Admissibility Alignment: A Distributional Framework

• Admissibility Alignment is a framework that defines AI alignment as an admissibility property of policies, using probabilistic risk measures and distributional metrics to guide decision making.
• The MAP-AI implementation employs Monte Carlo rollouts and a control-plane architecture to stress-test policies against predefined risk thresholds and governance constraints.
• It extends classical decision theory by integrating uncertainty, value ambiguity, and regulatory requirements to ensure transparent and auditable AI deployments.

Admissibility Alignment is a formal framework reframing AI alignment as an admissibility property of policies, based on the probabilistic assessment of action and decision selection over distributions of outcomes under uncertainty. This concept requires that decision policies meet institutionally specified thresholds on risk, unacceptable outcomes, and tail events, operationalized through distributional metrics and enforced via a control-plane architecture distinct from static, model-level constraints. The canonical implementation, MAP-AI (Monte Carlo Alignment for Policy), defines and evaluates alignment as a property of policies over ensembles of plausible futures, integrating uncertainty modeling, intervention effects, value ambiguity, and governance constraints for trustworthy, auditable AI deployment (Duffey, 5 Jan 2026).

1. Formal Basis: Admissibility Alignment Definition and Distributional Metrics

Admissibility alignment replaces classical, point-estimate notions of AI alignment with an outcome-distribution-centric paradigm. For a scenario generator $G$ producing worlds $w \sim P_G(w)$, a (possibly stochastic) policy $\pi$, and value parameters $\theta \in \Theta$ drawn from $P(\theta)$, a trajectory $\tau = (s_0,a_0,\dots,s_T)$ is characterized by: $$P(\tau \mid \pi) = \int_{w,\theta} P(\tau \mid \pi, w, \theta) \; P_G(w) \; P(\theta) \; dw \, d\theta$$ Given utility $U(\tau; \theta)$, loss $L(\tau; \theta) = -U(\tau; \theta)$, and a set of hard constraints $C$ (or unacceptable outcomes $M$), empirical evaluation proceeds via Monte Carlo rollouts: $$(w_i,\theta_i) \sim P_G(w)P(\theta),\, \tau_i \sim P(\tau | \pi, w_i, \theta_i),\, U_i = U(\tau_i; \theta_i),\, L_i = -U_i,\, c_i = \mathbf{1}\{\tau_i \in M\}$$ The core risk and alignment metrics are:

• Expected utility: $\widehat{\mathbb{E}[U]} = \frac1N \sum_{i=1}^N U_i$
• Utility variance: $$\widehat{\mathrm{Var}[U]} = \frac1{N-1}\sum_{i=1}^N (U_i - \widehat{\mathbb{E}[U]})^2$$
• Constraint-violation probability: $\hat{p}_{\mathrm{viol}} = \frac1N\sum_{i=1}^N c_i$
• Tail risk (CVaR): $$\widehat{\mathrm{CVaR}_\alpha} = \frac{1}{|\mathcal{I}_\alpha|} \sum_{i\in\mathcal{I}_\alpha} L_i$$ where $$\mathcal{I}_\alpha = \{ i: L_i \geq \hat{q}_{1-\alpha} \}$$

A policy $\pi$ is admissible if: $$\hat{p}_{\mathrm{viol}} \leq \epsilon \quad\text{and}\quad \widehat{\mathrm{CVaR}_\alpha} \leq K$$ for governance thresholds $\epsilon$ (maximum acceptable violation rate) and $K$ (maximum acceptable tail loss). Among admissible policies, selection prioritizes highest $\widehat{\mathbb{E}[U]}$ (Duffey, 5 Jan 2026).

2. MAP-AI Architecture: Control-Plane for Admissibility-Governed Alignment

MAP-AI operationalizes admissibility alignment in three system components:

• Part I: Monte Carlo Uncertainty Engine—Generates sample rollouts under varying world states $w$ and value parameters $\theta$.
• Part II: Distributional Alignment Stress Testing—Computes empirical risk metrics above over the induced trajectory distribution.
• Part III: Decision Integration (Champion–Challenger/Admissibility Filter)—Filters admissible policies and selects optimal via a proof-carrying admissibility compiler (PCAC), with possible outcomes: Act, Escalate, or Abort.

