Evidence Markets: Dynamic Evidence-Driven Pricing

• Evidence markets are mechanisms that combine probability-belief vectors and atomic evidence to dynamically adjust market liquidity and pricing.
• They generalize traditional prediction markets by incentivizing both belief accuracy and evidence provision, enabling resolution through external outcomes or crowd-sourced verification.
• Their design employs a dynamic LMSR, LLM-based evidence verification, and asynchronous execution to aggregate information in settings lacking clear ground truth.

An evidence market is a dynamic mechanism in which sequentially arriving traders submit both probability-belief vectors over possible outcomes and “atomic” evidence items, with platform incentives structured to reward both belief accuracy and informative evidence provision. Unlike classical prediction markets, which function solely as aggregators of crowd beliefs and require exogenous, time-resolved outcomes, evidence markets directly incorporate submitted evidence into both the pricing mechanism and, when needed, endogenous outcome resolution. This approach generalizes the canonical logarithmic market scoring rule (LMSR) to dynamically adjust market liquidity based on total verified evidence quality. Evidence markets support bounded loss guarantees, truth-telling equilibria up to $\varepsilon$-dominance, automated or crowd-sourced verification, and LLM-judged scalable workflows, permitting new forms of information aggregation particularly suitable for settings that lack clear external, ground-truth events (Hossain et al., 5 Jun 2026).

1. Definition and General Construction

An evidence market is defined by requiring each trader interaction to include (a) a belief vector $q\in\Delta_{n-1}$ (probabilities over outcomes) and (b) a collection of atomic evidence items $E=\{e_1,\dots,e_\ell\}$. The mechanism aggregates all submitted beliefs into current prices via a generalization of the LMSR, and the platform assigns rewards both for increasing forecast accuracy (movement toward truth) and for contributing evidence that reduces uncertainty. If an external truth is available at event resolution, the market pays out using the realized outcome; otherwise, the outcome is determined endogenously by a pre-specified evidence-resolving procedure acting on the crowdsourced evidence set.

This design generalizes traditional prediction markets, which elicit only probabilistic beliefs over outcomes and pay solely based on ex post outcome resolution. Evidence markets (i) incentivize evidence contribution via explicit information-theoretic reward mechanisms, and (ii) enable endogenous, crowd-resolved settlement in the absence of exogenous ground-truth (Hossain et al., 5 Jun 2026).

2. Core Mechanism: Dynamic Evidence-Augmented LMSR

At the mechanism’s core lies the evidence-augmented logarithmic market scoring rule (LMSR). Let $s \in \mathbb{R}_{\ge 0}^n$ denote the outstanding share vector and $\beta>0$ a liquidity parameter. The classical cost function is $C(s)=\beta\,\ln\sum_{i=1}^n e^{s_i/\beta}$, with instantaneous prices $\pi_i(s)\propto e^{s_i/\beta}$. In the evidence market, $\beta$ becomes a non-increasing function of the total verified evidence quality $R=\sum_{e\in E^{\text{tot}}} r(e)$: $\beta(R)$. This design ensures that additional evidence reduces market liquidity as uncertainty falls, increasing price-responsiveness and aligning platform loss with the information coarse structure of the environment.

A trader’s expected payoff decomposes as the sum of (1) a belief update term proportional to the Kullback–Leibler divergence between updated and previous prices (expressing reward for price movement toward reality given the true outcome), and (2) an explicit “evidence reward” proportional to the reduction in entropy due to evidence, $q\in\Delta_{n-1}$0, where $q\in\Delta_{n-1}$1 and $q\in\Delta_{n-1}$2 is the Shannon entropy of post-trade beliefs. Critically, even risk-averse traders who do not move prices (i.e., submit no shares but do submit evidence) are rewarded proportionally to the remaining market uncertainty, provided their evidence is judged to be informative (Hossain et al., 5 Jun 2026).

The table below summarizes key mechanism components:

Component	Classic Prediction Market (LMSR)	Evidence Market (Augmented LMSR)
Inputs per trade	Probability vector $q\in\Delta_{n-1}$3	Probability vector $q\in\Delta_{n-1}$4, evidence $q\in\Delta_{n-1}$5
Liquidity parameter	Fixed $q\in\Delta_{n-1}$6	$q\in\Delta_{n-1}$7 non-increasing in $q\in\Delta_{n-1}$8
Rewards	Belief accuracy at resolution	Belief accuracy + information gain via $q\in\Delta_{n-1}$9
Outcome resolution	Exogenous only	Exogenous or evidence-based (endogenous)

3. Incentive Properties and Loss Guarantees

Evidence markets preserve bounded-loss properties from the base LMSR: for $E=\{e_1,\dots,e_\ell\}$0 outcomes, the maximum platform loss is $E=\{e_1,\dots,e_\ell\}$1, where $E=\{e_1,\dots,e_\ell\}$2 is the initial evidence quality. Theorem 1 demonstrates that evidence-related payoffs scale with remaining uncertainty, and the cost-function mechanism can always be equivalently expressed as an automated market maker (AMM) for practical deployment.

