{"id":"6518ef52-f4ea-4581-b0c5-434cac2c5f1a","arxiv_id":"1908.02988","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The leftmost leaves cake-cutting mechanism is proportional and not-obviously manipulable, making a weak truthfulness guarantee compatible with fairness.","lead":"This paper shows that a weak form of manipulation resistance, called non-obvious manipulability, can be combined with fair cake division, contradicting the usual idea that fairness forces manipulability. A simple sequential cutting procedure, leftmost leaves, achieves both properties, and the result helps explain why people manipulate some cake-cutting rules less than others.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Leftmost leaves is specified with a cut query whose parameter is the agent's private remaining value, which is not a valid Robertson–Webb query; Theorem 2's scope is undefined until the protocol is re-specified with explicit eval queries.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the leftmost leaves protocol, as written, uses a cut query whose parameter is the agent's own private remaining value, which is not a legitimate Robertson–Webb cut query under the paper's own definition in Section 3.2. This is not a mere notational quibble: the paper's framework and its comparison with Brânzei–Miltersen depend on the RW model, and Theorem 2 claims existence of a mechanism in that model. If the protocol is interpreted as a relative-cut query, it lies outside RW; if it is interpreted as first eliciting the remaining value, that elicitation step is omitted and the manipulation space changes. The paper's own Remark 1 demonstrates that replacing the private-value parameter with a public constant destroys NOM, so one cannot silently reinterpret the query without re-examining the proof. I checked the rest of the argument in good faith: the proof of the worst-case bound in Lemma 1 is informal and has some quantifier sloppiness around conditioning on histories, but it is plausibly repairable; the best-case bound in Lemma 2 is essentially vacuous because a truthful agent can already obtain utility approaching the whole cake, so condition (2) is automatic; the direct-revelation results in Lemmas 3–4 and Theorem 3 are standard and appear correct. The experimental evidence is consistent with NOM but is not load-bearing for the main theorem. Thus the central mathematical claim is likely correct under a natural re-specification, but the paper as written has a genuine formal gap in the statement of the mechanism. The reader's CONDITIONAL verdict is therefore appropriate; my stress-test does not move that verdict, so I recommend UNCHANGED.","tokens_in":1130,"tokens_out":993,"duration_ms":219843,"concrete_test":"Re-derive Theorem 2 for a fully specified Robertson–Webb extensive form in which each period t begins with eval(i; x_{t-1}, 1) for every remaining agent, followed by cut(i; x_{t-1}, v_i/(n-t+1)) using the reported value v_i. Then verify that Lemmas 1 and 2 and the manipulation argument in Section 5 remain valid when the agent may lie about the eval answer, so that the cut parameter is partially under the agent's control. If the NOM proof cannot be adapted, the theorem as stated is not about any well-defined RW mechanism. If it can, the paper should be amended to include the explicit eval queries and the re-verified proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4, the leftmost leaves mechanism asks, at each period t, the query cut(i; x_{t-1}, u_i([x_{t-1},1];θ_i)/(n-t+1)). The Robertson–Webb model in Section 3.2 defines a cut query as cut(i; x, α) with α a known number in [0,1]; the mechanism cannot use the agent's true remaining value u_i([x_{t-1},1];θ_i) as α because θ_i is private and has not been elicited. If the intended query is instead a relative-cut query ('cut so that the piece has value 1/(n-t+1) of your remaining value'), that query is outside the stated RW model. The paper never supplies an eval query that would make this parameter publicly available. This matters because Lemmas 1 and 2 and the manipulation proof in Section 5 all rely on the relative-cut form, and Remark 1 shows that the query choice is substantive: replacing the parameter with the constant 1/n destroys NOM. As written, the mechanism in Theorem 2 is not a well-defined object of the framework the paper explicitly adopts, so the exact scope of the central existence claim is ambiguous. The issue is fixable by inserting eval(i; x_{t-1}, 1) before each cut and using the reported value in the cut parameter, but that re-specification changes the manipulation space and must be checked afresh. This is the most load-bearing concern because it sits at the boundary between 'the theorem is true under a clear intended interpretation' and 'the theorem is about a protocol that cannot be expressed in the declared query model.