Scaling Legal AI: Benchmarking Mamba and Transformers for Statutory Classification and Case Law Retrieval

Anuraj Maurya

REVIEW 6 major objections 5 minor 25 references

Scaling Legal AI: Benchmarking Mamba and Transformers for Statutory Classification and Case Law Retrieval

T0 review · 6 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Linear-time state-space models match or beat transformer baselines on legal benchmarks while running 2–3 times faster.

desk verdict Useful first-pass benchmark numbers, but the central efficiency claim is confounded by unmatched context lengths and pooling, and the paper is not reproducible without code or protocol details. read the letter →

arxiv 2509.00141 v1 pith:D74RYRX6 submitted 2025-08-29 cs.CY cs.AIcs.LG

Anuraj Maurya This is my paper

classification cs.CYcs.AIcs.LG

keywords state-spacemodelsMambalegalNLPstatutoryclassificationcaselawretrievallong-contextmodelingtransformersbenchmarking

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 asks whether state-space models with linear-time selective mechanisms can replace quadratic-attention transformers for legal AI over very long texts. It benchmarks Mamba and SSD-Mamba against BERT, DeBERTa, and Longformer on statutory classification and case-law retrieval using the ECtHR, EUR-Lex, SCOTUS, and ILDC corpora. The results show that SSD-Mamba matches or beats transformer baselines on most long-context tasks while processing 35–49k tokens per second versus 9–18k for transformers, and that transformers keep a small edge on moderate-length tasks such as SCOTUS issue classification. If this holds, large-scale statutory analysis and precedent retrieval could run on more modest hardware, with whole documents read without truncation and without losing accuracy.

What carries the argument

The central mechanism is Mamba's selective state-space recurrence, which updates a hidden state in linear time rather than computing pairwise attention, letting the model read much longer legal documents without quadratic cost. The benchmarking pipeline applies a sliding window with 20% overlap to every model, then aggregates window-level embeddings—averaging probabilities for multi-label classification, averaging logits for single-label classification, and mean-pooling window embeddings for retrieval—so that the comparison hinges on how each architecture represents long documents.

What would settle it

A controlled comparison in which SSD-Mamba and Longformer run on the same documents with identical window sizes, overlap, and aggregation; if SSD-Mamba's quality advantage vanishes when effective context length is equalized, the claimed tradeoff is an artifact of windowing rather than evidence about state-space models.

Watch

Extended reading notes

Core claim

The core claim is that SSD-Mamba achieves the best overall balance of scalability, accuracy, and efficiency for legal classification and retrieval, making it a strong candidate for large-scale legal AI over statutes and long-form case law. Across the benchmark tables, SSD-Mamba matches Longformer on ECtHR classification, edges out both DeBERTa and Longformer on ILDC statute tagging, leads on ECtHR case retrieval, and stays competitive on ILDC retrieval, while everywhere delivering roughly 3 times the throughput of the transformer baselines. The paper also reports complementary strengths: attention-based models excel on short, structured inputs, while state-space models preserve global cohere

Load-bearing premise

The central conclusion assumes that mean-pooling the same 20%-overlapping windows produces equally faithful document representations for every architecture, but the actual context lengths used by Mamba are never reported, so the throughput and accuracy gaps could come from windowing or pooling choices rather than from the architecture itself.

Editorial extensions

If this is right

Legal AI systems can process full statutes and judgments without truncation or heavy window fragmentation, at a fraction of the compute cost of transformer baselines.
Resource-constrained institutions could deploy statutory tagging and precedent retrieval at near-transformer accuracy with much higher throughput, lowering the hardware barrier to legal AI.
Hybrid architectures that combine selective state-space layers with attention are a natural next step for capturing both global coherence and fine-grained local semantics.
For short, structured legal inputs such as Supreme Court issue classification, transformer baselines remain competitive and should not be discarded in favour of state-space models.
If the efficiency result generalizes, state-space models are a practical backbone for retrieval-augmented legal reasoning, not just classification and retrieval.

