A Machine Learning Theory Perspective on Strategic Litigation

Aloni Cohen; Avrim Blum; Han Shao; Melissa Dutz

REVIEW 4 major objections 4 minor 8 references

A Machine Learning Theory Perspective on Strategic Litigation

T0 review · 4 major / 4 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 paper models strategic litigation as a data-selection problem and proves exact optimality results in two simple settings: a quadratic-time dynamic program for nearest-neighbor courts, and a two-case achievability condition for SVM…

desk verdict Novel model with a real counterexample in the main SVM characterization—worth a revise. read the letter →

arxiv 2506.03411 v2 pith:VOXW7GEO submitted 2025-06-03 cs.LG cs.GT

Melissa Dutz , Han Shao , Avrim Blum , Aloni Cohen This is my paper

classification cs.LGcs.GT MSC 68Q32

keywords strategiclitigationimpactcommonlawprecedentnearestneighborclassifiersupportvectormachineteachingdimensionclean-labeldatapoisoningalgorithmicgametheory

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

Strategic litigation, the paper argues, is a data-selection problem: a litigator who brings cases to a high court is choosing training points for a lower court's learning algorithm, and later cases are decided by whatever rule that algorithm infers from precedent. The paper shows that even in simple versions of this model, the optimal choice of cases has exact structure. In a one-dimensional setting where the lower court uses nearest-neighbor classification, the litigator can find a subset of a finite pool of cases that minimizes expected error with respect to their goal in $O(|P|^2)$ time. In a $d$-dimensional setting where the lower court uses a maximum-margin linear classifier, a desired linear rule is achievable exactly when it is less than $90^\circ$ from the high court's rule and no historical data obstructs it, and when it is achievable two cases are enough. The paper also shows that bringing a case the high court will decide against can be arbitrarily helpful, and that short-sighted selection strategies fail irreversibly.

What carries the argument

The key object is the abstraction of a case as a labeled point in instance space and of precedent as a training set, so the lower court's rule is $f = A(S_h \cup S_l)$ for a deterministic learning algorithm $A$. In the one-dimensional case, nearest-neighbor classification makes each selected case a boundary point, and the dynamic program works because only the rightmost selected point affects the learned rule to its right, so error decomposes interval by interval. In higher dimensions, the SVM's separating hyperplane is determined by support vectors, and its normal vector is a nonnegative linear combination of labeled points; that fact alone rules out any goal more than $90^\circ$ away from the high court's rule, while placing one positive and one negative point equidistant from the target hyperplane makes those two points the only support vectors and yields exactly $g$. The overturning-precedent construction uses convex hulls: the litigator adds points so that any linear separator consistent with them must label old conflicting precedents according to $f^*$, forcing the high court to delete those precedents.

What would settle it

In the one-dimensional nearest-neighbor setting, run the dynamic program against an exhaustive search over all subsets of a small random pool and compare the resulting expected errors; any mismatch would refute Theorem 5.1. For the SVM setting, generate random high-court functions, goal functions, and historical data, test the conditions of Theorem 5.3, and brute-force search over added pairs of points: one case where the conditions fail yet the goal is learned, or hold yet the goal is not, would refute the characterization.

Watch

Extended reading notes

Core claim

The central claim is that strategic impact litigation can be formalized as the problem of choosing a set of fact patterns to feed to a known learning algorithm, with labels assigned by a known high-court function $f^*$. Given that formalization, the paper proves exact characterizations of the litigator's power. For one-dimensional nearest-neighbor lower courts, the optimal subset of any finite pool $P$ can be computed in $O(|P|^2)$ time by dynamic programming (Theorem 5.1), and the same recurrence handles a budget of at most $k$ cases. For $d$-dimensional SVM lower courts, a linear goal $g$ is achievable if and only if the angle between the normal vectors of $g$ and $f^*$ is less than $90^\circ$, no historical point lies in the disagreement region or on the boundary of $g$, and, in the parallel case, no historical point lies within a margin $\delta$ of $g$; whenever these conditions hold, two cases suffice (Theorem 5.3). When conflicting precedent exists and the high court may overturn it, the same goal remains achievable with at most $2d+1$ cases (Theorem 6.1).

Load-bearing premise

The entire analysis assumes the litigator knows the high court's labeling function and the lower court's learning algorithm exactly; all optimality and achievability results depend on that perfect knowledge.