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends the notion of not-obvious manipulability (NOM) from direct-revelation mechanisms to indirect cake-cutting mechanisms. It introduces the leftmost leaves mechanism, a simultaneous variant of the Dubins-Spanier moving knife, and claims in Theorem 2 that this mechanism is proportional and NOM. It further claims in Theorem 1 that several classical mechanisms, including cut-and-choose, cut-middle, last diminisher, and the original moving knife, are obviously manipulable. Finally, Section 6 argues that a Nash-optimal direct-revelation mechanism is NOM, envy-free, and Pareto-optimal, and Section 7 cites experimental evidence connecting NOM to observed manipulation rates.","tokens_in":12101,"tokens_out":24827,"duration_ms":279524,"significance":"If the main theorem is correct, the paper makes a meaningful contribution by showing that a weak incentive property, NOM, is compatible with the strong fairness guarantee of proportionality in cake-cutting, using a simple connected mechanism. The paper also offers a plausible explanation for observed differences in manipulation rates across cake-cutting procedures. The direct-revelation result in Section 6 is a nice application of the framework. However, the formal specification of the main mechanism is not valid within the Robertson-Webb model as defined, and the treatment of the moving knife in Theorem 1 appears to be incorrect; these issues must be resolved before the paper's claims can be relied upon.","major_comments":[{"comment":"The leftmost leaves mechanism is not well-defined in the Robertson-Webb model. The protocol asks cut(i; x_{t-1}, u_i([x_{t-1},1];θ_i)/(n-t+1)), but Section 3.2 defines a cut query cut(i; x, α) with α ∈ [0,1] a number known to the mechanism, not a quantity that depends on agent i's private type. Since θ_i has not been elicited at that point, the mechanism cannot compute the query parameter. The proofs of Lemmas 1 and 2 and the manipulation argument for Theorem 2 all rely on this relative-cut form, and Remark 1 shows that the choice of query is substantive. The protocol should be re-specified, for instance by asking eval(i; x_{t-1}, 1) before each cut and using the reported value in the cut parameter, and the NOM proof must then be re-checked to account for manipulations of the eval reports as well as the cut reports.","section":"Section 4"},{"comment":"Lemma 1 is not stated precisely enough to be a valid equality. The right-hand side, u_i([x_{t-1},1];θ_i)/(n-t+1), depends on θ_{-i} through x_{t-1}, while the left-hand side is an infimum over θ_{-i}. As written, the lemma asserts that an infimum equals a quantity that varies with the variable being integrated out. The statement should be conditional on the realized history (or on the realized previous cut points), with the infimum taken over continuations from period t onward. The same issue affects Lemma 2. The proof of Theorem 2 needs this clarification to be formally coherent.","section":"Section 5, Lemma 1"},{"comment":"The proof that the Dubins-Spanier moving knife is obviously manipulable is internally inconsistent. The text says that in the best case the other agent cuts at ε, giving the truthful agent almost 1, and then says that a truthful agent gets 0.5 with certainty; these two statements cannot both hold, since if the other agent cuts at ε, the truthful agent does not stop the knife at his fair-share point. Moreover, the proposed manipulation of stopping at 1−ε does not improve the best case, which is already arbitrarily close to 1, and it does not improve the worst case: other agents can stop just before 1−ε and leave the manipulator with only a small remaining piece. For n=2, the truthful cutoff at 1/2 is exactly the maximin strategy, so inequality (1) is not violated. Consequently, the claim that the moving knife is obviously manipulable is not established and appears to be false. This claim should be corrected or removed, and the abstract and related-literature statements that depend on it should be adjusted.","section":"Theorem 1, Dubins-Spanier moving knife"}],"minor_comments":[{"comment":"The name 'Morrill' is misspelled as 'Morril' in the abstract and in the acknowledgments; the correct spelling appears elsewhere in the text.","section":"Abstract and Acknowledgments"},{"comment":"The phrase 'E(θ) is an edge cover' is nonstandard: an edge cover usually means a set of edges that touches every vertex, whereas here the intended meaning is a selection of one outgoing edge at every node. Please use different terminology to avoid confusion.","section":"Section 3.2"},{"comment":"The proof of Lemma 1 compresses an induction over periods into a single sentence. Expanding the induction explicitly, especially the 'otherwise' case and the recursive step, would improve readability and make the argument easier to verify.","section":"Section 5, Lemma 1 proof"},{"comment":"Theorem 1 relies on Figures 2 and 3, but these figures are not included in the arXiv text. The piecewise-uniform constructions used in the proof are therefore difficult to verify; please ensure that the figures are available or add a textual description of the valuations.","section":"Section 1 and Theorem 1"},{"comment":"The text says that a Nash-optimal allocation is 'hence also proportional' after noting it is Pareto-optimal and envy-free; since envy-freeness implies proportionality for additive utilities, this is correct, but the one-line justification would be helpful for readers.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"The core construction of leftmost leaves is appealing and likely salvageable, but the paper currently fails to specify the mechanism within its own formal query model. The Dubins-Spanier claim in Theorem 1 appears to be wrong and, as a formal theorem, needs to be corrected. I do not see the issues as fatal to the main result, but they require a substantive revision, not just copy-editing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is the straightforward take. The core idea is right: NOM is compatible with proportionality, and the leftmost leaves mechanism is the witness. That is a real conceptual contribution, and the contrast with the classical proportional mechanisms (Theorem 1) plus the direct-revelation Nash-optimal result (Theorem 3) round it out.