Reading between the lines

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

The reported speedups may partly come from Mamba using far longer effective windows than the 512- or 4096-token windows of transformers; equalizing effective context length and windowing would test whether the accuracy gains are architectural or an artifact of the comparison setup.
Mean pooling of window embeddings assumes every part of a long judgment contributes equally; a learned or weighted aggregation could change which architecture wins on retrieval, so the pooling choice is a testable lever.
A natural extension not explored in the paper is to pretrain SSD-Mamba on multilingual legal corpora, which could widen its lead on non-English and multi-jurisdictional tasks.
The benchmark suggests an immediate follow-up experiment: use SSD-Mamba document embeddings as the retriever in a retrieval-augmented generation pipeline for legal question answering, where the long-context advantage could matter most.

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

Short version: the paper gives genuinely new per-model numbers for Mamba and SSD-Mamba on four legal NLP tasks, and the benchmark question is worth asking. But the manuscript as written cannot support its central efficiency claim, and the promised code and data are missing.

What's new: the specific Mamba/SSD-Mamba scores on ECtHR, EUR-Lex, SCOTUS, and ILDC are not in the prior literature, as far as I can tell. The setup is standard sliding-window encoding plus mean-pooled embeddings, which is simple, and that simplicity is a virtue for a first pass. The throughput numbers are also consistent across tables, so the rough direction of the result—SSMs faster than transformers, close in accuracy—is plausible.

Where it falls down: every table reports point estimates with no variance, no seeds, no run counts. Training details (learning rate, epochs, batch size, optimizer) are absent. Hardware and throughput measurement protocol are absent. The abstract promises open-source code and datasets; none are provided. That alone would require major revision, but there is also a load-bearing confound. Transformers are capped at 512 or 4096 tokens with 20% overlap; Mamba's context length is listed as 'Flex' throughout and the actual sequence lengths used are never reported. So the 'several times longer' claim is unquantified, and the 2–3x throughput gap could come from fewer windows and less overlap rather than linear-time inference. The paper's own language about Mamba 'reducing window fragmentation' says the same thing. Without an ablation matching context length, overlap, or pooling, the numbers cannot distinguish 'SSD-Mamba is architecturally better' from 'full-document input with single-vector pooling is better.' The stress-test note is right about that, though I'd soften one point: as a deployment comparison—what a practitioner could actually run—the mismatch is a legitimate design choice, not a sleight of hand. As a scientific benchmark about architecture, it needs the controlled condition.

Minor but worth noting: the 'first comprehensive benchmark' label collides with their own reference [21], which already benchmarks Mamba for legal document ranking; they need to say what this paper adds beyond it.

Who should read it: people choosing between SSMs and transformers for legal NLP deployments might get a rough sketch. I would not cite the current numbers. The paper deserves peer review more than desk rejection: the question is real, the measurements are new, and the flaws are fixable by releasing artifacts and adding matched-context ablations. But it needs major revision first.

Referee Report

6 major / 5 minor

Summary. The paper compares Mamba and SSD-Mamba with BERT/DeBERTa/Longformer on ECtHR, EUR-Lex, SCOTUS, and Indian ILDC/ILC corpora for statutory classification and case-law retrieval. The methodology uses sliding windows with 20% overlap and pools window-level outputs; results are reported as F1, AUC, MAP, MRR, R@10, nDCG@10, and tokens/sec. The paper claims that Mamba's linear scaling processes legal documents several times longer than transformers and that SSD-Mamba offers the best balance of scalability, accuracy, and efficiency. As submitted, the manuscript lacks essential experimental documentation and the design confounds model architecture with input segmentation and aggregation, so the empirical basis for the central claims is not verifiable.

