Disaster-management AI is commonly governed before deployment through ethics reviews, data-sharing agreements, and standards checklists. This paper argues that design-time governance is necessary but incomplete because decisive failures occur while systems run: a prediction must be overridden, a model rolled back, a decision audited, an incident escalated, or a system kept safe when connectivity degrades. We reframe governance as a runtime dependency and synthesize a purposively screened 80-record set of standards, policy documents, surveys, primary studies, and technical sources. Ten runtime controls are coded through 232 non-exclusive source-control assignments. Coverage averages 23.2 sources per control (median 21.5; range 6 to 43); the five most-covered controls account for 80.2% of assignments, whereas the five least-covered controls account for 19.8%. A six-family standards crosswalk contains support in 31 of 60 cells (51.7%), with each control represented by two to four instrument families and no family covering all controls. Auditability, co-production, and human override are prominent, while ownership, rollback, incident reporting, and degraded-mode operation remain comparatively sparse. The contribution is an Operational AI Readiness Matrix with pass/fail questions, evidence requirements, and four maturity levels. An analytical evaluation checks completeness, traceability, maturity discrimination, non-compensation, and scenario coverage, while explicitly recognizing that the matrix has not yet been validated against deployment outcomes. The result is a practical instrument for testing whether disaster-AI systems remain governable during drills and live events, not merely documented before use.

Runtime governance; disaster AI; AI readiness matrix; auditability; human override; model rollback; degraded-mode operation; standards crosswalk