{"id":"57e64e67-277e-493d-a7a6-2bf8ea77375d","arxiv_id":"2607.17428","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A uniform-LVR AMM exists for every separable win-martingale price process, and every sufficiently regular pool-value function is made uniform by an induced win-martingale, with liquidity schedules controlling loss timing.","lead":"This paper defines 'uniform' automated market makers for binary prediction markets, where the cost of subsidizing price discovery—loss-versus-rebalancing—is spread evenly across all price states after normalizing for information arrival. It proves that any smooth belief process can be paired with an AMM curve that achieves this balance, and that liquidity can be scheduled over time to hit a target loss profile.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 10 (minimax optimality) is false as stated: for G(p)=min(p,1-p), the uniformity BVP admits positive C^2 solutions for every beta in (0,1/4), giving uniform AMMs with arbitrarily small worst-case relative LVR.","rationale":"The reader's weakest assumption was the continuous/no-jump price process. That is a legitimate scope limitation, but the paper is explicit about focusing on continuous diffusions, so it is not an internal inconsistency. The more serious issue is that Theorem 10, one of the paper's normative justifications for uniform LVR, is mathematically false as stated. The proof assumes H^1_0 regularity that the theorem's hypotheses do not require, and the singular endpoints admit classical non-H^1 solutions of the uniformity BVP for a continuum of eigenvalues. For G(p)=min(p,1-p), these solutions are positive, concave, C^2 on (0,1), and yield uniform relative LVR equal to beta for any beta in (0,1/4). Thus the claimed minimax optimality fails; in fact the infimum of achievable relative LVR rates is 0. This does not refute the central bidirectional correspondence (Theorems 11 and 13) or the dynamic scheduling result (Theorem 14), both of which appear sound. It does mean the paper's stated minimax fairness property needs either a corrected statement (e.g., restricting to H^1_0 value functions or to the principal eigenpair) or removal. The reader's CONDITIONAL verdict remains appropriate: the main construction is credible, but the manuscript contains a substantive false theorem that must be fixed before acceptance.","tokens_in":23696,"tokens_out":55340,"duration_ms":531036,"concrete_test":"Verify the explicit counterexample for G(p)=min(p,1-p), beta=1/8. Set r_+=(1+sqrt(1-4beta))/2=(1+1/sqrt2)/2, r_-=(1-1/sqrt2)/2, and rho=r_+/r_-. Define V(p)=p^{r_+} on [0,1/2]; on [1/2,1] define V(p)=C(1-p)^{r_+}+D(1-p)^{r_-}, with C=(1+rho)/(1-rho) and D=(1-C)/2^{r_+-r_-}. Check that V and V' match at p=1/2, that V is positive and C^2 on (0,1), and that -V''(p)G(p)^2/V(p)=1/8 identically. This is a uniform AMM with worst-case relative LVR 1/8, below the Hardy principal value 1/4, contradicting Theorem 10 as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The bidirectional correspondence (Theorems 11 and 13) is internally plausible, but the paper's minimax optimality claim, Theorem 10, is false as stated. The theorem allows any V in C([0,1]) cap C^2((0,1)) with V(0)=V(1)=0 and V>0, but the proof applies a Rayleigh-quotient inequality over H^1_0(0,1). These classes differ at singular endpoints. For G(p)=min(p,1-p), which satisfies Assumption 3, the uniformity BVP G(p)^2 V''(p)+beta V(p)=0 has, for every beta in (0,1/4), a positive classical solution V_beta that is C^2 on (0,1), continuous on [0,1], zero at both endpoints, and not in H^1_0. Explicitly, with r_+=(1+sqrt(1-4beta))/2 and r_-=(1-sqrt(1-4beta))/2, set V(p)=A p^{r_+} on [0,1/2], and on [1/2,1] set V(p)=C(1-p)^{r_+}+D(1-p)^{r_-}, choosing C,D to match V and V' at p=1/2. Then V solves the BVP pointwise and is concave, so its relative-LVR ratio -V''(p)G(p)^2/V(p) equals beta for every p. Since beta can be any number in (0,1/4), the set of achievable uniform relative-LVR rates contains (0,1/4); the infimum is 0 and no uniform AMM minimizes worst-case relative LVR. The paper itself (Section 4.1) acknowledges a family of admissible eigenpairs, but Theorem 10's statement and proof do not restrict to the principal eigenpair or to H^1_0, so the minimax claim is unsupported and contradicted by these explicit solutions.