Right-to-Act: A Pre-Execution Legitimacy Protocol

• Right-to-Act is a pre-execution non-compensatory protocol that grants AI outputs the intrinsic right to be executed only if all required legitimacy constraints are met.
• It operates as a deterministic decision layer placed between AI decision generation and execution, independently verifying a structured set of conditions.
• The protocol enhances reversibility, auditability, and error prevention by blocking any action that fails a hard, non-compensable legitimacy test.

Right-to-Act is a pre-execution non-compensatory decision protocol for AI systems that determines whether an AI-generated decision is permitted to be realized at all. In its most technical formulation, it is a deterministic decision layer placed between decision generation and downstream validation, certification, or execution, and it treats legitimacy as satisfaction of required structural constraints rather than attainment of a favorable aggregate score. The concept was introduced to address what is termed the Pre-Action Legitimacy Gap: the absence of a formal layer answering the question “Should this specific decision be allowed to proceed at all?” before ordinary safety, compliance, or authorization checks (Lavi, 27 Apr 2026).

1. Conceptual basis and architectural placement

The Right-to-Act protocol is motivated by the observation that AI systems increasingly produce outputs that directly trigger real-world actions. Existing governance and safety layers typically ask whether a user is authorized, whether an output is harmful, whether a path is compliant, or whether behavior remains within certified risk bounds. The Right-to-Act proposal isolates a prior question: whether a specific decision has the intrinsic right to exist as an executable event. The lack of this prior legitimacy test is identified as the Pre-Action Legitimacy Gap (Lavi, 27 Apr 2026).

Architecturally, Right-to-Act is explicitly pre-execution. It is neither a post-hoc review nor a runtime audit. Its placement is immediately before execution and before later validation or certification stages. In the formulation given in the protocol paper, the decision flow is:

AI Decision $\rightarrow$ Right-to-Act $\rightarrow$ Validation/Certification $\rightarrow$ Execution,

with failure at the Right-to-Act boundary resulting in a non-execution outcome rather than conditional execution. This location is central to the protocol’s purpose: it governs admissibility, not performance quality, risk scoring, or after-the-fact accountability (Lavi, 27 Apr 2026).

The protocol is also defined as model-independent. It operates on structured decisions and is stated to be agnostic to model architecture, model type, training data, and proprietary logic. This suggests a protocol-level abstraction that can be superimposed on heterogeneous AI or agentic systems without requiring changes to the underlying inference mechanism (Lavi, 27 Apr 2026).

2. Formal decision structure

An executable AI decision is represented abstractly as

$a = (x, c, t, s, \tau)$

where $x$ is the operation, $c$ the context, $t$ the target, $s$ the scope, and $\tau$ the timing. The paper states that this tuple is illustrative rather than exhaustive, but it serves to formalize the idea that admissibility is assessed over a structured action candidate rather than an undifferentiated model output (Lavi, 27 Apr 2026).

Let

$C = \{C_1, \ldots, C_n\}$

be the set of required legitimacy constraints, with each constraint defined as

$\rightarrow$0

Thus, $\rightarrow$1 when the $\rightarrow$2-th required condition holds for decision $\rightarrow$3, and $\rightarrow$4 otherwise. The Right-to-Act decision function is then

$\rightarrow$5

The corresponding feasible region is

$\rightarrow$6

so that

$\rightarrow$7

and, equivalently,

$\rightarrow$8

This formalization is the core of the protocol. Legitimacy is treated as membership in a feasible set, not as success on a scalar score or threshold. The output is deterministic: once the constraints are specified and evaluated, the result is unambiguous. The non-allow outcome is denoted NON, and may be operationalized as DEFER, ESCALATE, or REQUEST more info (Lavi, 27 Apr 2026).

3. Non-compensatory logic and its distinction from score-based systems

The defining property of Right-to-Act is its non-compensatory character. If any required legitimacy constraint fails, execution is not permitted, and no favorable signal elsewhere can compensate. High model confidence, low estimated risk, or strong evidence on other dimensions does not override a failed required check. In the paper’s formulation,

$\rightarrow$9

This is contrasted with a compensatory scoring regime of the form

$\rightarrow$0

with $\rightarrow$1 and $\rightarrow$2, where execution is allowed when

$\rightarrow$3

In such a system, poor performance on one factor can be offset by strength in others. Right-to-Act rejects this logic for required legitimacy constraints. The paper’s Non-Equivalence Theorem states that compensatory scoring systems cannot guarantee the non-compensatory Right-to-Act property unless they are forcibly reduced to hard constraints. The accompanying corollary states that no choice of weights can repair the problem unless the scoring system is replaced by explicit hard constraints (Lavi, 27 Apr 2026).

A recurrent misconception addressed by this distinction is that sufficiently conservative risk scores or calibration can simulate strict admissibility rules. The protocol argues otherwise. Thresholded aggregate scores may still authorize execution in cases where a required structural condition fails. By construction, Right-to-Act does not allow this. Its central maxim is therefore that legitimacy is a boundary, not a score (Lavi, 27 Apr 2026).

4. Pre-execution legitimacy boundary and operational consequences

The pre-execution legitimacy boundary is the set of decisions satisfying every required structural constraint. The paper describes this as a hard boundary: only decisions inside the feasible region acquire the right to be executed. Outside that region, execution is blocked, deferred, or escalated. The geometric contrast given in the paper is between threshold-above-line score systems and feasibility-based systems that permit only the intersection of all required conditions (Lavi, 27 Apr 2026).