Significance. A careful, controlled comparison of state-space models and transformers on public legal benchmarks would be a useful contribution to legal NLP. The paper selects relevant tasks and datasets and addresses a timely question. However, the reported results are not reproducible: hyperparameters are omitted, no variance or significance information is given, the 'Flex' context length is never quantified, and throughput conditions are unspecified. More importantly, the comparison is confounded by the use of different window sizes and pooling strategies across architectures. If a revised version supplied exact protocols, ablations, and matched settings, the work could be valuable; in its current form the empirical conclusions are not established.

major comments (6)

[Tables I–VI] All reported metrics are single point estimates with no standard deviation, number of runs, seeds, or significance testing. Several headline differences are small (e.g., Table IV: SSD-Mamba Acc 77.1 vs DeBERTa 76.1; Table VI: nDCG@10 68.1 vs DeBERTa 67.8). Without repeated runs or error bars, the claim that SSD-Mamba 'matches or surpasses' transformers is not supported by the data presented.
[Section II.B / Tables I–VI] The context length for Mamba/SSD-Mamba is reported as 'Flex' and never quantified. The abstract's claim of processing documents 'several times longer' than transformers depends on the actual sequence lengths used and the resulting number of windows per document. The paper should report the maximum sequence length used for each Mamba model, the distribution of document lengths and window counts, and how overlapping tokens are counted. Without these numbers, the Tok/s comparisons and the scalability claim are untestable.
[Sections II.A, II.C, II.E / III.E] The comparison conflates architecture with input segmentation and aggregation. Transformers are restricted to 512/4096-token windows while Mamba processes longer contexts and generates fewer windows. Since classification outputs and retrieval embeddings are mean-pooled across windows, the observed gains could reflect the benefits of full-document processing or the pooling rule rather than the selective state-space mechanism. No ablation matches Mamba at transformer-length windows, varies the overlap fraction, or isolates pooling choices. The paper's own statement that 'Mamba and SSD-Mamba reduce window fragmentation' (Section II.D) confirms this confound, so the central architectural claim in Section III.E is not justified.
[Tables I–VI / Section II.A] Throughput (Tok/s) is reported without any measurement protocol: no hardware type, batch size, model precision, warm-up, sequence length, or statement about whether overlapping tokens are counted. These choices strongly affect Tok/s; for example, larger batches alone can change throughput by several factors. Without a documented protocol, the reported 2–3x (and up to 5x) throughput advantages over transformers cannot be interpreted or reproduced.
[Section II.A] Training details are missing: learning rate, optimizer, batch size, number of epochs, warmup, maximum sequence length per model, early stopping, and checkpoint selection. The paper states only that models are 'fine-tuned' or 'evaluated.' Without these hyperparameters, the results cannot be reproduced, and differences among models may be due to differential tuning rather than architectural properties. This omission is load-bearing because the paper's conclusions are entirely empirical.
[Section I.I / References [21], [22]] The paper claims to present 'the first comprehensive benchmark of Mamba versus transformer models for legal NLP,' but it cites [21], 'Benchmarking Mamba's document ranking performance on legal data,' without discussing or comparing against that prior work. The novelty claim is therefore overstated, and the contribution should be positioned relative to existing Mamba-legal benchmarks. Reference [21] also appears to be a placeholder ('J. Doe'), which further weakens the literature review.

minor comments (5)

[Section I.F] The statement that 'each additional token increases memory and computation exponentially' is inaccurate; standard transformer self-attention has quadratic, not exponential, complexity in sequence length.
[Abstract / Section IV] The abstract promises 'open-source code and datasets to support reproducibility,' but no repository, URL, or data access link appears in the manuscript. Provide the links or remove the claim.
[Section II.A.4] The ILDC/ILC task list includes 'legal reasoning extraction (multi-label),' but no results for this task appear in Tables I–VI. Either report results or clarify that this task was not included in the current evaluation.
[References] Reference [21] appears to be a placeholder, and reference [22] is a blog post rather than a peer-reviewed source. These are insufficient for supporting the claim that prior Mamba legal benchmarks exist; replace with verifiable academic references.
[Tables I–VI] The 'Len' column lists 'Flex' for Mamba models, which is not a length. Specify the actual maximum context length used, or replace the entry with the average/maximum document length and window count.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical benchmark whose reported advantages are direct measurements, not quantities defined by fitted parameters or self-citations.

