AI Vigilance continuously observes Evaluation Tags and classifies governance states against the criteria defined by applicable Policies.

An Evaluation Tag does not, in itself, indicate whether something is “allowed” or “prohibited.” Evaluation represents information about the meaning, risk, or relevance identified in relation to an observed Fact. To connect an evaluation result to an actual governance determination, it must be assessed against applicable Policies. AI Vigilance continuously monitors this relationship between Evaluation and Policy and classifies the resulting state according to the applicable conditions.

The figure represents the classification results in three states.

White / Allowed indicates a state that remains within the applicable criteria and in which normal processing may continue.

Gray / Review Required indicates a state that cannot be conclusively determined automatically due to insufficient information, ambiguity, exception conditions, conflicting rules, or similar factors. In such cases, human review or additional Evidence is required.

Black / Prohibited indicates a state in which execution is not permitted under the applicable Policy.

Importantly, this classification is not based on the AI model’s unconstrained judgment. AI Vigilance does not receive an Evaluation and stop an action merely because it “appears risky.” Rather, classification is performed on the basis of explicitly defined Criteria and applicable Policies. Accordingly, the same Evaluation Tag may result in different classifications depending on the applicable country, organization, use case, Authority, or Policy.

White / Gray / Black are not types of Governance Tags, but classifications of governance state. They are treated as a separate dimension from the information structure represented by Fact Tags, Evaluation Tags, Policy Tags, and Process Tags. This distinction makes it possible to record separately “what the information concerns” and “what governance state currently applies,” without conflating the two.

The classification result does not itself constitute the final Execution. The resulting state of Allowed / Review Required / Prohibited is passed to the subsequent Decision and Execution Boundary. AI Vigilance therefore operates between observation and evaluation on the one hand and execution control on the other, continuously connecting changing conditions to governance determinations.