This entire process is external to the model internals—requiring no retraining or parameter updates of policy models. Algorithm 1 in (Duffey, 5 Jan 2026) specifies the full pseudocode, emphasizing the tractable and auditable encapsulation of admissibility filtering and metrics logging.

3. Uncertainty, Value Ambiguity, and Governance in Admissibility Alignment

The admissibility alignment paradigm explicitly decomposes sources of uncertainty:

• World uncertainty: $w \sim P_G(w)$ models environmental and systemic variation.
• Policy stochasticity: Randomized components within $\pi$.
• Trajectory evolution: Stochastic transitions given $(\pi, w)$.
• Value uncertainty: $\theta \sim P(\theta)$ represents institutional ambiguity in tradeoffs (risk vs reward, multi-stakeholder aims).
• Constraint realization: Adversarial or stochastic satisfaction of hard constraints.

Interventions (Escalate, Abort) are treated as policies with specified utility costs, ensuring that exceptional handling paths are fully stress-tested within the same distributional framework. Constraints are operationalized as real-valued functions $g_j(\tau) \leq 0$, and governance thresholds $(\epsilon,K)$ are externally specified and auditable (Duffey, 5 Jan 2026).

4. Distributional Assessment: Alignment, Tail Risk, and Policy Dominance

MAP-AI distinguishes:

• Misalignment probability: Probability that $c_i = 1$, i.e., constraint violation or serious misalignment event.
• Tail risk: CVaR$_\alpha$ as a distributional measure of exposure to catastrophic outcomes in the worst $\alpha$ fraction of possible trajectories.
• Comparative alignment: Distributional dominance is defined: $\pi^a$ dominates $\pi^b$ if it is no worse in all risk metrics and strictly better in at least one; the governance-efficient frontier consists of all admissible non-dominated policies.

This framework supports direct, interpretable tradeoff analysis among candidate policies and transparent justification for action selection, well-suited for institutional or regulatory settings (Duffey, 5 Jan 2026).

5. Embedding Admissibility into Decision Procedures and Auditing

Admissibility-controlled action selection is formalized by a decision functional: $$D\bigl(\pi, \widehat{E}[U], \hat{p}_{\rm viol}, \mathrm{CVaR}\bigr) \in \{ \mathrm{Act}, \mathrm{Escalate}, \mathrm{Abort} \}$$ with selection privileging admissible, highest-utility policies; otherwise, escalate or abort protocols.

The proof-carrying admissibility compiler (PCAC) performs:

1. Filter policies violating hard constraints.
2. Prune to the Pareto (governance) frontier using lexicographic risk–utility order.
3. Deterministically break ties.
4. Emit a certificate with input metrics, governance spec, and dominance witnesses for auditable compliance.

This approach provides end-to-end auditability of decision-making under explicit risk and alignment constraints, decoupled from the learning system’s weight or code updates (Duffey, 5 Jan 2026).

6. Practical Implementation and Extension

Standard usage proceeds as follows:

1. Define scenario generator $G$ and value prior $P(\theta)$.
2. Specify trajectory-level utility $U(\tau;\theta)$ and constraints $C$.
3. Set Monte Carlo parameters $(N,\alpha)$ and governance thresholds $(\epsilon,K)$.
4. For each decision epoch:
   • Propose policies $\{\pi_k\}$.
   • Run MAP-AI core loop (Algorithm 1) to evaluate metrics.
   • Select and log admissible action via PCAC.
5. Continuously monitor and recalibrate $G$, $P(\theta)$, and $(\epsilon,K)$ as real-world context evolves.

MAP-AI is designed for scalable evaluation, institutional traceability, and model-agnostic deployment, providing a compositional standard for integrated alignment, risk management, and regulatory auditing (Duffey, 5 Jan 2026).

7. Broader Impact and Relationship to Classical Admissibility

Admissibility alignment builds on and generalizes the classical decision-theoretic notion of admissibility (no other policy performs uniformly better) to the setting of distributed, uncertain, and value-ambiguous AI operations. Unlike Bayes optimality under standard priors or minimax analysis, admissibility alignment is intrinsically distributional and quantitative, unifying traditional statistical safety (e.g., via tail risk) and institutional risk thresholds within executable system control-planes. The resulting architecture facilitates practical, trustworthy alignment governance for complex, high-impact AI deployments, where single-point predictions or static safety constraints are insufficient (Duffey, 5 Jan 2026).