Right-to-Act: A Pre-Execution Non-Compensatory Decision Protocol for AI Systems

1. Problem Statement: The Pre-Action Legitimacy Gap

As AI systems evolve from advisory tools to autonomous agents capable of executing real-world actions (e.g., fund transfers, account suspensions, infrastructure changes), current architectural paradigms face a critical gap. Existing mechanisms—such as authorization, safety alignment, policy enforcement, and statistical certification—evaluate who is allowed to act, if the output is harmful, or if the system behaves within risk bounds.

However, these mechanisms fail to answer a distinct, prior question: "Should this specific AI-generated decision be allowed to exist as an executable event at all?"

The paper defines the Pre-Action Legitimacy Gap as the absence of a formal mechanism to determine whether a decision has the "right to act" before it enters the execution pipeline. A decision may be authorized, safe, and low-risk, yet still be structurally illegitimate due to missing contextual verification, insufficient justification, or premature timing. Current systems often rely on compensatory scoring, where a high confidence score or strong policy match can offset a missing critical condition, leading to potentially irreversible errors.

2. Methodology and Formal Model

The author proposes a minimal, formal framework called the Right-to-Act protocol. This approach treats legitimacy not as a probabilistic score to be optimized, but as a feasibility condition defined by hard constraints.

2.1 Formal Definitions

• Decision Space ($A$): A set of candidate executable decisions $a$, represented as structured tuples containing operation, context, target, scope, and timing.
• Legitimacy Constraints ($C$): A finite set of deterministic predicates $\{C_1, C_2, ..., C_n\}$ where $C_i(a) \in \{0, 1\}$. A value of 1 indicates the condition is met; 0 indicates failure.
• Feasible Region ($F$): The set of all decisions where all required constraints are satisfied:
  $$F = \{a \in A : \forall i, C_i(a) = 1\}$$
• Right-to-Act Function ($D$): A deterministic decision function:
  $$D(a) = \begin{cases} \text{ALLOW} & \text{if } a \in F \\ \text{NON-ACTION} & \text{otherwise} \end{cases}$$

2.2 Non-Compensatory Logic

The core methodological innovation is the non-compensatory nature of the boundary. In this system, if any single required constraint fails ($C_j(a) = 0$), the decision is rejected, regardless of how many other constraints are satisfied or how high the confidence score is.

• Outcome: "Non-Action" is not treated as a system error but as a valid, auditable state (e.g., DEFER, ESCALATE, or REQUEST-INFO).

3. Key Contributions

The paper makes four primary contributions:

1. Problem Definition: It formally identifies the "Pre-Action Legitimacy Gap" as a distinct architectural primitive separate from authorization, safety, or runtime governance.
2. Formalization: It introduces a minimal mathematical model for a Right-to-Act boundary that enforces structural feasibility rather than optimizing scores.
3. Non-Equivalence Theorem: It proves that compensatory scoring systems cannot guarantee pre-action legitimacy.
   • Theorem: If a system uses a weighted sum $S(a) = \sum w_i x_i(a)$ with a threshold $\theta$, it is possible to construct a decision where a critical constraint fails ($C_k=0$) but the aggregate score still exceeds $\theta$ due to other high signals. Thus, a scoring system cannot replicate the hard-blocking behavior of a non-compensatory constraint system.
4. Case Study Application: It demonstrates the practical impact of this boundary in an AI-driven account suspension scenario.

4. Results and Case Study Analysis

The paper validates the framework using a case study of AI-driven account suspension.

• Scenario: An AI system proposes disabling a user account based on high confidence, pattern similarity, and prior flags.
• Compensatory Baseline: A standard risk-scoring system calculates a high aggregate score ($S(a) \ge \theta$) and immediately executes the suspension.
• Right-to-Act Intervention: The system checks for a specific structural constraint: Contextual Verification (e.g., ensuring the activity isn't due to a compromised device or false correlation).
  • In the case study, this constraint fails ($C_{context} = 0$).
  • Result: The Right-to-Act boundary blocks execution ($D(a) = \text{NON-ACTION}$), triggering an escalation or request for more info, whereas the baseline system would have caused an unjustified suspension.

Key Finding: A decision can be statistically justified and policy-compliant yet structurally illegitimate. The non-compensatory boundary prevents this by treating missing conditions as hard blockers rather than score deductions.

5. Significance and Implications

The paper argues that shifting from "scoring" to "boundary enforcement" fundamentally changes AI governance:

• Inaction as a First-Class Outcome: Non-execution is redefined as a valid, controlled operational state rather than a failure mode.
• Improved Auditability: Because decisions are blocked by specific failed constraints rather than vague low scores, audit trails become clearer and more defensible.
• Model Independence: The framework applies to any AI system (LLMs, agents, classifiers) as long as the decision can be represented structurally, making it compatible with heterogeneous environments.
• Mitigation of Recursive Data Risks: In domains where AI generates data that feeds back into training (recursive data), statistical certification may become self-referential. A pre-action constraint boundary enforces structural validity on individual decisions, providing a safeguard against certification of distorted distributions.

Conclusion

The paper concludes that legitimacy is a boundary, not a score. By introducing a deterministic, non-compensatory decision primitive, the Right-to-Act protocol fills a critical gap in AI safety architecture, ensuring that decisions only proceed to execution when they satisfy all necessary structural conditions, regardless of their aggregate confidence or risk profile.