Editorial extensions

If this is right

In the one-dimensional nearest-neighbor model, a litigator with any finite pool of available cases and any distribution over future cases can compute an optimal set of cases in polynomial time, and can solve the constrained version that uses at most $k$ cases.
In the SVM model, an achievable goal never requires more than two new cases; if the literal goal is not achievable, the best achievable proxy can be approximated within $\epsilon$ by enumerating a sample of size $\tilde{O}(d/\epsilon^2)$, using the shattering bound and a uniform convergence argument.
Bringing a case the high court will rule against can be exactly the right move, and the benefit can be arbitrarily large when the litigator's favorable region is small.
Myopic selection and one-step-lookahead selection are provably insufficient; some goals are reachable only by sequences that temporarily increase error, so case selection must be planned globally.
In the extension with precedent that contradicts the current high court, achieving a desired rule requires at most $2d+1$ cases, independent of how many conflicting precedents exist.

Reading between the lines

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

The $90^\circ$ angle condition suggests a practical heuristic for real impact litigation: when the target doctrine is more than orthogonal to the sitting court's apparent doctrine, direct installation of that target is impossible under max-margin lower courts, so litigators should pursue an intermediate achievable rule or first shift the court's own doctrine.
The model is a special case of clean-label data poisoning in which the attacker chooses inputs but an oracle chooses labels; the same 'oracle-labeled poisoning' abstraction could transfer to other settings where data is labeled by an authority rather than the attacker, though the paper does not draw this generalization.
Because the analysis assumes a single active litigator, the optimal case sets are best understood as best responses in a larger game with competing litigators, where the DP and achievability tests would enter as components of a strategic equilibrium.
The inducible-set characterization yields a testable empirical prediction: if lower courts behave like max-margin learners, successful strategic litigation should rarely produce doctrines at or beyond $90^\circ$ from the prevailing high-court doctrine, and a systematic study of landmark impact cases could check this.

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

Quick take: this is a genuinely new model—strategic litigation as teaching a known learner, with the twist that examples are labeled by the court's f*, not the litigator's goal. The 1D nearest-neighbor DP is correct as far as I can tell, and the connection to teaching dimension is apt. But the main SVM achievability theorem (Thm 5.3) is false as stated: the θ=0 condition 'no historical data less than distance δ from g' is too weak. Consider d=1, f* positive for x>-1 with f*(-1)=+1, negative for x<-1; g positive for x>0, negative for x<0; historical point (1,+1). Then δ=max(dist to positive region, dist to negative region)=max(0,1)=1, and the historical point at distance exactly 1 is allowed. But no negative training point exists at distance 1 from g (the nearest negative points are all >1 away), so any finite labeled set makes SVM output a separator strictly left of 0; g is approachable only as a limit. The proof's claim that any pair of equidistant points at least δ away suffices fails because the only point at distance δ on the negative side is labeled positive under this f*. So the characterization needs either a strict 'within distance δ' condition, a boundary-label specification, or a generic-position assumption. This is not a philosophical quibble; it's a counterexample to a central theorem. It is probably fixable, but it has to be fixed before the results can be treated as established. Also, Theorem 6.1 uses infinite points and limiting arguments that are sketches rather than proofs, and the O(|P|^2) runtime assumes constant-time geometric subroutines. The model's knowledge assumptions (litigator knows f* and A) are strong but honestly stated. Bottom line: the paper deserves serious peer review—the model is worth having—but it needs a revised theorem and tightened proofs. I'd send it out.

Referee Report

4 major / 4 minor

Summary. The paper introduces a stylized model of strategic impact litigation in a common-law system, in which a lower court applies a learning algorithm A to precedents labeled by a high court rule f*, and a strategic litigator with full knowledge of f* and A chooses cases to steer the learned rule toward a goal g. The paper studies two instantiations: one-dimensional nearest-neighbor classification and d-dimensional linear separators learned by an SVM. Structural examples show that myopic case selection can be suboptimal and that short-term losses may be necessary. Theorem 5.1 claims an O(|P|^2) dynamic program for optimal case selection in the 1D setting; Theorem 5.3 characterizes when a linear goal is exactly achievable with at most two cases; Section 6 extends the model to overturning precedent, with Theorem 6.1 claiming achievability using at most 2d+1 cases. The paper is framed as an initial ML-theoretic exploration of precedent-based litigation.