full rationale

This paper does not contain a derivation chain of the kind circularity analysis targets. It reports fine-tuned/evaluated model scores (Tables I–VI) for classification and retrieval; the central claims (e.g., 'Mamba matches or surpasses transformers while maintaining 2–3x higher throughput') are summary readings of those measured tables, not quantities re-derived from their own inputs. No parameter is fitted to a subset and then called a prediction: no fine-tuning or fitting procedure is described in a way that would create a fitted-input/prediction pair. The 'Flex' context length for Mamba/SSD-Mamba in Tables I–VI is an input configuration (Section II.B), not a predicted output, so saying that Mamba processes longer contexts is a capacity property of the chosen settings, not a circular derivation. There are no self-citations by the sole author; the only relevant prior Mamba-legal benchmark is the external J. Doe preprint [21], cited in Section I.G, and its existence merely tensions the 'first comprehensive benchmark' novelty claim in Section I.I. That is an overstatement/scope issue, not circularity. Likewise, the concern that windowing and pooling choices may confound architectural comparisons is an experimental-design threat to validity, not a reduction of the result to its inputs. Under the stated rules, no circular step can be quoted, so the appropriate finding is no significant circularity (score 0).

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

The central claims rest on an unreleased benchmark setup: window overlap and pooling are hand-chosen, training configurations are unreported, and throughput is presented without measurement conditions. These are assumptions the reader cannot verify from the paper.

free parameters (4)

Unreported training hyperparameters (learning rate, epochs, batch size, warmup, seed)
Fine-tuning results depend on these; Section II reports none.
Sliding-window overlap fraction = 20%
Chosen by hand in Section II.A without ablation.
Embedding aggregation rule = mean pooling / averaging of probabilities and logits
Section II.C/II.E selects averaging without experimental justification.
Throughput measurement configuration
Tokens/sec numbers in Tables I-VI lack hardware and batch-size reporting.

assumptions (4)

domain assumption Sliding windows with 20% overlap preserve all document content and apply uniformly across all models
Stated in Section II.A; no information-loss analysis provided.
domain assumption Mean-pooling of window embeddings yields a faithful document representation for classification and retrieval
Section II.C and II.E select averaging without comparing to max-pooling or attention pooling.
domain assumption Throughput measured in tokens/sec is comparable across models under the same unstated conditions
Tables I-VI list Tok/s without hardware, batch size, or measurement method.
domain assumption The public datasets (LexGLUE, EUR-Lex, ILDC) are used with correct labels and standard splits
Section II.A names datasets but gives no versions or split scripts.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Scaling Legal AI: Benchmarking Mamba and Transformers for Statutory Classification and Case Law Retrieval." pith.science (2026). https://pith.science/paper/D74RYRX6

@misc{pith2026250900141,
  author       = {Pith},
  title        = {Pith review of: Scaling Legal AI: Benchmarking Mamba and Transformers for Statutory Classification and Case Law Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D74RYRX6}},
  note         = {Machine review of arXiv:2509.00141}
}

read the original abstract

The rapid growth of statutory corpora and judicial decisions requires scalable legal AI systems capable of classification and retrieval over extremely long contexts. Transformer-based architectures (e.g., Longformer, DeBERTa) dominate current legal NLP benchmarks but struggle with quadratic attention costs, limiting efficiency and scalability. In this work, we present the first comprehensive benchmarking of Mamba, a state-space model (SSM) with linear-time selective mechanisms, against leading transformer models for statutory classification and case law retrieval. We evaluate models on open-source legal corpora including LexGLUE, EUR-Lex, and ILDC, covering statutory tagging, judicial outcome prediction, and case retrieval tasks. Metrics include accuracy, recall at k, mean reciprocal rank (MRR), and normalized discounted cumulative gain (nDCG), alongside throughput measured in tokens per second and maximum context length. Results show that Mamba's linear scaling enables processing of legal documents several times longer than transformers, while maintaining or surpassing retrieval and classification performance. This study introduces a new legal NLP benchmark suite for long-context modeling, along with open-source code and datasets to support reproducibility. Our findings highlight trade-offs between state-space models and transformers, providing guidance for deploying scalable legal AI in statutory analysis, judicial decision support, and policy research.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 16 canonical work pages

