{"id":"2f4508ba-ada7-4450-9465-60874c7695da","arxiv_id":"2411.19431","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"In mediated communication with transparent sender motives, allowing the sender to burn money strictly raises his equilibrium payoff for almost all priors where commitment has value, and the optimal value equals the worst-case Bayesian persuasion payoff.","lead":"A theory paper in economics shows that a sender who can burn money through a mediator can strictly improve strategic communication compared with mediation alone, whenever commitment is valuable. The paper characterizes the sender's best achievable payoff and links it to robust Bayesian persuasion and smart-contract settings.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 5's boundary-continuity proof invokes Lemma 4 on a non-full-support face; without an induction on support size, Proposition 7 and hence Theorem 1 are not fully proven.","rationale":"The paper's central claim, Theorem 1, is well motivated and the main proof architecture is coherent: an upper bound from nonnegative burning, an approximating construction, and a minimax characterization via Proposition 3. The reader's stated concerns about Proposition 10 and Lemma 5's 'all gamma' claim do not withstand scrutiny: Sion's minimax theorem allows a non-compact convex Y, and the uniqueness claim for all gamma in (0,1) follows from linearity and strict slack at mu2. The genuine weakness is the boundary-continuity gap in Lemma 5, which is load-bearing because Proposition 12's contradiction argument requires continuity on lower-dimensional faces even when the initial prior is full support. This is a proof gap rather than a demonstrated falsehood, so the verdict remains CONDITIONAL; the authors should either supply the inductive argument or restrict Proposition 7 to full-support priors if the theorem's almost-everywhere statement can be proven without boundary continuity. I did not find a counterexample to the central theorem, and the paper's self-identified limitations are consistent with the analysis.","tokens_in":32978,"tokens_out":32676,"duration_ms":298613,"concrete_test":"Provide a complete proof of Lemma 5 by induction on |supp(mu)|, with the base case |supp|=1 trivial and the induction step applied inside the face Delta(supp(mu)); then verify that the same induction yields the continuity at lower-support beliefs mu2 that Proposition 12 Case 2 requires. If the induction goes through, the proof gap is closed; if it needs an extra assumption not entailed by Definition 3, Proposition 7 is unproven as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 7 is the bridge needed for Theorem 1: under Definition 3, V*=V*_MD implies V*_CT=V*. The proof in Appendix B.3 first proves Proposition 12 under the assumption that V* is continuous, then tries to prove continuity of V* on all of Delta(Theta) via Lemma 5. In Lemma 5, for a boundary belief mu and a direction mu1 with supp(mu1) subset of supp(mu), the proof says 'by Lemma 4' we get the limit. But Lemma 4 establishes continuity only at full-support beliefs mu(theta)>0 for all theta. The perturbed prior (1-epsilon)mu + epsilon mu1 has support supp(mu), a proper face when mu is boundary, so Lemma 4 does not apply. This is not cosmetic: Proposition 12's Case 2 uses continuity at the lower-support belief mu2 to choose epsilon>0 with V*((epsilon mu1 + p2 mu2)/(epsilon+p2)) > R. The manuscript supplies no induction on support size to repair the gap. Unless such an induction or an alternative proof of boundary continuity is provided, the proof of Proposition 7, and therefore the almost-sure strict-improvement conclusion of Theorem 1, is incomplete.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces \"mediated communication with money-burning mechanisms\" (MDMB), in which a Sender with state-independent preferences commits to a mediator who sends messages and burns money contingent on the Sender's report. It develops a revelation principle for this protocol, characterizes the Sender's value as V*(µ0) = max_π min_θ V_π(θ) = min_λ cav(V̂_λ)(µ0) (Proposition 3), and compares this value with the values of cheap talk, mediated communication without burning, and Bayesian persuasion. The main result, Theorem 1, claims that under a generic condition on Receiver payoffs, for almost all priors either cheap talk already attains the maximum possible Sender payoff or money burning strictly improves mediated communication. The paper also analyzes