From Vibes to Verifiability: Review-Gated Automation for Supporting Situated Judgment in Post-Generation Creative Systems

Abstract

Generative AI has made aesthetically plausible creative proposals inexpensive, but it has not removed the work required tomake those proposals semantically coherent, accessible, culturally appropriate, selectively repairable, and suitable for release and explicitly authorized to progress. This paper argues that a central human–computer interaction problem now lies after generation: how automation should prepare evidence for accountable judgment, and how that judgment should control production. This paper introducesreview-gated automation, a model in which a system emits an evidence-bearing candidate while a decision bound to an exact version and scope determines whether work advances, returns for a targeted correction, or remains blocked. The model is examined through a deterministic animated-sticker compiler developed through Research through Design, and through four purposeful production-trace families. A longitudinal Pace trace makes the state transitions explicit. An initial engineering-valid “3D” proof was rejected for art direction and retained only as enhanced 2.5D; exactly six native Blender/GLB proofs were then generated and reviewed; approval of that visual language authorized only the remaining eighteen translations; and the resulting 24-sticker candidate was submitted for another review while platform export and public release remained absent. Other traces show technically valid assets failing because of phrase occlusion, weak edition differentiation, identity drift, or incorrect assistive-device topology. The case demonstrates pipeline-level, fail-closed progression and selective preservation through explicit scopes, locked source records, parity hashes, evidence boards, and owner decisions. It does not show that human judgment is infallible, that the interface has been validated with external creators, or that the system establishes cultural legitimacy. Finally, implications are derived for post-generation HCI centered on repairability, evidence-bearing candidates, authenticated and consequential decisions, risk-based review, culturally situated authority, and the operational power to refuse release.