Calibrated Mechanism Design

• Calibrated mechanism design is a framework that maintains incentive compatibility by addressing information leakage through repeated interactions.
• It utilizes a two-stage mechanism where a Bayes-plausible signal is disclosed before executing a direct allocation rule based on agents' posterior beliefs.
• Design procedures involve optimizing payoff functions via concavification, ensuring alignment of agent learning with incentive compatibility and individual rationality.

Calibrated mechanism design is a framework in which mechanisms are required to remain incentive compatible under the constraints imposed by repeated interaction and the information revealed to agents over time. This approach integrates classical mechanism design with information design, formalizing the constraints faced when agents learn about hidden states (such as auction parameters or environmental conditions) from repeated allocations and reports. The goal is to characterize, construct, and optimize mechanisms that are "calibrated," meaning they are robust to the information leakage about underlying states caused by repeated use, while maintaining incentive compatibility and individual rationality (Doval et al., 19 Dec 2025).

1. Foundations of Calibrated Mechanism Design

A calibrated mechanism is defined in a setting with a finite set of states $\Theta$ (prior $\pi \in \Delta(\Theta)$), agent types $t \in T$ (distribution $f \in \Delta(T)$), and allocations $a \in A$. The designer commits to a mechanism $$m: T \times \Theta \times [0,1] \rightarrow \Delta(A)$$. The agent observes only the allocation resulting from their reports. Over repeated use, the agent can experiment with different reports, learning about the hidden state by observing the mapping $\hat t \mapsto m(\hat t, \theta, \cdot)$.

The key concept is the calibrated information structure: the distribution over interim allocation rules $$r_\theta(\hat t) = \mathbb{E}_{\omega}[m(\hat t, \theta, \omega)]$$ induced by the mechanism and the prior on states. Calibration requires that the agent's ex ante probability of seeing a given interim rule matches the underlying prior. Mechanisms are calibrated if they remain incentive compatible and individually rational with respect to what agents can learn through repeated interaction (Doval et al., 19 Dec 2025).

2. Characterization: Two-Stage Mechanisms and Calibration

The central result is that in the presence of agent learning, the set of implementable outcomes coincides with those implementable by "two-stage mechanisms." In this construction, the designer first publicly reveals a Bayes-plausible signal about the underlying state (formally, an experiment $\beta \in \Delta(\Delta(\Theta))$ with $\int b\,\beta(db) = \pi$), and then commits to a direct mechanism $$\rho: \Delta(\Theta) \times T \rightarrow \Delta(A)$$ that does not depend on the true state, but only on the agent's posterior $b$ induced by the signal.

This yields a full allocation rule

$$\mu(a, t, \theta) = f(t) \int_{b \in \Delta(\Theta)} \rho(a \mid t, b) \beta(db) \pi(\theta).$$

A mechanism is calibrated if and only if it is implementable via such two-stage mechanisms, with truthfulness and participation constraints enforced at every possible posterior belief over the state that the agent may obtain from learning (Doval et al., 19 Dec 2025).

3. Constructive Procedures and the Design Problem

The process for constructing a calibrated mechanism involves:

1. For each candidate posterior $b \in \Delta(\Theta)$, solving the standard mechanism design problem (e.g., revenue maximization subject to incentive compatibility and individual rationality under prior $b$), yielding a value or payoff function $W(b)$.
2. "Concavifying" $W(\cdot)$ over $\Delta(\Theta)$: computing its concave envelope $$\mathrm{cav}\,W(\pi)=\max_{\beta:\int b\,\beta(db)=\pi} \int W(b)\,\beta(db)$$.
3. Constructing the Bayes–plausible experiment $\beta$ whose support achieves the concave envelope at the prior $\pi$. The mechanism $\rho$ uses the optimal solution at each $b$.

This reduction to information design aligns the calibrated mechanism design problem with the classic menu of Bayesian persuasion, subject to the additional constraints of incentive compatibility and individual rationality with respect to induced beliefs (Doval et al., 19 Dec 2025).

4. Implications in Private-Values and Quasilinear Environments

In private values settings ($u(a, t, \theta) = u(a, t)$), the calibration constraint enforces full transparency: the designer must fully reveal the state to the agent. The optimal calibrated mechanism is "state-by-state Myerson": for every state $\theta$, run the optimal mechanism tailored to that state, as if it were common knowledge. Any form of correlation-based surplus extraction (e.g., à la Crèmer–McLean) fails, as pooling types and partial information becomes infeasible under calibration (Doval et al., 19 Dec 2025).

The intuition is that the agent's payoff is independent of the state, so $W(b)$ is linear in $b$ and the concavification is trivial. Hence, only full disclosure (i.e., revealing the true state) is compatible with calibrated incentive compatibility.

5. Microfoundation: Repeated Interaction and Learning

Calibrated mechanisms capture precisely what is implementable in dynamic settings where an agent repeatedly interacts with a fixed mechanism. Over time, the agent, by experimenting with reports and observing allocations, infers the "interim rule" associated with the hidden state. By martingale convergence properties of Bayesian learning, the agent's belief sequence converges almost surely, so the long-run occupation measure over observed outcomes reduces to the two-stage structure: first, learn a signal (posterior), then play the associated direct mechanism $\rho$.

A key theorem is that the set of distributions $\mu(a, t, \theta)$ implementable via repeated learning by a patient agent coincides exactly with those implementable statically by a calibrated mechanism (Doval et al., 19 Dec 2025).

6. Dynamic Mechanisms Versus Static Calibration

Dynamic or history-dependent mechanisms can potentially expand the set of feasible allocations by employing punishments for detectable deviations (those that alter the empirical report distribution). However, in transferable utility (e.g., quasilinear) environments, this distinction vanishes—dynamic commitment does not enlarge implementable outcomes compared to static calibrated mechanisms. The cyclic-monotonicity conditions for incentive compatibility collapse to those enforced by static calibration, so dynamic mechanisms offer no extra advantage in these cases. In contrast, in interdependent-value or non-quasilinear settings, dynamically weakening incentive constraints may strictly expand the set of implementable outcomes (Doval et al., 19 Dec 2025).

7. Connections to Learning-Augmented and Output-Advice Mechanism Design

A related research stream addresses the impact of imperfect predictions in mechanism design, aiming for mechanisms whose performance "degrades gently" with the prediction error but provides worst-case guarantees. Notable contributions establish quantitative relationships among three metrics: consistency (accuracy when predictions are perfect), robustness (worst-case guarantee), and error-tolerance (range of prediction error for a given guarantee) (Christodoulou et al., 2024, Xu et al., 2022). These are closely related to calibration in that mechanisms must interpolate smoothly between high-performance regimes (accurate predictions/advice) and robust safety nets (arbitrary or adversarial advice).

A universal "quality of recommendation" $\delta$—defined as ratio between optimal and advised objective—serves as a calibration measure for output advice, allowing performance guarantees to be parametrized by advice quality (Christodoulou et al., 2024). Tables of canonical bounds for facility location, scheduling, and combinatorial auctions formalize this connection.

• "Calibrated Mechanism Design" (Doval et al., 19 Dec 2025)
• "Mechanism design augmented with output advice" (Christodoulou et al., 2024)
• "Mechanism Design with Predictions" (Xu et al., 2022)