A scenario-based case study concerns AI-driven account suspension on an online platform. In the compensatory baseline, repeated flags, similarity to malicious patterns, high model confidence, and prior suspicious activity are aggregated into a score

$\rightarrow$4

and suspension occurs if $\rightarrow$5. The paper emphasizes that this can yield suspension even when direct contextual verification is absent. Under a Right-to-Act boundary, insufficient contextual verification makes the outcome NON rather than suspension, with the system returning a reversible non-action such as escalation or a request for additional information (Lavi, 27 Apr 2026).

This case study is used to illustrate three operational consequences. First, reversibility is preserved: irreversible or high-impact actions are withheld until all required conditions are met. Second, auditability is improved because the failed requirement can be recorded explicitly rather than buried inside a low or borderline aggregate score. Third, error prevention is reframed structurally: the protocol is intended to prevent catastrophic or unfair outcomes caused by missing critical information even when overall signals look favorable (Lavi, 27 Apr 2026).

The minimal rule can be expressed procedurally as:

$\rightarrow$6

This algorithm is deliberately austere. Its significance lies not in computational novelty but in the insistence that the admissibility decision is logically prior to ordinary execution and cannot be softened by favorable collateral evidence (Lavi, 27 Apr 2026).

5. Governance significance and rights-based extensions

Within AI governance, Right-to-Act is presented as a mechanism for explicit, testable pre-execution legitimacy. Because it is independent of model internals, the protocol is described as integrable with multiple systems through explicit representation of required constraints. The paper also notes a recursive data scenario: if AI-generated content feeds back into training or evaluation data, statistical certification may be distorted, whereas a Right-to-Act boundary remains decision-local and constraint-based (Lavi, 27 Apr 2026).

In adjacent governance literature, the phrase “Right-to-Act” is also used in a broader, stakeholder-centered sense. The HEAT framework for the EU AI Act defines the Right-to-Act as the ability for stakeholders, including directly affected individuals and organizations, to meaningfully engage, intervene, or contest AI systems’ operation and decisions. HEAT operationalizes this through stakeholder engagement, complaint and redress mechanisms, affirmative and revocable consent practices, accessibility and design justice, and collective deliberation grounded in distributed expertise (Hollanek et al., 24 Apr 2025).

A related rights-based analysis of the EU AI Act argues that fundamental rights function as legal thresholds and procedural triggers across the AI lifecycle. On that account, prohibited practices, high-risk classification, transparency obligations, human oversight, Fundamental Rights Impact Assessments, and the right to explanation collectively create mechanisms of empowerment, contestation, and redress. This is not identical to the protocolic Right-to-Act of pre-execution admissibility, but it is normatively aligned with the idea that execution of AI-mediated decisions is conditioned by prior legal and procedural thresholds rather than left entirely to ex post review (Pavlidis, 24 Mar 2026).

This broader governance literature suggests an important distinction. In the protocol paper, Right-to-Act denotes a machine-interpretable admissibility boundary over decisions. In the policy and justice literature, it can denote human and institutional capacities to challenge, shape, or halt those decisions. The two senses are compatible but not interchangeable.

6. Related concepts and distinct technical usages

Several recent works employ neighboring concepts that clarify the semantic range of “Right-to-Act.” In multi-agent enterprise architectures, RP-ReAct describes a system in which a Reasoner Planner Agent holds the sole authority to decide what should be done, while Proxy-Execution Agent(s) may act only on explicit delegated sub-questions. There, the “right to act” is implemented through tagged delegation boundaries and role separation: the executor cannot autonomously initiate new plans or exceed the granted remit. This usage concerns authority delegation inside an agent architecture rather than legitimacy screening of a decision prior to execution (Molinari et al., 3 Dec 2025).

A complementary but distinct notion is the Right to History, defined as the entitlement to a complete, verifiable, and tamper-evident record of every AI agent action on an individual’s own hardware. That work formalizes append-only logging, completeness, integrity, boundary enforcement, and energy conservation, and implements them in a sovereignty kernel with Merkle audit logs and human approval. Relative to Right-to-Act, Right to History governs verifiable post-action traceability rather than pre-action admissibility. The two are therefore orthogonal: one determines whether an act may proceed, the other ensures that an executed act is cryptographically reconstructible (Zhang, 23 Feb 2026).

A further neighboring discussion concerns the legal demand for technical interpretability. Work on the AI Act proposal argues that GDPR, Convention 108+, and the AI Act move toward stronger transparency and interpretation requirements, but do not yet establish an enforceable right to technical interpretability. This matters because insufficient understanding of reasoning can impede objection, complaint, and redress. A plausible implication is that interpretability-oriented rights and Right-to-Act constraints can be mutually reinforcing: one supports comprehension and contestation, the other governs admissibility at the point of execution (Gallese, 2023).

Taken together, these strands indicate that “Right-to-Act” has become a polysemous term in current AI discourse. Its most exact technical meaning is the deterministic, non-compensatory pre-execution protocol introduced in 2026 (Lavi, 27 Apr 2026). Surrounding literature extends the phrase toward stakeholder agency, delegated execution authority, and rights-preserving governance. The common thread is a shift away from treating AI outputs as presumptively executable and toward treating action as something that must first be justified, bounded, and, where necessary, contestable.