The problem
Agents can interpret a human goal, consult tools, and propose a plan. The same fluidity creates a risk: a plan can appear authorized simply because it was expressed confidently in the conversation.
A governed mission needs a stable contract underneath that conversation: intended outcome, permitted effects, limits, evidence, approval, and stop conditions.
Four separate records
Context says what the system knows about the task. Authority says which actor may perform which effect. Evidence shows what was actually observed. The decision record binds the first three to an accountable choice.
Keeping these records distinct allows an agent to learn from a failed step without inheriting new permissions.
From design to operation
Describe hostile cases before implementation: missing identity, stale evidence, wrong recipient, changed target, partial execution, and recovery after a stop. Test each case at the effect boundary.
A design document alone does not establish runtime containment. The evaluation corpus, integrated implementation, and independent review are separate milestones. This paper is an editorial method draft; it is not a report of deployed WWAIOS behavior.