bounded burning budgets, relates MDMB to robust/cautious Bayesian persuasion, and discusses applications to Web 3.0 smart contracts.","tokens_in":33098,"tokens_out":9313,"duration_ms":87228,"significance":"If the main characterization and Theorem 1 are correct, the paper makes a substantive theoretical contribution: it identifies a limited-commitment communication protocol whose value has a clean geometric formula, and it shows a generic strict-improvement result relative to ordinary mediation. The proof strategy is largely constructive, and the connection to cautious Bayesian persuasion and heterogeneous-prior persuasion is insightful. The paper is also commendably explicit about the scope of its conclusions: Section 7 states the transparent-motives assumption, and Appendix A.4 gives a counterexample without it. The formal apparatus, including the new revelation principle and the minimax argument, is a serious derivation rather than a list of assertions. The paper does not provide machine-checked proofs or computational verification, but the written proofs are extensive.","major_comments":[{"comment":"The first equality V*(µ0) = max_π min_{θ∈Θ} V_π(θ), together with the definition of V̂_λ in Eq. (10), is not well-posed for priors with zero-probability types. The ratio µ(θ)/µ0(θ) is undefined when µ0(θ) = 0, and for such types π(·|θ) is unconstrained by the model yet still enters the minimum over Θ. The statement should either restrict attention to supp(µ0), restrict λ to be supported on supp(µ0), or explicitly assume a full-support prior. Lemma 2, which derives the upper bound, has the same issue: incentive compatibility is only required for types with positive prior probability.","section":"Appendix B.3, Lemma 5"}],"minor_comments":[{"comment":"In the paragraph after Proposition 3, \"min-mas\" should read \"min-max\".","section":"Section 4.1"},{"comment":"The word \"Lagragian\" appears twice and should be \"Lagrangian\".","section":"Section 6.3"},{"comment":"The sentence beginning \"Because we can merge the posteriors that induce the same action...\" is repeated verbatim in consecutive paragraphs; please delete the duplicate.","section":"Appendix A.4"},{"comment":"The proof of Proposition 9 refers to \"Cauchy inequality\" without qualification; this should be \"Cauchy-Schwarz inequality\".","section":"Appendix B.4"},{"comment":"Several proofs cite \"Proposition 9 of the working paper version of Kamenica and Gentzkow [2011]\" with no formal reference in the bibliography; please provide a citable published version or state the result explicitly.","section":"Appendix B"},{"comment":"The sentence \"Technically, we provide an example in Appendix A. Example 3 shows...\" is awkwardly punctuated and could be read as referring to two different examples; please rephrase for clarity.","section":"Section 5"},{"comment":"The figure captions are minimal; please label all curves and indicate which lines correspond to which protocol or value function, as this would substantially improve readability.","section":"Figures 2-5"}],"recommendation":"major_revision","confidential_remarks":"The Lemma 5 gap is substantive but appears repairable, so I am not recommending rejection. The main theorem is interesting enough to justify a revision. The Web 3.0 discussion and the bounded-budget extension broaden the paper's scope, but the core of the contribution is the value characterization and the generic strict-improvement result; I would focus the revision on closing the boundary-continuity argument."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Good paper to know about, but not one you can trust as written. The main idea is that a sender with state-independent preferences can use report-contingent money burning to gain commitment power, and the optimal value equals a robust Bayesian persuasion payoff: V*(µ0) = min_λ cav(^V_λ)(µ0). That is a clean, genuinely new characterization. The binary example is a nice illustration, and the connection to cautious Bayesian persuasion is a real contribution. The paper is also honest about its limits: transparent motives matter, and the quantitative magnitude of the improvement is left open.