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces uniform automated market makers for binary prediction markets, defined by the requirement that instantaneous loss-versus-rebalancing (LVR) is proportional to pool value and independent of price, after normalizing by an information clock. The central object is the uniformity boundary value problem G(p)^2 V''(p) + β V(p) = 0 with V(0)=V(1)=0, V>0. For separable win-martingales satisfying Assumption 3, the paper claims a bidirectional correspondence: Theorem 11 constructs a concave pool-value function from a volatility profile, and Theorem 13 constructs a win-martingale from a concave pool-value function. Theorem 14 extends the framework to dynamic liquidity, giving an explicit deterministic liquidity schedule that implements any prescribed target expected cumulative loss schedule. The paper also claims a minimax optimality property for uniform LVR (Theorem 10) and illustrates the constructions on canonical volatility profiles and AMMs.","tokens_in":24164,"tokens_out":16125,"duration_ms":170363,"significance":"If the main existence and correspondence results hold, this is a valuable framework: it gives a constructive link between belief-process dynamics and AMM geometry, yields explicit closed-form liquidity schedules for dynamic loss targeting, and recovers canonical mechanisms such as CPMM and LMSR as uniform for specific win-martingales. The paper is generally transparent about its modeling scope (continuous Itô price processes, risk-neutral zero-rate setting) and it explicitly acknowledges non-uniqueness of admissible eigenpairs. The dynamic implementation result is a genuinely useful design tool. However, the claimed minimax optimality is false as stated, and because this claim is used in the introduction and Section 4 as a normative justification for uniform LVR, the paper requires substantive revision before it can be accepted.","major_comments":[{"comment":"The minimax claim is false as stated. For G(p)=min(p,1-p), which satisfies Assumption 3, every β∈(0,1/4) admits a positive solution Vβ of G²V''+βV=0 with V(0)=V(1)=0. With r±=(1±√(1−4β))/2, take Vβ=A p^{r+} on [0,1/2] and match to C(1−p)^{r+}+D(1−p)^{r−} on [1/2,1]; the matching constants give a positive concave C² interior function whose relative-LVR ratio is identically β. Since β can be arbitrarily small, the theorem's lower bound fails. The proof's Rayleigh-quotient integration by parts is over H^1_0, but Vβ∉H^1_0 (near 1, V∼D(1−p)^{r−}, r−<1/2, so ∫(V')²=∞). Section 4.1 itself acknowledges a family of admissible eigenpairs. The theorem must specify the principal eigenpair and restrict the admissible class, or the minimax claim should be withdrawn.","section":"Section 4, Theorem 10 / Appendix B.6"}],"minor_comments":[{"comment":"The absorbed Brownian motion profile G(p)=1_{0<p<1} does not satisfy Assumption 3(i), which requires continuity on [0,1] with G(0)=G(1)=0. The paper calls this example 'formal,' but since it is used to derive an explicit invariant in §4.2, the mismatch should be acknowledged explicitly or the example should be presented as a limiting case.","section":"Section 2.2 / Section 4.2"},{"comment":"The statement that the square-root clock h(t)=√(T−t) has 'remaining quadratic variation of the process is (T−t)' is misleading for the price process: by Proposition 4(ii) the conditional remaining quadratic variation equals P_t(1−P_t), which is stochastic. The claim is only true for the latent driving Brownian motion or for a specific underlying process, and should be clarified.","section":"Section 2.3"},{"comment":"For the absorbed Brownian motion invariant, the text says 'After a series of substitutions, we can show' and gives F(x,y;L). The derivation is entirely omitted. Since this is a claimed explicit construction, a sketch or a reference to an appendix is needed.","section":"Section 4.2"},{"comment":"The phrase 'The ratio −V(p)/V''(p) yields the dynamics' omits the role of β. For the CPMM the displayed dynamics dP_t=P_t(1−P_t)/h(t)dW_t is correct only after fixing β=1/4; for the LMSR the displayed dynamics uses β=1. Please state the chosen β explicitly in these examples.","section":"Section 4.4"},{"comment":"There are several typos and small notational inconsistencies: 'ahve' (§2.3), 'fiar' (§4.1), 'It¯o' (§3.2), 'is has' / 'has is' (§4.4), 'loses' for 'losses' (§5). Also, LVR_t is defined as an instantaneous rate in Proposition 8 but used as both a rate and an increment in the proof of Theorem 14; the notation dLVR_t would remove the ambiguity.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The main construction theorems (11, 13, 14) appear defensible and are worth publishing after repair. The false Theorem 10 is prominent in the paper's normative motivation, so I cannot accept the current version; however, the flaw is localized and fixable by restricting the theorem to the principal eigenpair / correct domain or withdrawing the minimax claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's what you should know: the core construction — every separable win-martingale in their class admits a pool-value function making LVR uniform, and conversely any sufficiently regular pool-value function induces a compatible win-martingale — is real and well argued. Theorems 11 and 13 hold up on inspection. That is a genuine generalization of Moallemi and Robinson's Gaussian-only result, and the dynamic liquidity schedule in Theorem 14 is a clean, useful contribution. The examples (Wright-Fisher/sqrt invariant, logistic/CPMM, Gaussian/normal density) are informative, and the paper is honest about its scope.