{"id":"59572454-bb11-44f7-b384-4e3f9ccf1d39","arxiv_id":"2412.18138","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper argues that less discriminatory algorithms cannot be defined by quantitative performance metrics alone and must incorporate a reasonableness standard, with feasible but computationally hard search problems.","lead":"This paper analyzes the legal concept of a less discriminatory algorithm (LDA) and argues that any formal definition must depend on judgments about whether an alternative model would reasonably generalize, not just on measurements of its accuracy and disparity on data. It also proves mathematical and computational limits on searching for such algorithms, while presenting evidence that practical searches can still work.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Definition 1's 'reasonableness standard' is never operationalized; without it, Proposition 1 only rules out naive measured-performance definitions, so the paper's positive account of what constitutes an LDA is unsupported.","rationale":"I agree with the reader's weakest-assumption diagnosis. The paper's formal results are mostly sound: Proposition 1 is correct, and the NP-hardness reduction in Appendix C.2 is detailed. But the constructive part of the paper—Definition 1 and the argument that courts can evaluate LDAs by 'projecting' performance—rests entirely on an unformalized reasonableness standard. The paper explicitly acknowledges this in Section 2.2, so the concern is not manufactured; it is a self-identified gap. This is load-bearing because the paper's central claim is not merely that naive quantitative definitions fail (which Proposition 1 establishes), but that a reasonableness-based definition is the right alternative. Without an account of what makes a projection reasonable, the definition cannot be applied, tested, or falsified; 'significantly lower' adds a second indeterminate term. The appropriate response is to keep the paper conditional: either narrow the contribution to the negative result, or provide a minimal operationalization. I would not reject the paper, since the negative results and the empirical demonstrations have independent value, and a fully general reasonableness standard may indeed be a legal rather than a formal question. But the current positive framework is incomplete.","tokens_in":22983,"tokens_out":7928,"duration_ms":73737,"concrete_test":"Ask the authors to specify a minimal formal 'reasonable projection' predicate R(h', h0, D_pre, D_post) that (a) excludes the Proposition 1 pathological rule and (b) admits the random-search alternatives reported as successful in Section 5. Then determine whether R can be expressed without using a held-out dataset, numerical thresholds on accuracy/disparity gaps, or model-class complexity. If every such R remains essentially quantitative, or if none satisfies both (a) and (b), the central dichotomy between quantitative and reasonableness-based LDA definitions is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Definition 1 (Section 2.1) is the paper's central formal proposal: h' is an LDA if it is 'reasonably projected' to have accuracy at least as high as h0 and 'significantly lower' disparity. Both predicates are undefined. Section 2.2 concedes that formalizing reasonable projections 'in full generality is beyond the scope this work.' This is not a peripheral caveat: the entire argument that LDA definitions cannot be purely quantitative is cashed out through this reasonableness predicate. Proposition 1 shows that a purely measured-performance definition admits a pathological rule; but Definition 1 excludes that rule only by invoking an unanalyzed notion of reasonable projection. Without an operational criterion, a court has no way to distinguish the Proposition 1 rule from a legitimate multiplicity-based alternative, and any proposed quantitative surrogate (complexity penalty, holdout-style evaluation, disparity threshold) reintroduces the purely quantitative definition the paper claims is impossible. The paper may establish a negative thesis (measured performance alone is insufficient), but it does not establish the positive thesis that a reasonableness-based definition can be meaningfully applied; 'significantly lower' also lacks any threshold, so the LDA predicate is indeterminate even after projection is specified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper examines how to define a \"less discriminatory algorithm\" (LDA) in the legal disparate-impact sense. It argues that LDA definitions based on measured accuracy and disparity on observed datasets are untenable: Proposition 1 constructs a decision rule that achieves perfect accuracy on pre-deployment data and zero disparity on post-deployment data, and the authors contend that any purely quantitative measured-performance definition admits