Delegation without abdication: HACO as a standard for mixed-initiative, human-in-the-lead design

We address a persistent ambiguity in co-creative tooling: who acts when (agency) and who decides with which guarantees (control). Synthesising mixed-initiative HCI and governance evidence, we propose Human-AI Collaboration (HACO), a taxonomy that binds five agency bands (A0-A4) to non-negotiable control guarantees and task-criticality tiers. The lifecycle lens (intent capture → generation → selection → revision → handover) turns “human in the lead” into an auditable property via paired human-centred and operational metrics: time-to-desired change (TTC), override cost, recovery success, provenance completeness, and harmful-output rates. Design patterns progressive disclosure of power, Explain→Edit links, cheap branching with fine-grained rollback, and preference plasticity convert explanations into steerable controls and create a recovery routine. For professional and regulated contexts, HACO requires signed content credentials at export and escalates safeguards for dual control and conformance testing. We specified falsifiable, band-specific thresholds (for example, ≥15% TTC gains at A2; provenance completeness ≥95% at A3; harmful output ≤0.1/100 at A4) and a study recipe combining targeted-edit tasks, survival curve reporting for TTC, and telemetry-backed soft launches. HACO aligns with established guidance on human-AI interaction and mixed-initiative design while operationalising provenance through Content Credentials (C2PA). The result is a pragmatic, testable standard for ensuring that humans remain genuinely in the lead across creative domains.