[21]

Benchmarking mamba’s document ranking performance on legal data,

J. Doe, “Benchmarking mamba’s document ranking performance on legal data,” 2024, preprint, Legal NLP Workshops

work page 2024
[2]

Data-centric and logic-based models for automated legal problem solving,

L. K. Branting, “Data-centric and logic-based models for automated legal problem solving,” Artificial Intelligence and Law , vol. 25, no. 1, pp. 5–27, 2017. [Online]. Available: https://doi.org/10.1007/s10506- 017-9190-4

doi:10.1007/s10506- 2017
[3]

K. D. Ashley, Artificial Intelligence and Legal Analytics: New Tools for Law Practice in the Digital Age . Cambridge University Press, 2017. [Online]. Available: https://www.cambridge.org/core/books/artificial- intelligence-and-legal-analytics/

work page 2017
[4]

Legislative updates in the digital era,

A. Smith, “Legislative updates in the digital era,” Law & Technology Review, 2020

work page 2020
[5]

Taxman: An experiment in artificial intelligence and legal reasoning,

M. Sarner and L. McCarty, “Taxman: An experiment in artificial intelligence and legal reasoning,” in Proceedings of the 9th International Joint Conference on Artificial Intelligence , 1985

work page 1985
[6]

Hypo: A case-based reasoning system for argumentation,

E. Rissland, K. Ashley, and R. Loui, “Hypo: A case-based reasoning system for argumentation,” in Proceedings of AAAI , 1987

work page 1987
[7]

A. v. d. L. Gardner, An Artificial Intelligence Approach to Legal Reasoning . MIT Press, 1987. [Online]. Avail- able: https://mitpress.mit.edu/9780262570754/an-artificial-intelligence- approach-to-legal-reasoning/

arXiv 1987
[8]

Machine learning in legal document classification,

C. Bruce, “Machine learning in legal document classification,” Law, Probability and Risk , 2004

work page 2004

Show all 25 references

[9]

Semantic retrieval of legal documents,

L. Bing and H. Chan, “Semantic retrieval of legal documents,” Artificial Intelligence and Law , 2010

2010
[10]

Predictive analytics and law: Models, outcomes, and fairness,

K. Goodman, “Predictive analytics and law: Models, outcomes, and fairness,” Stanford Technology Law Review, 2007

2007
[11]

Tetlock, Expert Political Judgment: How Good Is It? Princeton University Press, 2007

P. Tetlock, Expert Political Judgment: How Good Is It? Princeton University Press, 2007

2007
[12]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems , 2017

2017
[13]

A general approach for predicting the behavior of the supreme court of the united states,

D. M. Katz, M. J. Bommarito, and J. Blackman, “A general approach for predicting the behavior of the supreme court of the united states,” PLoS ONE , vol. 12, no. 4, p. e0174698, 2017. [Online]. Available: https://doi.org/10.1371/journal.pone.0174698

2017 doi
[14]

Legal summarization models and their practical performance,

I. Chalkidis et al. , “Legal summarization models and their practical performance,” Journal of Artificial Intelligence Research , 2022

2022
[15]

Predicting judicial decisions of the european court of human rights: A natural language processing perspective,

N. Aletras, D. Tsarapatsanis, D. Preot ¸iuc-Pietro, and V . Lampos, “Predicting judicial decisions of the european court of human rights: A natural language processing perspective,” PeerJ Computer Science , vol. 2, p. e93, 2016. [Online]. Available: https://doi.org/10.7717/pee...