similar pathologies. The paper therefore proposes Definition 1, under which a model is an LDA if it is \"reasonably projected\" to have at least the accuracy of the baseline and \"significantly lower\" selection-rate disparity, while acknowledging that the projection standard is not fully formalized. It then analyzes the feasibility of finding LDAs: Theorem 1 characterizes the attainable utility-disparity frontier under full information and shows a zero-disparity alternative exists unless the baseline utility is above a threshold; Theorem 2 proves the full-information LDA existence problem NP-complete via a reduction from subset sum, with an accompanying (1+epsilon)-approximation claim; and Section 5 presents empirical simulations showing that simple multiplicity-based searches can reduce out-of-sample disparity on the Adult dataset, though not on German Credit.","tokens_in":23263,"tokens_out":8387,"duration_ms":78152,"significance":"If the results hold, the paper makes three useful contributions: it gives a crisp negative result showing why measured-performance LDA definitions are insufficient, it provides a formal characterization of when accuracy/disparity trade-offs do and do not bind, and it offers evidence that practical LDA searches are often computationally feasible despite worst-case hardness. The NP-completeness reduction and the empirical open-source evaluation are concrete and reproducible. The main caveat is that the paper's positive definition rests on an unspecified \"reasonableness\" predicate, so the conceptual core is not yet fully operational; the paper's significance is therefore stronger as a negative/conceptual result than as a complete formal framework.","major_comments":[{"comment":"The definition's central predicates, \"reasonably projected\" and \"significantly lower,\" are never given a formal or operational meaning, and Section 2.2 explicitly defers their full formalization as beyond the scope of the work. This is load-bearing: Proposition 1 only rules out purely measured-performance definitions, and without an operational account of reasonable projection, a court (or a firm) has no criterion for distinguishing the pathological Proposition 1 rule from a legitimate multiplicity-based alternative. The paper may establish a negative thesis, but the positive account of what constitutes an LDA is not yet established. The authors should either provide a formal operationalization for at least a restricted but nontrivial setting, or explicitly reframe the contribution as a diagnosis plus a research agenda.","section":"Section 2.1, Definition 1"},{"comment":"The claim that a (1+epsilon)-approximate full-information LDA can be identified in polynomial time O(n^3 epsilon^{-1}) is stated without proof: Appendix C.3 says \"the proof is deferred\" and points only to standard subset-sum approximation schemes. This claim is load-bearing for the paper's conclusion that the NP-hardness result is \"weak\" and unlikely to be prohibitive in practice. As written, the reader cannot verify the central \"weak computational limits\" argument. A complete proof or a clear designation of the claim as conjectural is needed.","section":"Section 4 / Appendix C.3, Claim 5"},{"comment":"The second displayed formula in Theorem 1 is garbled: \"Delta(h') = 1 - min[ n1/n+, lambda n2/n- ] Delta(h*) - Delta(h0)\" is missing parentheses and at least one factor, so the claimed characterization of the minimum-disparity alternative at a given utility level cannot be verified from the statement as printed. The appendix proof may be correct, but the theorem statement must be restated cleanly before the feasibility result can be used as the paper intends.","section":"Section 3.2, Theorem 1"}],"minor_comments":[{"comment":"There is a typo in the first paragraph: \"representative dastaset\" should be \"representative dataset.\"","section":"Section 2.2"},{"comment":"The arithmetic derivation of the condition on alpha contains a garbled line with missing parentheses and a dropped factor (the line beginning \"-lambda 1/N ...\"), which makes the reduction harder to follow; the derivation should be rewritten cleanly.","section":"Appendix C.2"},{"comment":"The column \"Freq. min-disp.\" is not defined in the caption or the surrounding text; the caption should explain what this frequency measures.","section":"Table 1"},{"comment":"Reference [54] contains a typo: \"uidance\" should be \"Guidance.