Significance. The paper's formal model is transparent and its main structural observations are pedagogically useful. The 1D dynamic program and the SVM achievability characterization are natural first results for a new problem, and the connections to teaching dimension and clean-label poisoning are appropriate. The paper has no fitted parameters and the proofs proceed from the model definitions, which is a strength. However, the central SVM characterization is false as stated, and the proof of the overturning-precedent extension is not fully rigorous; these issues must be repaired before the paper's main claims can be accepted.

major comments (4)

[Section 5.2.1, Theorem 5.3] The θ=0 bullet of Theorem 5.3 is false as stated. In d=1, take f*(x)=+1 for x≥-1 and f*(x)=-1 for x<-1, g(x)=+1 for x>0 and -1 for x<0, and S_h={(1,+1)}. Here θ=0, the distance from g to f*'s positive region is 0 and to its negative region is 1, so δ=1; the historical point at distance exactly 1 from g is not 'less than' δ, so the condition is satisfied. Yet g is not achievable: every negative training point lies strictly left of -1 (distance >1 from g), while the closest positive point is the historical point at distance 1, so for any finite S_l the SVM's max-margin threshold lies strictly left of 0. The proof's sufficiency claim that 'any pair of equidistant points at least δ away on either side of g will suffice' fails because the point at distance δ on the negative side (x=-1) is labeled positive by f* under this convention. The theorem needs either a strict 'no historical data within distance δ' condition, an explicit specification of f*'s boundary labels, or a generic-position assumption excluding points at distance exactly δ.
[Section 6.2, Theorem 6.1] The statement of Theorem 6.1 is ambiguous or incorrect when S_h contains points whose labels disagree with f* and g. It says any g satisfying the conditions of Theorem 5.3 is achievable in the overturning setting, but any disagreeing point lies in the disagreement region of f* and g and therefore violates the first bullet of Theorem 5.3; if no such points exist, the theorem reduces to Theorem 5.3. The intended use of Theorem 5.3's conditions on the subset S_h \ S_dis should be stated explicitly. Independently, the proof is not a finite constructive proof: it adds points at ∞·e_i and uses informal limits ('when ε→0 and α→∞', 'when ε is small enough'). For a theorem asserting achievability with at most 2d+1 points, the proof must show for every finite S_dis that finite parameter choices exist with strictly positive margins; this can likely be fixed by writing 'for all sufficiently small ε and sufficiently large α', but the current text is not a complete proof.
[Section 5.1 and Appendix A, Theorem 5.1] The claimed O(|P|^2) runtime is not established by the proof as written. The correctness proof in Appendix A treats nearest_neighbor(interval, S) and discrepancy(f,g,interval,D) as constant-time utilities, but these are nontrivial geometric and integration subroutines: discrepancy requires measuring disagreement with respect to an arbitrary distribution D, and nearest-neighbor predictions over an interval depend on all of S_h and on the selected points. Without a stated representation of D or explicit implementations of these utilities, the O(|P|^2) bound is vacuous. Please either state a concrete computation model (e.g., piecewise-constant densities and arithmetic oracles) or provide preprocessing and implementation details with their costs.
[Sections 5.2.2 and 5.2.3] The proposed method for handling unachievable goals is under-specified. Section 5.2.2 says to draw a sample S, enumerate all achievable linear classifiers corresponding to distinct labelings of S, and choose the one with minimal sample error, but no procedure is given for enumerating only the achievable classifiers; this depends on the achievability conditions, which currently suffer from the error in Theorem 5.3. The sample-complexity bound is also garbled in the text (the formula appears as |S| = e^{O(d/ε^2)}); the standard uniform-convergence bound for VC dimension d+1 is polynomial, roughly O((d log(1/ε) + log(1/δ))/ε^2). Please clarify the enumeration step and correct the bound.

minor comments (4)

[Section 2.1, Eq. (1)] The assumption that the litigator knows f* and A exactly is very strong and is only acknowledged in the Discussion; it should be stated prominently in the model section as a simplifying idealization.
[Section 2.2] The nearest-neighbor classifier is not fully specified in the case of ties (points exactly equidistant from two training points); this can affect the discrepancy computation when D has point masses. A tie-breaking convention (or a continuity assumption on D) would make the 1D results rigorous.
[Algorithm 1] There are small typographical inconsistencies, including 'case fact patters' and the use of '𝑓' versus 'A' in the abstract and main text; these should be cleaned up.
[Section 5.2.1, proof of Theorem 5.3] The phrase 'at least as close to g as any historical data' should be clarified as 'strictly closer than every historical data point' or 'at distance no larger than the minimum historical distance', because the current wording is what leads to the incorrect inclusion of points at distance exactly δ.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all main theorems are proved from the model definitions, and the only overlapping-author citation is related work, not load-bearing.