2016 doi
[17]

Longformer: The long-document transformer,

I. Beltagy, M. Peters, and A. Cohan, “Longformer: The long-document transformer,” in arXiv preprint arXiv:2004.05150 , 2020

2004 arXiv
[18]

Big bird: Transformers for longer sequences,

M. Zaheer, G. Guruganesh, A. Dubey, J. Ainslie, C. Alberti, S. Ontanon, P. Pham, A. Ravula, Q. Wang, L. Yang, and A. Ahmed, “Big bird: Transformers for longer sequences,” in Advances in Neural Information Processing Systems, 2020

2020
[19]

Combining recurrent, convolutional, and continuous-time models with linear state-space layers,

A. Gu, K. Goel, and C. R ´e, “Combining recurrent, convolutional, and continuous-time models with linear state-space layers,” in Advances in Neural Information Processing Systems (NeurIPS) , 2021. [Online]. Available: https://arxiv.org/abs/2110.13985

2021 arXiv
[20]

Mamba: Linear-time sequence modeling with selective state spaces,

A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752, 2024. [Online]. Available: https://arxiv.org/abs/2312.00752

2024 arXiv
[22]

Mamba explained—a potential replacement for transform- ers,

T. Hossain, “Mamba explained—a potential replacement for transform- ers,” The Gradient, 2024

2024
[23]

Legal-bert: The muppets straight out of law school,

I. Chalkidis, M. Fergadiotis, P. Malakasiotis, N. Aletras, and I. Androutsopoulos, “Legal-bert: The muppets straight out of law school,” in Findings of EMNLP , 2020, pp. 2898–2904. [Online]. Available: https://aclanthology.org/2020.findings-emnlp.261

2020
[24]

Lexglue: A benchmark dataset for legal language understanding in english,

I. Chalkidis, M. Fergadiotis, and N. Aletras, “Lexglue: A benchmark dataset for legal language understanding in english,” arXiv preprint arXiv:2110.00976, 2021

2021 arXiv
[25]

The open case law project: Open data for legal ai benchmarking,

P. Henderson, S. Zhang et al. , “The open case law project: Open data for legal ai benchmarking,” 2023, preprint

2023
[26]

Benchmarking the ability of large language models to ground legal reasoning in statutory text,

S. Leivaditi, P. Malakasiotis, and I. Androutsopoulos, “Benchmarking the ability of large language models to ground legal reasoning in statutory text,” in LegalNLP Workshop, NAACL, 2021

2021
[2019]

Available: https://arxiv.org/abs/1904.10509

[Online]. Available: https://arxiv.org/abs/1904.10509

1904 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [21]

Benchmarking mamba’s document ranking performance on legal data,

J. Doe, “Benchmarking mamba’s document ranking performance on legal data,” 2024, preprint, Legal NLP Workshops

work page 2024

[2] [2]

Data-centric and logic-based models for automated legal problem solving,

L. K. Branting, “Data-centric and logic-based models for automated legal problem solving,” Artificial Intelligence and Law , vol. 25, no. 1, pp. 5–27, 2017. [Online]. Available: https://doi.org/10.1007/s10506- 017-9190-4

doi:10.1007/s10506- 2017

[3] [3]

K. D. Ashley, Artificial Intelligence and Legal Analytics: New Tools for Law Practice in the Digital Age . Cambridge University Press, 2017. [Online]. Available: https://www.cambridge.org/core/books/artificial- intelligence-and-legal-analytics/

work page 2017

[4] [4]

Legislative updates in the digital era,

A. Smith, “Legislative updates in the digital era,” Law & Technology Review, 2020

work page 2020

[5] [5]

Taxman: An experiment in artificial intelligence and legal reasoning,

M. Sarner and L. McCarty, “Taxman: An experiment in artificial intelligence and legal reasoning,” in Proceedings of the 9th International Joint Conference on Artificial Intelligence , 1985

work page 1985

[6] [6]

Hypo: A case-based reasoning system for argumentation,

E. Rissland, K. Ashley, and R. Loui, “Hypo: A case-based reasoning system for argumentation,” in Proceedings of AAAI , 1987

work page 1987

[7] [7]

A. v. d. L. Gardner, An Artificial Intelligence Approach to Legal Reasoning . MIT Press, 1987. [Online]. Avail- able: https://mitpress.mit.edu/9780262570754/an-artificial-intelligence- approach-to-legal-reasoning/

arXiv 1987

[8] [8]

Machine learning in legal document classification,

C. Bruce, “Machine learning in legal document classification,” Law, Probability and Risk , 2004

work page 2004