\"","section":"References"},{"comment":"The sentence \"These results imply that it may not be a good idea to punish firms for having considered and rejected models that are ultimately less discriminatory\" goes beyond what the experimental simulation can support; the claim should be softened to a hypothesis or explicitly labeled as a policy interpretation.","section":"Section 5"},{"comment":"Definition 2 lists the LDA input as <X, sigma, rho_g, h0> but the utility function U(h; lambda) is also part of the problem instance; the tuple should include lambda or the text should state that lambda is fixed.","section":"Definition 2 vs. Theorem 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a good fit for CS&Law and the conceptual message is timely. My main concern is the gap between the negative results and the positive framework: the reasonableness-based definition is not operationalized, and a key algorithmic claim in the computational-limits section is unproved. I would encourage the editors to require the authors to either provide a formal partial operationalization or explicitly reframe the contribution as a diagnosis plus an agenda, and to supply the missing approximation proof."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe thing to know: this paper earns its place in the debate. The NP-hardness result for the full-information LDA problem (Theorem 2) appears new, and the measured-vs-projected performance framing is genuinely useful. The authors are right that any purely data-measured definition can be gamed by a pathological rule, and that courts therefore need some notion of \"reasonable projection.\" The feasible-set characterization (Theorem 1) is also a clean, nice result.\n\nWhere it gets soft: the positive definition is under-specified. Definition 1 says an LDA must be \"reasonably projected\" to meet accuracy and \"significantly lower\" disparity, but neither predicate is formalized, and the authors concede this in Section 2.2. That is not a small omission. The entire argument against purely quantitative definitions is cashed out through that unanalyzed standard. The stress-test note is right: without an operational criterion, a court has no principled way to distinguish the pathological rule in Proposition 1 from a legitimate multiplicity-based alternative. I would phrase it less harshly: the paper's negative thesis is solid, but its positive thesis—that a reasonableness standard can be meaningfully applied—is a sketch, not a working definition.\n\nSecond soft spot: the (1+epsilon)-approximation claim in Section C.3 is stated with the proof deferred. That is load-bearing for the \"weak hardness\" argument, and deferring it to an exercise is unsatisfying. The empirical demonstration in C.4 is suggestive, but not a proof. A referee should ask for the proof or a clear citation to a known result.\n\nThe empirical section is appropriately modest. Two datasets, simple heuristics; the results are plausible and honestly reported. That is not a weakness worth dwelling on.\n\nWho this is for: people working at the intersection of algorithmic fairness and anti-discrimination law, and theorists interested in the complexity of fairness-constrained search. It deserves peer review; it will provoke useful discussion even if the reasonableness standard needs much more work.\n\nMy recommendation: send it out, with a request that the authors either sharpen Definition 1 or clearly scope it as a proposal that requires legal judgment to operationalize. Also get the approximation proof on the table.","headline":"A genuinely useful formal analysis of LDA search, with a central 'reasonableness' predicate that remains a placeholder.","tokens_in":23734,"tokens_out":2416,"would_cite":true,"duration_ms":21905,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that no data-only measurement can define a less discriminatory algorithm, and that courts must instead judge whether a proposed model would reasonably generalize before accepting it as an alternative to the status quo.","keywords":["less discriminatory alternative","disparate impact doctrine","algorithmic fairness","model multiplicity","selection rate disparity","reasonableness standard","NP-hardness","fairness-utility trade-off"],"falsifier":"Try to construct the Proposition 1 rule inside a restricted hypothesis class, for instance shallow decision trees of a fixed depth, on the Adult and German Credit datasets; if no such rule can simultaneously achieve perfect training accuracy and zero selection-rate disparity across groups, the impossibility result depends on unbounded model complexity rather than on the absence of held-out data, and a quantitative standard might suffice.","tokens_in":22792,"feed_emoji":"⚖️","tokens_out":8468,"duration_ms":82642,"temperature":0.7,"pith_summary":"This paper asks what should count as a less discriminatory algorithm (LDA) in disparate-impact law, a question regulators now press on firms and plaintiffs. Its central claim is that no purely quantitative definition can work: because courts lack held-out data, a plaintiff can always construct a decision rule that is perfectly accurate on the firm's training data and has zero selection-rate disparity on post-deployment data, so measured accuracy and disparity alone cannot