EmDash uses Clef to moderate the plugin registry

EmDash now screens every plugin listing in its registry with Cloudflare’s Clef decision model, which evaluates both text and images. Clef inspects listing metadata, outbound links (evaluated by link text and destination), icons, and screenshots for categories such as explicit sexual content, hateful or dehumanizing content, graphic violence, phishing or credential requests, impersonation, scams, spam, deceptive links or misleading media, and attempts to manipulate moderation. The model returns probabilities for each category; any score at or above 0.45 sends the listing to a human operator for review, and Clef never has unilateral blocking authority. Because releases are signed and published by authors via AT Protocol, automated screening lets EmDash preserve an open publishing model while controlling what appears in the public catalog.

Clef was chosen after comparisons with the previous multi-model pipeline and alternatives; it is multimodal, faster, and required fewer components. In regression testing it produced consistent pass-or-review decisions across three runs of 21 public text fixtures (63/63) with no invalid outputs and a p95 end-to-end text moderation latency of 1.64 seconds. It also passed EmDash’s protected corpus and the live registry samples (37 plugin profiles and 43 images). Replacing the earlier bundle simplifies implementation and inspection while maintaining a fail-closed human review path; the Clef adapter and question set are available on GitHub, and publishing via the EmDash CLI is unchanged.