This presentation addresses the challenges posed by possible inferences of confidential information and the resulting decisions on the releasability of documents. I will discuss my efforts seeking not to enable but to prevent specific inferences by withholding only the information used to make an inference about the secret information rather than entire documents, or, as we sometimes see, withholding everything out of an abundance of caution because an inference of the secret information might be possible. My work is motivated by what has been called the "mosaicing" problem in declassification review for documents that, in the past, were withheld from public access for national security reasons. I will discuss the unique challenges posed by the overall inference prevention task as well as the special considerations that come from considering government documents. I will then describe my ongoing efforts to identify the key elements of these inferences which could be redacted, the usefulness (and limitations) of LLMs in this line of work, and the considerations for creation of specialized datasets focusing on this particular task.