. Flores, and C. Li, “Faultexplainer: Leveraging large language models for interpretable fault detection and diagnosis,”Computers & Chemical Engineering, p. 109152, 2025

work page 2025

[4] [4]

Fd-llm: Large language model for fault diagnosis of machines,

H. A. A. M. Qaid, B. Zhang, D. Li, S.-K. Ng, and W. Li, “Fd-llm: Large language model for fault diagnosis of machines,”arXiv preprint, 2024

work page 2024

[5] [5]

Integrating llms for explainable fault diagnosis in complex systems,

A. J. Dave, T. N. Nguyen, and R. B. Vilim, “Integrating llms for explainable fault diagnosis in complex systems,”arXiv preprint, 2024

work page 2024

[6] [6]

Llm-based framework for bearing fault diagnosis,

L. Tao, H. Liu, G. Ning, W. Cao, B. Huang, and C. Lu, “Llm-based framework for bearing fault diagnosis,”Mechanical Systems and Signal Processing, vol. 224, p. 112127, 2025

work page 2025

[7] [7]

Maintagt: Sim2real-guided multimodal large model for in- telligent maintenance with chain-of-thought reasoning,

H. He, J. Huang, Q. Li, X. Wang, F. Zhang, K. Yang, L. Meng, and F. Chu, “Maintagt: Sim2real-guided multimodal large model for in- telligent maintenance with chain-of-thought reasoning,”arXiv preprint, 2024

work page 2024

[8] [8]

Artificial intelligence for it operations in era of large language model,

D. Pei, S. Zhang, Y . Sun,et al., “Artificial intelligence for it operations in era of large language model,”ZTE Technology Journal, vol. 30, no. 2, pp. 56–62, 2024

work page 2024