\n\nThe paper does well in identifying exactly which query choice is responsible for the NOM property: Remark 1, showing that switching from a relative-cut to an absolute-cut parameter destroys NOM, is the best part of the paper. The proof strategy—separate inductions for worst-case and best-case utility—is sound under the intended reading of the protocol.\n\nThe soft spot is formal and it is real. The mechanism in Section 4 asks agents for cut(i; x_{t-1}, u_i([x_{t-1},1];θ_i)/(n-t+1)). In the Robertson-Webb model, α in a cut query must be a publicly known number; the agent's private remaining value is not known to the mechanism unless it first asks an eval query. No such eval query is specified. Because Remark 1 shows the query parameter is load-bearing, the theorem as written is about a protocol that is not well-defined in the model the paper claims to be using. This is fixable: insert an explicit eval(i; x_{t-1}, 1) before each cut and use the reported value in the cut parameter. But that changes the manipulation space—agents can lie about the eval—and the lemmas need to be rechecked under that richer game.\n\nThe proof of Lemma 1 also contains a slightly hand-wavy step ('a recursive formulation shows...'), but the induction is convincing enough. The NOM definition's extension to indirect mechanisms is conceptually fine, though the wording around 'profitable manipulation' could be tightened.\n\nOn the experimental side, the paper cites the authors' own lab study as supporting evidence. That's acceptable as illustration but doesn't carry the load; the honest discussion of the Selfridge-Conway anomaly is to the authors' credit.\n\nWho is this for? Fair division specialists and mechanism design people working on bounded rationality and prior-free guarantees. It deserves serious peer review. The gap is not a dead end; it is a specification problem that a good referee can help the authors fix. I would accept the paper for review, with the expectation of substantive revision on the protocol definition.","headline":"Genuinely new result with a fixable formal gap in the query specification; deserves peer review.","tokens_in":12591,"tokens_out":6085,"would_cite":true,"duration_ms":66130,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B32"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that the leftmost leaves cake-cutting mechanism gives every agent a fair share and leaves no obvious way to gain by lying.","keywords":["cake-cutting","not-obvious manipulability","proportionality","leftmost leaves mechanism","moving knife","fair division","strategy-proofness","mechanism design"],"falsifier":"A counterexample would be a non-atomic valuation profile and a misreport for which, under the relative-cut leftmost leaves protocol, a truthful agent's worst-case utility is strictly below what the manipulation guarantees, or the manipulation's best-case utility exceeds the value of the remaining cake; an exhaustive search over piecewise-constant valuations would find such a profile if the theorem is false.","tokens_in":11534,"feed_emoji":"🍰","tokens_out":10480,"duration_ms":98240,"temperature":0.7,"pith_summary":"This paper asks whether a cake can be divided fairly without giving agents an obvious reason to lie. Strategy-proofness, the guarantee that truth-telling is always best, is known to be incompatible with even minimal fairness in cake-cutting: with two agents it forces a dictatorship, and with three or more it leaves one agent with nothing. The authors show that a weaker standard, non-obvious manipulability (NOM), is compatible with proportionality, the guarantee that every truthful agent receives $1/n$ of the cake. Their vehicle is the leftmost leaves mechanism, a discrete simultaneous variant of the classical moving-knife procedure, which they prove is proportional and NOM. The paper also shows that most classical proportional mechanisms, including the original moving knife, fail NOM, which offers an explanation for why leftmost leaves is manipulated less often in experiments.","feed_headline":"This cake-cutting method is fair and can't be obviously gamed","feed_subtitle":"A moving-knife variant guarantees every agent 1/n and blocks lies that improve best or worst case.","key_machinery":"The key object is the leftmost leaves mechanism, a discrete variant of the moving-knife procedure in which all remaining agents cut simultaneously and the smallest cut leaves with the interval to its left. Its defining feature is the relative cut query: at period $t$ each agent is asked for the point $x$ such that $u_i([x_{t-1},x])=u_i([x_{t-1},1])/(n-t+1)$, the point that gives the agent a fair share of the cake that remains. This query does the work: Lemma 1 pins the truthful worst-case utility to exactly that fair share, Lemma 2 pins the truthful best-case utility to the whole remaining cake, and comparing these two extremes rules out every obvious manipulation. The relative form is essential, as Remark 1 demonstrates.","core_discovery":"The central claim is Theorem 2: the leftmost leaves mechanism is proportional and not-obviously-manipulable. In each round, every remaining agent marks the point from the current left boundary where the value equals the value of the remaining cake divided by the number of remaining agents, and the smallest such cut wins that interval and leaves. The proof shows that a truthful