Large language models (LLMs) have strong reasoning capabilities, but current architectures often connect their outputs directly to external execution environments such as APIs, databases, and business workflows. In such structures, safety depends too heavily on the correctness of the model’s inferred intent. If missing conditions, ambiguous intent, conflicting constraints, or unverified authorization boundaries remain unresolved, the system may proceed to erroneous execution, inappropriate operations, information leakage, or sequential risk formation. This paper proposes the Intent Structure Control Layer (ISCL) as a front-end control layer that structurally separates natural language reasoning from execution. ISCL shifts the safety boundary from post-hoc output inspection to pre-execution intent validation. It represents intent extracted from natural language in the form of object. method (args): responseₛchema where conditions and evaluates its execution eligibility using the three-way decision set execute, block, clarify. This allows incomplete condition states to be handled explicitly, enables clarification when required information is missing, and prevents dangerous or inconsistent intents from reaching execution systems. The paper further applies the condition-preservation problem identified in prior work on Constraint-Bound Queries (CBQ) to control-layer design. In ISCL, conditions are treated not as independently removable constraints, but as a condition bundle whose completeness must be verified before execution. In addition, by treating not only single intents but also session-level Intent Trajectories as evaluation targets, ISCL can detect latent risks formed by sequences of individually acceptable operations. ISCL does not restrict the reasoning capability of LLMs. Rather, by structuring and validating intent before execution, it provides a safer connection basis for heterogeneous execution systems, including LLMs, SQL, APIs, and business logic.
Yuji Takahashi (2026) studied this question.