separate a genuine LDA from an overfit artifact. The paper therefore defines an LDA as a model that is reasonably projected to be at least as accurate as the baseline and to have significantly lower disparity, and it argues that courts must apply a reasonableness standard for generalization. It then shows the mathematical limits on achieving zero disparity are real but rarely binding, that finding the least discriminatory alternative is NP-hard yet a close approximation is efficiently findable, and that simple randomized searches can reduce disparity out-of-sample, sometimes at no accuracy cost. The payoff, if correct, is a framework in which firms can proactively search for fairer models and plaintiffs can challenge failures to do so without pretending that data alone can settle the question.","feed_headline":"Fairer algorithms can't be judged by data alone","feed_subtitle":"Data-only tests of fairer algorithms can be gamed; courts must judge generalization instead","key_machinery":"The argument runs on two engines. The first is the feasible utility-disparity polygon: for a finite population with group sizes and base rates, every binary classifier corresponds to a point in the $(\\Delta, U)$ plane with $U(h;\\lambda) = \\mathrm{TPR} - \\lambda\\mathrm{FPR}$, and the Pareto frontier is traced by two types of label swaps away from the perfect classifier; this pins down the threshold $U^*$ in Theorem 1. The second is the reasonableness standard in Definition 1, which replaces measurement with projection: because no held-out data exists at litigation time, a candidate LDA must be judged by whether a court would reasonably expect it to generalize, with model complexity relative to the baseline as the main heuristic. On top of these, the paper's NP-hardness result is driven by a reduction from subset sum, while the positive approximation guarantee comes from a polynomial-time approximation scheme for the same knapsack-like problem.","core_discovery":"On the paper's own terms, the discovery is that the LDA concept must move from measured performance to projected performance. Proposition 1 constructs a rule that is perfectly accurate on one dataset and selects a constant fraction of each group on another, showing that any definition based solely on observed accuracy and disparity can be satisfied by a pathological model. The paper's formal definition therefore says $h'$ is an LDA relative to $h_0$ when $h'$ is reasonably projected to have at least $h_0$'s accuracy and significantly lower selection-rate disparity; what makes the projection reasonable is left to a case-by-case judgment, not to a formula. The accompanying results give the boundaries of what is achievable: a utility threshold $U^*$ below which a zero-disparity alternative always exists even when base rates differ, an NP-completeness result for the full-information LDA search, a polynomial-time $(1+\\epsilon)$-approximation that guarantees finding a significantly better model whenever the baseline is not already near-optimal, and empirical evidence on common datasets that simple random-seed or resampling searches within one model class reduce out-of-sample disparity, sometimes with a utility gain.","pith_inferences":["The paper leaves implicit that the same gameability argument would apply to other fairness metrics, not just selection-rate disparity, so any purely data-based fairness definition in litigation may need an accompanying reasonableness standard.","Outside the courtroom, the reasonableness standard could be made operational by benchmarking generalization through temporal or cross-domain shifts, turning an opaque legal judgment into a measurable inductive question.","The approximation guarantee is proved for the full-information setting, where the firm knows the true population distribution; converting it into practical training algorithms for finite samples and restricted model classes is a natural next step the paper leaves open."],"forward_implications":["Courts should evaluate candidate LDAs by whether they would be expected to generalize, not by how they score on observed data, which means post-deployment measurements alone cannot establish liability.","When group base rates differ, perfect accuracy necessarily has nonzero disparity, but a zero-disparity alternative exists whenever the baseline's utility is below $U^*$; the paper's reading of Theorem 1 is that this trade-off is vacuous except at unusually high accuracy, so it rarely shields a firm.","Since the least discriminatory alternative is NP-hard to find but a $(1+\\epsilon)$-approximation is computable in polynomial time, a firm's claim that the search is computationally impossible is a weak defense as long as the legal standard permits de minimis slack.","Simple searches by random seed or resampling within the same model class can reduce disparity on unseen