\n\nThe trouble is in the proof of Theorem 1. The stress-test note is on target. Lemma 5 claims that for a boundary belief µ and a direction µ1 supported inside supp(µ), the limit of V* at (1−ε)µ+εµ1 equals V*(µ) by Lemma 4. But Lemma 4 proves continuity only at full-support beliefs. When µ is on the boundary, (1−ε)µ+εµ1 has support supp(µ), a proper face, so Lemma 4 does not apply. There is no induction on support size, and the proof of Proposition 12's Case 2 explicitly needs continuity at a lower-support belief. As written, the almost-sure strict-improvement result is incomplete. This is a load-bearing gap, not a cosmetic one.\n\nThe bounded-budget section has a separate, smaller problem: Proposition 10 applies Sion's minimax with λ ranging over aff(Θ), which is not compact. Without a boundedness or coercivity argument, the interchange is not justified. The reader is right that this can likely be repaired, but it needs to be done.\n\nWhat is good: the upper bound and the approximating construction are solid; the value formula in Proposition 3 is derived from primitives, not assumed; the generic condition is clearly stated, and the counterexamples in the appendix show the authors know where their assumptions bite. The self-citation to Liu-Wu is fine—it is a related implementation paper, not a hidden foundation.\n\nWho should read it: anyone working on limited commitment, robust persuasion, or the Web3/blockchain communication story. It deserves a serious referee. My recommendation is to send it to peer review, but the referee should insist on a repaired proof of Lemma 5 / Proposition 7 before acceptance. If the gap cannot be closed, Theorem 1 as stated does not hold up.","headline":"The main idea and value formula are genuinely new and likely right, but the proof of Theorem 1 has a real gap in the boundary-continuity lemma that needs repair before the result is fully convincing.","tokens_in":33729,"tokens_out":3389,"would_cite":false,"duration_ms":29860,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A28","91B44"],"pacs":[],"model":"deepseek-v4-flash","headline":"Money burning can strictly improve mediated communication.","keywords":["money burning","mediated communication","Bayesian persuasion","cheap talk","commitment","mechanism design","transparent motives","incentive compatibility"],"falsifier":"A computational search over finite receiver payoff matrices that satisfy the generic condition could look for a positive-measure set of priors where commitment is valuable, $V^*_{CT}(\\mu_0)<\\max V$, yet $V^*_{MD}(\\mu_0)=V^*(\\mu_0)$; Theorem 1 asserts no such set exists. The paper's Example 3 shows the equality can occur outside the generic class, so the search must enforce Definition 3. A sharp experiment would compare sender payoffs under mediation with and without burning in a lab: under transparent motives the burning treatment should strictly dominate for almost all priors, and the gap should disappear when sender payoffs depend on his type.","tokens_in":32648,"feed_emoji":"💸","tokens_out":11726,"duration_ms":96284,"temperature":0.7,"pith_summary":"This paper asks whether deliberately wasteful money burning can make an informed sender more persuasive when he can commit to a mediator but not to a full persuasion strategy. The answer is yes, under transparent motives: when the sender's payoff depends only on the receiver's action, mediated communication with report-contingent burning generically beats ordinary mediation for almost every prior belief at which commitment has any value. The paper characterizes the sender's optimal payoff as the best worst-case interim payoff over signaling schemes, which is also the payoff of a cautious sender or of a sender facing the worst subjective prior in Bayesian persuasion. If the characterization is right, burning money is a practical substitute for commitment power: it lets a sender without full commitment capture part of the Bayesian-persuasion surplus.","feed_headline":"Burning money can strictly improve mediated communication","feed_subtitle":"A sender who can burn cash through a mediator gains credibility and beats ordinary mediation for almost all priors.","key_machinery":"The central object is a canonical mediated-communication-with-money-burning mechanism: the sender reports his type to a mediator, who publicly outputs a posterior belief from a pre-committed signaling scheme and burns a deterministic amount of money that depends only on that posterior. Because burning carries no information beyond the posterior, the receiver's obedience constraint and the sender's incentive constraints separate. The load-bearing identity is Proposition 3, $V^*(\\mu_0)=\\max_{\\pi}\\min_{\\theta}V_\\pi(\\theta)=\\min_{\\lambda}\\operatorname{cav}(\\hat V_\\lambda)(\\mu_0)$, obtained by viewing