agent's worst-case utility is exactly an equal share of the remaining cake, and that in the best case the truth-teller can get arbitrarily close to the whole remaining cake, so no misreport can improve either extreme. The authors also prove that cut-and-choose, cut-middle, the last-diminisher procedure, and the original moving-knife mechanism are all obviously manipulable, and that replacing the relative cut query with an absolute $1/n$ cut makes leftmost leaves itself obviously manipulable.","pith_inferences":["The relative cut query is likely the general recipe: to make a sequential allocation protocol NOM, each agent's action should reference his own residual value of the remaining cake rather than a fixed fraction of the original cake.","NOM only constrains the two extreme scenarios, so a mechanism satisfying it can still be manipulated when agents hold intermediate beliefs; a testable prediction is that leftmost leaves will be manipulated more when agents have informative priors that make a misreport attractive.","The same simultaneous-relative-cut idea might transfer to other fair-division settings, such as allocating indivisible goods with cardinal utilities or rent division, as long as agents can report indifference points relative to their own residual value."],"forward_implications":["If Theorem 2 is correct, there is a deterministic, connected, procedurally fair cake-cutting mechanism that guarantees a truthful agent at least $1/n$ of the cake and admits no manipulation that improves the agent's worst-case or best-case utility.","The classical proportional mechanisms listed in Theorem 1 are all obviously manipulable, so NOM separates leftmost leaves from these allocation-equivalent procedures.","The choice of query matters: replacing the relative cut query with the absolute cut query $cut(i; x_{t-1}, 1/n)$ produces a variant that is obviously manipulable.","In direct-revelation mechanisms, proportionality alone guarantees the worst-case NOM condition and Pareto-optimality alone guarantees the best-case condition, so the Nash-optimal mechanism is NOM, envy-free, and Pareto-optimal, though it may return disconnected pieces."],"supporting_citations":[{"why":"Supplies the moving-knife procedure that leftmost leaves adapts, and the baseline that Theorem 1 shows is obviously manipulable.","marker":"Dubins and Spanier (1961)"},{"why":"Defines non-obvious manipulability, the solution concept the paper extends to indirect mechanisms.","marker":"Troyan and Morrill (2019)"},{"why":"Proves the strategy-proofness impossibility that motivates weakening the incentive requirement.","marker":"Brânzei and Miltersen (2015)"},{"why":"Defines maximin strategy-proofness, the weaker worst-case property that NOM strengthens with a best-case condition.","marker":"Brams et al. (2006)"},{"why":"Provides the experimental evidence that NOM mechanisms, including leftmost leaves, are manipulated less in practice.","marker":"Kyropoulou et al. (2019)"},{"why":"Shows that theoretically equivalent mechanisms can differ in incentive properties for boundedly rational agents, used to explain the leftmost leaves versus moving-knife difference.","marker":"Li (2017)"}],"fun_headline_variants":["Fair cake, no obvious gaming","Cake-cutting resists obvious manipulation","Proportional cake cut that's not obviously gameable","Leftmost leaves: proportional, non-obviously manipulable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof depends on the mechanism asking each agent to cut at a point computed from the agent's own private value of the remaining cake, not at a pre-specified fraction; if this relative cut query cannot be implemented with the standard cut and eval queries, the exact scope of the theorem is undefined.","fun_headline_variants_meta":{"raw":{"variants":["Fair cake, no obvious gaming","Cake-cutting resists obvious manipulation","Proportional cake cut that's not obviously gameable","Leftmost leaves: proportional, non-obviously manipulable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001812,"raw_usage":{"total_tokens":7078,"prompt_tokens":838,"completion_tokens":6240,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":6183}},"tokens_in":454,"tokens_out":6240,"duration_ms":51402,"temperature":1.0,"reasoning_tokens":6183,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:29:54.594819+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A counterexample would be a non-atomic valuation profile and a misreport for which, under the relative-cut leftmost leaves protocol, a truthful agent's worst-case utility is strictly below what the manipulation guarantees, or the manipulation's best-case utility exceeds the value of the remaining cake; an exhaustive search over piecewise-constant valuations would find such a profile if the theorem is false.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the moving-knife procedure that leftmost leaves adapts, and the baseline that Theorem 1 shows is obviously manipulable."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines non-obvious manipulability, the solution concept the paper extends to indirect mechanisms."},{"cited_title":"Jones, and C","cited_arxiv_id":null,"evidence_quote":"Defines maximin strategy-proofness, the weaker worst-case property that NOM strengthens with a best-case condition."},{"cited_title":"Ortega, and E","cited_arxiv_id":null,"evidence_quote":"Provides the experimental evidence that NOM mechanisms, including leftmost leaves, are manipulated less in practice."}],"review_version":1}