When the Same Advisory Rule Produces Different Decisions

TL;DR for operators

An AI agent’s proposed action, the decision to accept it, and the action eventually executed are three different events. A new shared-autonomy framework makes those stages explicit through reusable evaluation mechanisms while retaining standard Markov decision process (MDP) planning.

The formal contribution is a structured representation of interaction rules, not a more expressive decision model. In a synthetic navigation task, changing human attention alters the optimal recommendation at one bridge approach, even though the interaction remains advisory. All 20 reported state-attention configurations show positive modeled marginal value for their reference recommendations relative to silence.

For teams designing AI approval workflows, the architectural implication is to separate proposal generation, authorization or advisory evaluation, and actual execution. This could make common approval and fallback rules easier to maintain and inspect. However, the paper neither measures those maintenance benefits nor validates its human acceptance model in deployment.

The recommendation changes even when the interaction remains advice

An AI assistant recommends an action. A human can accept it, reject it, or take another action. The recommendation therefore does not determine the outcome on its own. What matters is how the proposal will be evaluated, what happens after rejection, and which action ultimately gets executed.

These details can change which recommendation is worth making, even when the human retains the same decision authority.

Consider a simulated navigation problem with several bridges. At one bridge approach, an AI recommends moving left when modeled human attention is very low, low, or medium. When attention becomes high or very high, the optimal recommendation switches to moving down. The human remains responsible for evaluating the suggestion throughout.

Shashank Shekhar, Laurent Jeanpierre, and Abdel-Illah Mouaddib examine this problem in From Proposal to Execution: Modeling Interaction Semantics in Shared Autonomy.¹ Their framework addresses a modeling issue: conventional MDPs can already represent acceptance, rejection, and execution through transition probabilities, but those probabilities may conceal the evaluation rules shared across different interactions.

The authors propose making those rules explicit instead of repeatedly embedding them in individual transition specifications.

Separate what an action means from what gets executed

The framework distinguishes three stages:

1. Proposal: Agents select interaction-level actions, such as advising, requesting authorization, executing autonomously, or remaining passive.
2. Evaluation: The interaction is interpreted under the current autonomy and attention context, selecting a semantic outcome and its associated evaluation mechanism.
3. Execution: The mechanism resolves the interaction into an executed action, potentially through stochastic acceptance, rejection, or fallback, before inducing task and context transitions.

An interpretation function determines which semantic outcome applies to a joint interaction under its context. It can also classify an interaction as invalid.

Importantly, the interpretation function does not directly execute a task action. It selects the evaluation principle governing how execution will be resolved.

The framework includes teleoperation, full autonomy, advisory interaction, authorization, and recovery as distinct semantic outcomes. These represent different allocations of control, evaluation, and execution responsibility.

In the advisory case, a proposed action $a_c$ is accepted with modeled probability $\rho(s,a_c)$. If rejected, execution follows the human fallback policy $\pi_j$:

The human evaluation model uses one-step lookahead, an approximation of future task value, and an attention-dependent effective discount factor. A stochastic decision rule then produces the acceptance probability.

This means that an AI’s expected advisory value depends on both acceptance and the consequences of the alternative action taken after rejection. Maximizing acceptance probability alone is not necessarily equivalent to maximizing decision value.

Reusable evaluation rules do not require identical outcomes

The framework’s first formal result, Proposition 1, groups valid joint action-context pairs into semantic regimes.

Two interactions belong to the same regime when the interpretation function assigns them the same semantic outcome. They can consequently use the same outcome-indexed evaluation mechanism.

For example, two advisory interactions may share the same general acceptance-and-fallback mechanism. Their actual acceptance probabilities, executed actions, rewards, and transitions can still differ because their task states, candidate actions, and attention contexts differ.

That separation permits a common evaluation principle to be specified once while retaining state-dependent execution dynamics.

The authors argue that this organization reduces redundant specification when multiple interactions share a mechanism. They do not quantify reductions in engineering effort, and reuse offers no guaranteed reduction if every interaction requires its own semantic outcome.

Proposition 2 establishes another important boundary: the semantics-aware process induces a conventional MDP when interpretation and induced dynamics depend only on the current extended state and admissible joint action.

The extended state includes both the task state and interaction context. Standard Bellman planning remains applicable.

Consequently, the framework does not establish greater expressive power or lower computational planning complexity than a suitably factored conventional MDP. Its contribution is an explicit interface for organizing interaction semantics.