In the endogenous-resolution regime, the settlement uses a “softmax” function of the normalized total evidence support for each outcome; e.g., with $E=\{e_1,\dots,e_\ell\}$3 verified items and $E=\{e_1,\dots,e_\ell\}$4 support indicators, $E=\{e_1,\dots,e_\ell\}$5 and $E=\{e_1,\dots,e_\ell\}$6. Lemmas bound the extent to which traders can manipulate resolution by withholding evidence, showing that truthful reporting of both beliefs and evidence is an $E=\{e_1,\dots,e_\ell\}$7-dominant strategy for any desired $E=\{e_1,\dots,e_\ell\}$8 with sufficient temperature parameter $E=\{e_1,\dots,e_\ell\}$9 (Hossain et al., 5 Jun 2026).

4. Operational Implementation: Evidence Verification and Asynchronous Execution

Verification is central to the proper functioning of an evidence market, given that evidence directly influences both rewards and, where applicable, endogenous resolution. The recommended protocol employs a LLM-as-judge structure: each evidence item is LLM-verified (Accept/Reject) with a subsequent dispute window enabling staking-based challenges. Incorrect or frivolous disputes result in forfeited collateral, while correct challenges yield rewards. This adversarial mechanism, with “skin in the game,” provides an incentive-compatible verification model for both exogenous and endogenous resolution regimes.

Execution is optimized through an asynchronous dual-queue architecture: trades are accepted and priced immediately under worst-case assumptions, while evidence verification and cost adjustment proceed in parallel, issuing refunds when the true evidence quality is certified. The correctness theorem guarantees that each trader's net cost equals the hypothetical cost under synchronous verification, and all evidence-free trades settle instantly at prevailing liquidity, supporting high-throughput applications (Hossain et al., 5 Jun 2026).

5. Example Application: LLM Model Evaluation

A salient application of evidence markets is benchmarking LLMs. Consider determining which model (A or B) is superior at a specific task, such as medical question-answering ($s \in \mathbb{R}_{\ge 0}^n$0 outcomes). Trades can submit question–answer pairs as atomic evidence, scored for discriminative power: an item is maximally discriminative if only one model answers correctly, in which case that model's relative support and the market’s liquidity parameter are adjusted accordingly.

For instance, submitting 10 questions only model A answers correctly reduces $s \in \mathbb{R}_{\ge 0}^n$1, increasing sensitivity; subsequent traders can more efficiently express directional beliefs and earn rewards for evidence provision or correct, belief-weighted forecast movements. When sufficient evidence items ($s \in \mathbb{R}_{\ge 0}^n$2) accumulate, endogenous resolution is triggered using a softmax mapping of support fractions ($s \in \mathbb{R}_{\ge 0}^n$3, $s \in \mathbb{R}_{\ge 0}^n$4) to outcome probabilities (Hossain et al., 5 Jun 2026).

6. Comparison to Classical Prediction Markets

Evidence markets extend prediction markets in several rigorous and operationally significant dimensions:

• Transparency: Markets record not only beliefs but also an auditable evidence ledger, enabling external verification and post hoc analysis.
• Self-Resolution: Events lacking exogenous ground-truth are resolvable via crowd-verified evidence.
• Direct Evidence Incentives: Mechanisms explicitly reward the addition of informative, uncertainty-reducing evidence; even agents unwilling to stake directional bets can earn via evidence submission.
• Flexible Use Modes: Interfaces allow any combination of pure prediction trading, pure evidence contribution, or hybrid participation.

However, the expanded structure introduces challenges not present in standard markets:

• Verification—particularly when using LLMs and staking-based disputes—may be subject to adversarial manipulation, model bias, or scalability constraints.
• Parameter selection, such as the evidence-resolution temperature $s \in \mathbb{R}_{\ge 0}^n$5 and $s \in \mathbb{R}_{\ge 0}^n$6 function, requires careful tuning to balance manipulation-resistance and resolution granularity.
• The effort or cost required to generate relevant evidence is not modeled in the base mechanism but has significant impact on real-world participation rates.
• Engineering asynchronous and parallelizable verification with real-time trading creates non-trivial overhead (Hossain et al., 5 Jun 2026).

7. Relationship to Evidence Aggregation and Market Efficiency

The evidence market paradigm formalizes and mechanizes the intuition that market-based systems can aggregate not just beliefs but the chains of reasoning or empirical justification underlying those beliefs. In traditional prediction markets, the aggregation is purely epistemic, with no explicit treatment of partial information, intermediating evidence, or rationale structure.

By elevating evidence to a first-class object—impacting both interim prices and eventual settlement—evidence markets operationalize principles of information theory in economic design. In contrast to real-world betting or financial markets, which function as reactive evidence-aggregators of observed variables but do not proactively encode latent foreknowledge (cf. findings in (Winkelmann et al., 27 May 2025)), evidence markets provide a structured setting for the aggregation and evaluation of intermediate, verifiable signals, particularly in domains where the resolution event is inherently ambiguous or absent (Hossain et al., 5 Jun 2026, Winkelmann et al., 27 May 2025).