full rationale

The paper's central claims are self-contained derivations. Theorem 5.1 is an optimization algorithm over the 1D nearest-neighbor classifier; its correctness proof (Appendix A) uses only the definition of nearest-neighbor intervals and the discrepancy objective, with no fitted parameters and no imported prior result. Theorem 5.3 characterizes SVM-inducible linear goals directly from the SVM dual (weights are signed convex combinations) and from a two-point symmetric construction; the citation to [5] appears for the special case f*=g with no historical data, but the proof immediately gives an independent general construction, so the citation is not load-bearing. The self-citation to [1] (Blum et al.) is a related-work pointer to clean-label poisoning and is never used as a premise for any theorem. Section 5.2.2's enumeration uses standard Sauer-Shelah and VC-dimension facts as external bounds, not as restatements of the paper's results. Section 6.1 and Theorem 6.1 are proven by explicit convex-hull/separator constructions. The only caveat worth noting is a technical boundary-condition concern in Theorem 5.3's parallel case (points exactly at distance delta can interact with f*'s boundary labels); that is a correctness issue, not a circularity. No equation reduces to its own input and no fitted value is renamed as a prediction.

Assumptions & free parameters 0 free parameters · 5 assumptions · 1 invented entities

The paper's results are derived within a deliberately stylized model, so the main intellectual burden is on the modeling assumptions rather than on fitted data: a two-level hierarchy, a known high-court function, a known deterministic learner, and realizability. The SVM results also invoke standard facts about max-margin separators and uniform convergence. The overturning-precedent extension in Section 6 introduces an additional ad hoc rule for removing inconsistent precedent.

assumptions (5)

domain assumption Two-level court hierarchy with fixed high-court rule f* and lower court learning from precedent via deterministic algorithm A.
Section 2.1 establishes this as the basic model; all results are stated within it.
domain assumption The strategic litigator knows f* and A and can choose any case from a finite pool P (or the whole instance space).
Section 2.1; the optimization in (1) assumes this knowledge. The paper lists uncertainty about f* and A as future work in Section 7.
domain assumption f* is realizable in the hypothesis class H and A always returns a consistent function.
Section 2.1; both the SVM characterization and the DP rely on consistency.
ad hoc to paper Overturning-precedent extension: high court removes a minimal set of disagreeing points only when a new case makes the data inconsistent.
Section 6 introduces this rule to model precedent overruling; it is not empirically grounded.
standard math Standard learning theory results: Sauer-Shelah lemma, VC dimension uniform convergence, SVM margin geometry, and the teaching dimension of SVM.
Used in Sections 5.2.2 and 5.3; Theorem 5.3 relies on the max-margin property.

invented entities (1)

Strategic litigator with full knowledge of f* and A
purpose: To study how a well-informed actor can shape precedent by selecting cases.
Idealized agent introduced in Section 2.1; the paper acknowledges real litigators have uncertainty about court behavior (Section 7), so this is a modeling construct without independent empirical backing.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Machine Learning Theory Perspective on Strategic Litigation." pith.science (2026). https://pith.science/paper/VOXW7GEO

@misc{pith2026250603411,
  author       = {Pith},
  title        = {Pith review of: A Machine Learning Theory Perspective on Strategic Litigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VOXW7GEO}},
  note         = {Machine review of arXiv:2506.03411}
}

read the original abstract

Strategic litigation involves bringing a case to court with the goal of having an impact beyond resolving the particular dispute at hand. In a common law system, one way a case may have far-reaching impact is by establishing new legal precedent that later courts must follow. In this paper, we explore strategic litigation from the perspective of machine learning theory. We consider an abstract model of a common law legal system where a lower court decides new cases by applying a decision rule learned from a higher court's past rulings. In this model, we explore the power of a strategic litigator, who strategically brings cases to the higher court to influence the decision rule applied by the lower court in future cases. We explore questions including: What impact can a strategic litigator have? Which cases should a strategic litigator bring to court? Does it ever make sense for a strategic litigator to bring a case when they are sure the court will rule against them? We show that this strategic case selection problem has interesting structure, with even simple settings exhibiting counterintuitive phenomena. When cases are represented by points in one dimension and the lower court's learning algorithm is nearest neighbor, or as points in d dimensions and the lower court's learning algorithm is a support vector machine, we characterize the set of inducible decision rules and develop algorithms for selecting an optimal set of cases to bring to the higher court given the strategic litigator's objectives.

Figures

Figures reproduced from arXiv: 2506.03411 by the authors.