Twenty simulated configurations show value, but only one policy switch

The evaluation uses an author-constructed stochastic 8-by-9 navigation grid with three bridges and irreversible corridor-selection decisions.

The reported experiment fixes the advisory interaction regime and examines four representative task states across five attention levels, producing 20 configurations.

The following results summarize Tables 3–6.

Source: Tables 3–6, discussed in Section 7.1. The marginal value $\Delta Q$ compares a fixed reference candidate action with no advice. The acceptance values and decision values are model-generated, not observations of human participants.

At the start state, acceptance changes by only 0.058 between the lowest and highest attention levels, yet the modeled advantage of the reference recommendation remains positive and increases. This illustrates how decision value can change without a large change in acceptance probability.

Above bridge B1, acceptance rises substantially, reaching 0.922 at high attention before declining slightly to 0.916 at very high attention. Higher attention therefore does not guarantee strictly increasing acceptance for every state.

The B2 approach offers the clearest policy result. The optimal advice switches from left to down between medium and high attention. The optimal modeled state value rises from 3.468 to 14.043 between the endpoint attention levels.

There is a reporting distinction here: Table 5’s marginal value is calculated for the fixed reference action down, even when left is optimal at lower attention levels. Its $\Delta Q$ must not be interpreted as the advantage of the optimal action in those rows.

These findings are main simulation evidence for context-dependent advisory planning. They are not an ablation demonstrating that explicit semantic modeling outperforms an equivalently structured MDP.

Acceptance has value only relative to execution consequences

The paper also provides a stylized three-bridge analytical example with different human and robot rewards, discount factors, and delayed payoffs.

Under its specified valuation scheme, acceptance of the proposed bridge action gives approximately 132 in value, while rejection and fallback give approximately 16.6. The full-autonomy reference is approximately 120.

Expected advisory value is therefore:

In this illustration, advisory exceeds the full-autonomy reference when acceptance reaches approximately 0.90.

That threshold is specific to the assumed rewards, fallback behavior, and discounting. It is not a general human acceptance requirement.

The analytical example clarifies why the value of a proposal cannot be assessed independently of the action executed after its evaluation.

Approval workflows can expose the same separation

Cognaptus inference: For an enterprise AI agent that recommends or initiates consequential actions, the framework suggests representing three workflow components explicitly: the proposed action, the authorization or advisory decision, and the action actually executed.

For instance, an operations team designing an automated payment-approval process could specify a common evaluation mechanism for proposals requiring approval, while preserving task-specific authority, rejection, and fallback behavior.

This structure could make changes to common approval policies easier to propagate and make execution decisions more inspectable. It could also help distinguish an agent’s intention from the action eventually committed to an external system.

These are architectural possibilities, not measured outcomes from the paper. Whether a semantic interface improves maintainability depends on how much evaluation logic is shared, how existing workflow engines already represent that logic, and how difficult application-specific mechanisms are to specify.

A production comparison would need to measure specification effort, policy-change effort, execution correctness, and system performance against a suitably factored existing architecture.

Formal compatibility is established; deployment benefit remains untested

The strongest results are the semantic-regime partition and the conditional proof of MDP compatibility.

The simulation demonstrates coherent behavior under one specified advisory evaluation model. It does not include observed human acceptance, real robotic deployment, learned evaluation mechanisms, or independent empirical replication.

Attention and autonomy context are assumed observable. Human evaluation relies on stipulated parameters, one-step lookahead, and a fixed future-value approximation. More complex human deliberation and partially observable interaction contexts remain outside the tested formulation.

Nor does explicit authorization or recovery modeling establish safety or compliance by itself.

For a team evaluating this approach, the unresolved tests are whether real evaluation behavior can be calibrated, whether decisions remain robust to context-estimation errors, and whether semantic reuse offers measurable advantages over existing implementations.

A modeling interface worth testing, not a new planning frontier

The central lesson is that an AI proposal and an executed action should remain distinguishable in the decision model. Evaluation determines the relationship between them, and recurring evaluation principles can be represented explicitly without replacing conventional planning.

The demonstrated result is structural clarity and compatible optimization, illustrated by context-dependent decisions in a synthetic environment.

For operational AI systems with recurring authorization, advice, and fallback pathways, this provides a concrete architecture to compare against existing designs. Its production value will depend on measured maintainability, calibrated interaction behavior, and reliable execution—not on the presence of a semantic layer alone.

Cognaptus: Automate the Present, Incubate the Future.