the mechanism-design problem as a zero-sum game between the sender and a fictitious chooser of the subjective prior $\\lambda$. The incentive-compatibility equation (5), $\\int_{\\mu}\\left(\\frac{\\mu(\\theta)}{\\mu_0(\\theta)}-\\frac{\\mu(\\theta')}{\\mu_0(\\theta')}\\right)(V(\\mu)-x(\\mu))\\,dp(\\mu)=0$, forces all types' net interim payoffs to be equal, and money burning $x(\\mu)$ is the transfer that makes this equalization possible. Proposition 11's construction, which burns money only on messages that reveal the type with small unconditional probability, attains this max-min value in the limit.","core_discovery":"On its own terms, the paper establishes Theorem 1: under a generic condition on the receiver's payoff functions (any action that is optimal at a belief is uniquely optimal at another belief with the same support), for almost all priors either cheap talk already reaches the sender's maximal feasible payoff, so commitment has no value, or money burning strictly raises the value of mediated communication. The companion characterization is $V^*(\\mu_0)=\\max_{\\pi}\\min_{\\theta}V_\\pi(\\theta)=\\min_{\\lambda\\in\\Delta(\\Theta)}\\operatorname{cav}(\\hat V_\\lambda)(\\mu_0)$, where $V_\\pi(\\theta)$ is type $\\theta$'s interim payoff under the signaling scheme $\\pi$ and $\\hat V_\\lambda(\\mu)=\\sum_{\\theta}\\lambda(\\theta)\\frac{\\mu(\\theta)}{\\mu_0(\\theta)}V(\\mu)$. This ties the money-burning value to cautious Bayesian persuasion and to Bayesian persuasion under the sender's worst subjective prior. The proof builds a revelation principle that separates a mechanism into a signaling scheme and a report-contingent burning schedule, then shows that whenever burning fails to beat ordinary mediation, cheap talk already attains the same value.","pith_inferences":["The equivalence with cautious Bayesian persuasion suggests money burning is a way to implement maxmin behavior without asking the sender to commit to ignoring his type: the burning schedule itself enforces equal interim payoffs.","Because the value is a minimum over subjective priors, algorithms that solve worst-prior persuasion could be repurposed to design optimal burning mechanisms, a computational shortcut the paper does not spell out.","A lab test could compare sender payoffs across cheap talk, mediation, and mediation with burning; under transparent motives the burning treatment should sit strictly between ordinary mediation and Bayesian persuasion for most priors, and the gap should disappear when sender payoffs are type-dependent.","The binary-type result (Proposition 6) suggests the worst subjective prior is always an extreme point in two-type environments; if that shortcut extends to small finite type spaces, optimal burning mechanisms become easy to compute."],"forward_implications":["At almost every prior where cheap talk cannot already deliver the sender's maximum feasible payoff, the optimal money-burning mechanism strictly outperforms ordinary mediation (Theorem 1).","Mediated communication with money burning has the same value as cautious Bayesian persuasion and as Bayesian persuasion under the sender's worst subjective prior, so robust-persuasion solution methods transfer directly (Corollaries 2 and 3).","With a finite burning budget, the value is $\\min_{\\lambda}\\operatorname{cav}(\\hat V_{\\lambda,C})(\\mu_0)$, and optimal mechanisms split messages into a costless persuasion group and a burning-for-credibility group (Proposition 10).","In smart-contract settings, commitment is valuable exactly when it is valuable in conventional communication, but the refined value of commitment is generically smaller (Corollary 4 and Theorem 1).","Money burning does not make cheap talk more credible; the credibility gain is specific to mediated communication (Section 6.2)."],"supporting_citations":[{"why":"Introduces cheap talk, the no-commitment benchmark that Theorem 1 must beat.","marker":"Crawford and Sobel [1982]"},{"why":"Supplies Bayesian persuasion value and the concave-envelope method used in Proposition 3.","marker":"Kamenica and Gentzkow [2011]"},{"why":"Defines the value of mediated communication without burning, the baseline MDMB is compared with.","marker":"Salamanca [2021]"},{"why":"Characterizes the bounds of mediated communication and gives the affine-combination comparison used to prove Theorem 1.","marker":"Corrao and Dai [2023]"},{"why":"Provides the transparent-motives cheap talk value qcav(V) on which the strict-improvement