Jev/diffusion/similar models seem deeply useful for redaction. The

Jev/diffusion/similar models seem deeply useful for redaction. The trick with redaction is that you have the entire thing in context, so you don't need causal masking.