Agent Sovereignty Scorecard

Agent Identification

Agent Name / ID Criticality Class Class A — Life-critical (medical, safety systems) Class B — Economic impact (business operations, data management) Class C — Low consequence (content generation, drafting)

Auditability

Decision trace available? Yes No Trace fidelity (quality/completeness of audit logs) Automated test rejection rate for AI output Human review rejection rate (if known)

Override Capability

Confirmation gates present? Yes No Gates enabled by default? No — can be disabled easily Yes — enforced by system Override disincentive score (how much using safety slows you down) Kill switch available? Yes No Kill switch activation latency (seconds)

Dependency Concentration

Single vendor dependency? Yes No — multi-vendor alternatives exist Percent of codebase AI-generated Percent of senior engineer time fixing AI output (correction tax) Alternative agent/tools available? No Yes — practical alternatives exist

Framing Asymmetry (from socrates_hemlock)

Default failure attribution to user when system fails? Yes — failures framed as user error Sometimes — case-by-case No — vendor takes responsibility