\n\nBut the minimax optimality claim, Theorem 10, is false as stated. The stress-test example is correct: for G(p)=min(p,1-p), which satisfies Assumption 3, the uniformity BVP G(p)^2 V'' + beta V = 0 admits positive concave classical solutions for every beta in (0,1/4) (power-law near each endpoint, matched at 1/2). Each gives uniform relative LVR equal to beta, and the functions are in C([0,1]) ∩ C^2((0,1)) with V(0)=V(1)=0. So the achievable uniform rates include all beta below the Hardy critical value 1/4; the infimum is 0 and no minimizer exists. The proof uses a Rayleigh-quotient argument over H^1_0, which implicitly selects the principal eigenvalue, but the theorem statement does not restrict to that. The fix is easy: state Theorem 10 for the maximal eigenvalue (or for the specific beta constructed in Theorem 11) and handle the non-attainment case separately. As written, it is a wrong side theorem. The main correspondence and dynamic results do not rely on it.\n\nThe other flagged concerns are less serious. The jump-diffusion limitation is a scope choice, not an error; the authors are explicit about continuous Itô paths. The omitted 'most exciting game' invariant is an example-level gap — minor. The reliance on cited Sturm-Liouville criticality results means the existence proof is as strong as those citations, which is acceptable for a theory paper, but it makes full independent verification more work. No code or data, but the theoretical contribution stands alone.\n\nBottom line: this is a serious paper that deserves a referee. I would not desk-reject it. But I would require the authors to fix Theorem 10 before acceptance — it is an embarrassing counterexample to a stated theorem, even if the paper's central claims survive. Anyone working on prediction-market mechanism design or AMM theory should read it, and cite it once Theorem 10 is corrected.","headline":"The main win-martingale/AMM correspondence is solid and worth engaging with, but Theorem 10's minimax claim is false as stated and needs a straightforward fix.","tokens_in":760,"tokens_out":749,"would_cite":true,"duration_ms":52453,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B26","60G44","60J60","91G80"],"pacs":[],"model":"deepseek-v4-flash","headline":"A second-order boundary value problem equates prediction-market AMM design with the choice of a belief process, so that loss is proportional to pool value at every price.","keywords":["uniform LVR","prediction markets","automated market maker","win-martingale","loss-versus-rebalancing","boundary value problem","liquidity schedule","pool-value function"],"falsifier":"Take a win-martingale with jumps, e.g., a compensated Poisson-driven belief process converging to 0 or 1, and compute the realized ratio LVR/V along a simulated path; if the ratio varies with price despite a pool-value function solving the continuity-based BVP for the diffusion part, the no-jump assumption is falsified. Alternatively, on real high-frequency prediction-market data, a uniform AMM fitted from the continuous model should show a flat LVR/V across price buckets; systematic elevation near p≈0 or p≈1 would indicate missing jump risk.","tokens_in":23585,"feed_emoji":"📉","tokens_out":3668,"duration_ms":35444,"temperature":0.7,"pith_summary":"The paper introduces uniform AMMs for binary prediction markets: mechanisms whose instantaneous loss-versus-rebalancing is proportional to pool value and independent of the current token price, after normalizing for the rate of information arrival. It proves a two-way equivalence: every separable win-martingale satisfying mild regularity admits a concave pool-value function that solves a boundary value problem, and every sufficiently regular concave pool-value function defines a win-martingale under which it is uniform. A deterministic liquidity schedule then implements any prescribed target expected cumulative loss over time, separating statewise from timewise subsidy. If correct, AMM designers and subsidizers can shape where and when losses occur instead of only bounding the total.","feed_headline":"A single boundary problem shapes AMM losses evenly across prices","feed_subtitle":"Prediction-market AMMs can match their geometry to belief dynamics so loss tracks pool value at every price.","key_machinery":"The uniformity BVP G(p)^2 V''(p)+βV(p)=0 with V(0)=V(1)=0 and V>0 