Unregulated
The model decides and the tool may run.
A proposed action can move directly from model output to tool execution, even when the action path is risky or irreversible.
Same model.Same task.Only admissible actions become real.Runtime authority for AI agents.
Admissible AI evaluates proposed AI actions before they execute or commit. It checks whether the action is authorized, supported, in scope, and safe to commit.
Admissible Runtime - demo snapshot
task: ambiguous_search unregulated_agent tool thrash: 7 reversals: 5 bad commits: 2 admissible_layer tool thrash: 1 reversals: 0 bad commits: 0
Not warnings. Runtime decisions.
Most systems treat model output like execution intent. Advisory governance can warn when something looks risky, but it does not always stop the action. Admissible checks scope violations, weak evidence, contradictions, ambiguity, stale context, retry loops, and unsafe commits before execution or commit.
Advisory governance
The system may warn, but execution can still continue.
A warning can identify the problem, but advisory checks do not necessarily stop the action before it reaches a real effect.
Admissible AI
The action only proceeds if it is admissible.
The action only proceeds if it is authorized, supported, in scope, and safe to commit.
Model proposes. Admissible evaluates. Only admissible actions proceed.