data, sometimes improving utility as well, so firms need not rely on specialized optimization to find reasonable LDAs.","Because a good-faith search often turns up and rejects models that later prove less discriminatory out-of-sample, treating considered-and-rejected models as evidence of liability would discourage proactive searching."],"supporting_citations":[{"why":"Supplies the legal framing of the less discriminatory alternative and the argument that firms should bear the burden of a reasonable search.","marker":"[7]"},{"why":"Provides the prior formalization of an LDA as an optimization over a fixed dataset that this paper extends and critiques.","marker":"[29]"},{"why":"Introduces D-hacking, the concern that fairness improvements measured on training data may not generalize, which motivates the projected-performance definition.","marker":"[8]"},{"why":"Establishes model multiplicity, the phenomenon that many equally accurate models make different predictions, which underpins the empirical search strategies.","marker":"[6]"},{"why":"Supplies the statistical-learning rationale that simpler models are more likely to generalize, the heuristic content of the reasonableness standard.","marker":"[68]"},{"why":"Provides the subset-sum approximation machinery used to prove that a near-least-discriminatory alternative can be found in polynomial time.","marker":"[38]"},{"why":"Raises the prior belief that finding the least discriminatory algorithm requires an unbounded and therefore intractable search, the position the paper's Theorem 2 and approximation result address.","marker":"[66]"}],"fun_headline_variants":["Data-only fairness tests are gameable","Fairness demands reasonableness beyond data","Less discriminatory algorithms need human judgment","Projected performance defines fairer algorithms","Math alone can't define a fairer algorithm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that courts can apply a reasonableness standard well enough to judge whether a proposed model would generalize, even though the paper does not define that standard precisely and concedes that formalizing it in full generality is beyond its scope.","fun_headline_variants_meta":{"raw":{"variants":["Data-only fairness tests are gameable","Fairness demands reasonableness beyond data","Less discriminatory algorithms need human judgment","Projected performance defines fairer algorithms","Math alone can't define a fairer algorithm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1429,"prompt_tokens":1003,"completion_tokens":426,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":364}},"tokens_in":619,"tokens_out":426,"duration_ms":4674,"temperature":1.0,"reasoning_tokens":364,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:58:43.028791+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Try to construct the Proposition 1 rule inside a restricted hypothesis class, for instance shallow decision trees of a fixed depth, on the Adult and German Credit datasets; if no such rule can simultaneously achieve perfect training accuracy and zero selection-rate disparity across groups, the impossibility result depends on unbounded model complexity rather than on the absence of held-out data, and a quantitative standard might suffice.","supporting_citations":[{"cited_title":"Less discriminatory algorithms","cited_arxiv_id":null,"evidence_quote":"Supplies the legal framing of the less discriminatory alternative and the argument that firms should bear the burden of a reasonable search."},{"cited_title":"Operationalizing the search for less discrim- inatory alternatives in fair lending","cited_arxiv_id":null,"evidence_quote":"Provides the prior formalization of an LDA as an optimization over a fixed dataset that this paper extends and critiques."},{"cited_title":"D-hacking","cited_arxiv_id":null,"evidence_quote":"Introduces D-hacking, the concern that fairness improvements measured on training data may not generalize, which motivates the projected-performance definition."},{"cited_title":"Model multiplicity: Opportunities, con- cerns, and solutions","cited_arxiv_id":null,"evidence_quote":"Establishes model multiplicity, the phenomenon that many equally accurate models make different predictions, which underpins the empirical search strategies."},{"cited_title":"Algorithm design","cited_arxiv_id":null,"evidence_quote":"Provides the subset-sum approximation machinery used to prove that a near-least-discriminatory alternative can be found in polynomial time."},{"cited_title":"Applying old rules to new tools: Employment discrimination law in the age of algorithms","cited_arxiv_id":null,"evidence_quote":"Raises the prior belief that finding the least discriminatory algorithm requires an unbounded and therefore intractable search, the position the paper's Theorem 2 and approximation result address."}],"review_version":1}