concave is the load-bearing object. It enforces that the state-dependent volatility G(p)^2 and the AMM curvature V''(p) cancel so that -½V''G^2 = (β/2)V at every price; β acts as the uniform loss rate per unit informational time. The singular Sturm-Liouville formulation provides existence of the eigenpair via Hardy-type bounds and the ground-state alternative, and its homogeneity allows arbitrary liquidity scaling and the closed-form dynamic liquidity schedules.","core_discovery":"The central result is the uniformity BVP, G(p)^2 V''(p)+βV(p)=0 on (0,1) with V(0)=V(1)=0, V>0, viewed as a singular Sturm-Liouville eigenpair. Theorem 11 constructs a concave unimodal V for any separable win-martingale with volatility profile G and information clock h(t) satisfying Assumption 3; Theorem 13 constructs G from V by G^2 = -βV/V''. Theorem 14 then shows that with a uniform V fixed, scaling liquidity by the schedule L_t ∝ h(t)^2 D'(t) exp(∫ β/(2h^2)) implements the target expected loss D(t). Examples: Wright-Fisher G=√(p(1-p)) yields constant-elasticity pool value p(1-p); logistic G=p(1-p) yields the constant product market maker; Gaussian score dynamics yield V=φ(Φ^{-1}(p)); the","pith_inferences":["My inference: the same BVP machinery extends naturally to non-uniform objectives by replacing the constant β with a price-dependent weight, which the paper gestures at; this would let a designer concentrate subsidy at, say, p=1/2.","My inference: empirical calibration of G and h from tick-level prediction-market data could test which canonical profile (logistic, Wright-Fisher, Gaussian score) real markets approximate; the uniform AMM would then be directly implementable.","My inference: in multi-outcome or combinatorial markets the scalar BVP would generalize to an elliptic eigenvalue problem on the simplex, with Hardy weights on the boundary; the 'uniform loss' criterion would become a normal-flux matching condition."],"forward_implications":["For any separable win-martingale in the class, a concave pool-value function exists that makes instantaneous LVR a constant fraction of pool value across all price states.","Conversely, every sufficiently regular concave pool-value function embeds a volatility profile under which it is uniform — AMM geometry encodes belief dynamics.","With a uniform value function, a deterministic liquidity schedule implements any target expected cumulative loss schedule in closed form, so designers can front-load or back-load losses without changing the statewise distribution.","Under uniform LVR, the subsidizer's loss rate per dollar deployed is the same in every price state, removing the incentive for sophisticated LPs to time entry and exit.","Uniform LVR is the minimax design: among all admissible pool-value functions it minimizes the worst-case relative loss rate across prices."],"fun_headline_variants":["Uniform-loss AMMs arise from a single boundary value problem","How to shape AMM losses evenly across all prices","Boundary problem yields uniform-loss AMM design","Uniform AMMs: loss tracks pool value at every price","Design AMMs so loss is uniform across price and time"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The fair price is assumed to move continuously via an Itô diffusion with no jumps; if real belief updates arrive as discrete jumps, the LVR decomposition and the uniformity BVP would need a different framework, and the stated correspondence would not hold as written.","fun_headline_variants_meta":{"raw":{"variants":["Uniform-loss AMMs arise from a single boundary value problem","How to shape AMM losses evenly across all prices","Boundary problem yields uniform-loss AMM design","Uniform AMMs: loss tracks pool value at every price","Design AMMs so loss is uniform across price and time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000577,"raw_usage":{"total_tokens":2599,"prompt_tokens":826,"completion_tokens":1773,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":1692}},"tokens_in":570,"tokens_out":1773,"duration_ms":11316,"temperature":1.0,"reasoning_tokens":1692,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T17:58:55.597958+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a win-martingale with jumps, e.g., a compensated Poisson-driven belief process converging to 0 or 1, and compute the realized ratio LVR/V along a simulated path; if the ratio varies with price despite a pool-value function solving the continuity-based BVP for the diffusion part, the no-jump assumption is falsified. Alternatively, on real high-frequency prediction-market data, a uniform AMM fitted from the continuous model should show a flat LVR/V across price buckets; systematic elevation near p≈0 or p≈1 would indicate missing jump risk.","supporting_citations":[],"review_version":1}