argument builds.","marker":"Lipnowski and Ravid [2020]"},{"why":"Supplies the limited-commitment revelation-principle structure that the canonical MDMB construction extends.","marker":"Doval and Skreta [2022]"},{"why":"Introduces truth-adjusted welfare functions and cautious persuasion, which give the interpretation of Vhat and Corollary 2.","marker":"Doval and Smolin [2024]"},{"why":"Establishes Bayesian persuasion with heterogeneous priors, the setting behind the worst-subjective-prior equivalence.","marker":"Alonso and Câmara [2016]"},{"why":"Shows money burning cannot improve cheap talk credibility, the contrast that motivates the mediated-communication setting.","marker":"Austen-Smith and Banks [2000]"},{"why":"Documents the failure of the revelation principle under imperfect commitment, motivating the paper's new principle.","marker":"Bester and Strausz [2001]"}],"fun_headline_variants":["Burning cash can strictly boost mediated talk","Why wasting money helps a sender persuade","Money burning can beat ordinary mediation","Burning money can make mediated talk credible","Burning can improve mediated communication"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result assumes the sender's payoff depends only on the receiver's action, not on his private type; if payoffs are state-dependent, money burning can fail to improve mediated communication, as the paper's buyer-seller example shows.","fun_headline_variants_meta":{"raw":{"variants":["Burning cash can strictly boost mediated talk","Why wasting money helps a sender persuade","Money burning can beat ordinary mediation","Burning money can make mediated talk credible","Burning can improve mediated communication"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001011,"raw_usage":{"total_tokens":4232,"prompt_tokens":866,"completion_tokens":3366,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":3305}},"tokens_in":482,"tokens_out":3366,"duration_ms":21592,"temperature":1.0,"reasoning_tokens":3305,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:13:05.146606+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A computational search over finite receiver payoff matrices that satisfy the generic condition could look for a positive-measure set of priors where commitment is valuable, $V^*_{CT}(\\mu_0)<\\max V$, yet $V^*_{MD}(\\mu_0)=V^*(\\mu_0)$; Theorem 1 asserts no such set exists. The paper's Example 3 shows the equality can occur outside the generic class, so the search must enforce Definition 3. A sharp experiment would compare sender payoffs under mediation with and without burning in a lab: under transparent motives the burning treatment should strictly dominate for almost all priors, and the gap should disappear when sender payoffs depend on his type.","supporting_citations":[{"cited_title":"Strategic information transmission","cited_arxiv_id":null,"evidence_quote":"Introduces cheap talk, the no-commitment benchmark that Theorem 1 must beat."},{"cited_title":"The value of mediated communication","cited_arxiv_id":null,"evidence_quote":"Defines the value of mediated communication without burning, the baseline MDMB is compared with."},{"cited_title":"Cheap talk with transparent motives","cited_arxiv_id":null,"evidence_quote":"Provides the transparent-motives cheap talk value qcav(V) on which the strict-improvement argument builds."},{"cited_title":"Mechanism design with limited commitment","cited_arxiv_id":null,"evidence_quote":"Supplies the limited-commitment revelation-principle structure that the canonical MDMB construction extends."},{"cited_title":"Persuasion and welfare","cited_arxiv_id":null,"evidence_quote":"Introduces truth-adjusted welfare functions and cautious persuasion, which give the interpretation of Vhat and Corollary 2."},{"cited_title":"Bayesian persuasion with heterogeneous priors","cited_arxiv_id":null,"evidence_quote":"Establishes Bayesian persuasion with heterogeneous priors, the setting behind the worst-subjective-prior equivalence."},{"cited_title":"Cheap talk and burned money","cited_arxiv_id":null,"evidence_quote":"Shows money burning cannot improve cheap talk credibility, the contrast that motivates the mediated-communication setting."},{"cited_title":"Contracting with imperfect commitment and the revelation principle: the single agent case","cited_arxiv_id":null,"evidence_quote":"Documents the failure of the revelation principle under imperfect commitment, motivating the paper's new principle."}],"review_version":1}