**Figure 6.** Frame A illustrates the first case from Theorem 5.3. The two points with bolded border lying on [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

8 extracted references · 5 canonical work pages

[1]

Avrim Blum, Steve Hanneke, Jian Qian, and Han Shao. 2021. Robust learning under clean-label attack. In Conference on Learning Theory . PMLR, 591–634

work page 2021
[2]

L Karl Branting, Craig Pfeifer, Bradford Brown, Lisa Ferro, John Aberdeen, Brandy Weiss, Mark Pfaff, and Bill Liao. 2021. Scalable and explainable legal prediction. Artificial Intelligence and Law 29 (2021), 213–238

work page 2021
[3]

Goldman and M.J

S.A. Goldman and M.J. Kearns. 1995. On the Complexity of Teaching.J. Comput. System Sci. 50, 1 (1995), 20–31. https://doi.org/10.1006/jcss.1995.1003

arXiv 1995
[4]

Hartline, Daniel W

Jason D. Hartline, Daniel W. Linna, Liren Shan, and Alex Tang. 2022. Algorithmic Learning Foundations for Common Law. In Proceedings of the 2022 Symposium on Computer Science and Law (Washington DC, USA) (CSLA W ’22). Association for Computing Machinery, New York, NY, USA, 109–117. https://doi.org/10.1145/3511265.3550438

arXiv 2022
[5]

Ji Liu and Xiaojin Zhu. 2016. The teaching dimension of linear learners. J. Mach. Learn. Res. 17, 1 (jan 2016), 5631–5655

work page 2016
[6]

Ejan Mackaay and Pierre Robillard. 1974. Predicting judicial decisions: The nearest neighbour rule and visual representation of case patterns . De Gruyter, Berlin, Boston, 302–331. https://doi.org/doi:10.1515/9783112320594-012

work page doi:10.1515/9783112320594-012 1974
[7]

N Sauer. 1972. On the density of families of sets. Journal of Combinatorial Theory, Series A 13, 1 (1972), 145–147. https://doi.org/10.1016/0097- 3165(72)90019-2

doi:10.1016/0097- 1972
[8]

Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein

Ali Shafahi, W. Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. 2018. Poison frogs! targeted clean-label poisoning attacks on neural networks. In Proceedings of the 32nd International Conference on Neural Information Processing Systems (Montréal, Canada) (NIPS’18). Curran Associates Inc., Red Hook, NY, USA,...

work page 2018

Pith tools

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

[1] [1]

Avrim Blum, Steve Hanneke, Jian Qian, and Han Shao. 2021. Robust learning under clean-label attack. In Conference on Learning Theory . PMLR, 591–634

work page 2021

[2] [2]

L Karl Branting, Craig Pfeifer, Bradford Brown, Lisa Ferro, John Aberdeen, Brandy Weiss, Mark Pfaff, and Bill Liao. 2021. Scalable and explainable legal prediction. Artificial Intelligence and Law 29 (2021), 213–238

work page 2021

[3] [3]

Goldman and M.J

S.A. Goldman and M.J. Kearns. 1995. On the Complexity of Teaching.J. Comput. System Sci. 50, 1 (1995), 20–31. https://doi.org/10.1006/jcss.1995.1003

arXiv 1995

[4] [4]

Hartline, Daniel W

Jason D. Hartline, Daniel W. Linna, Liren Shan, and Alex Tang. 2022. Algorithmic Learning Foundations for Common Law. In Proceedings of the 2022 Symposium on Computer Science and Law (Washington DC, USA) (CSLA W ’22). Association for Computing Machinery, New York, NY, USA, 109–117. https://doi.org/10.1145/3511265.3550438

arXiv 2022

[5] [5]

Ji Liu and Xiaojin Zhu. 2016. The teaching dimension of linear learners. J. Mach. Learn. Res. 17, 1 (jan 2016), 5631–5655

work page 2016

[6] [6]

Ejan Mackaay and Pierre Robillard. 1974. Predicting judicial decisions: The nearest neighbour rule and visual representation of case patterns . De Gruyter, Berlin, Boston, 302–331. https://doi.org/doi:10.1515/9783112320594-012

work page doi:10.1515/9783112320594-012 1974

[7] [7]

N Sauer. 1972. On the density of families of sets. Journal of Combinatorial Theory, Series A 13, 1 (1972), 145–147. https://doi.org/10.1016/0097- 3165(72)90019-2

doi:10.1016/0097- 1972

[8] [8]

Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein

Ali Shafahi, W. Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. 2018. Poison frogs! targeted clean-label poisoning attacks on neural networks. In Proceedings of the 32nd International Conference on Neural Information Processing Systems (Montréal, Canada) (NIPS’18). Curran Associates Inc., Red Hook, NY, USA,...

work page 2018