1. Articles are submitted to the journal, running on OJS for example, as XHTML. This could be authored using
   • a word processor such as OpenOffice or Word
   • a browser-based document editor, perhaps online to allow collaboration
   • DocBook
   • MultiMarkdown
   • LaTeX
2. The XHTML document is made available to reviewers, who can add notes—using Marginalia—and vote on whether to accept the article in its current form.
3. If necessary, the XHTML document is revised, submitted as a new version and presented to the reviewers again, until it is acceptable for publication.
4. The XHTML document is transformed to the NLM Journal Publishing format for storage and submission to external archives.
5. XSLT transformation of the NLM XML file to semantically structured XHTML, HTML, PDF and citation export formats for reading and re-use.

Added later:

Here's some of the reasoning behind that process... Firstly, the final format for the article has to be NLM's Journal Publishing XML, because it's a very capable standard that can be transformed into all the required output formats mentioned above. The second requirement is that people are able to author articles in whatever application/document format they want. So, the article has to get from Word, OpenDocument, DocBook, MultiMarkdown, LaTeX, etc into the NLM DTD and—at the moment—the way to do that seems to be via XHTML, which all those tools can produce. Ideally, there would be a tool that could process all those formats and produce NLM-formatted XML, but that seems unlikely in the near future. More likely is that all those tools will be able to produce OpenDocument files, which might be better, but I'd say that XHTML, with constrained use of microformats-like markup, can retain just as much of the semantics of the original document as an OpenDocument file, if not more so.

So then, either by asking authors to submit XHTML or by converting their submitted documents to XHTML, the journal has one document which it needs to get reviewed. Instead of emailing it out to a few people and collating their responses manually, the review process can take place centrally, with potentially lots of reviewers (and reviewees) all able to see each other's comments (which may or may not be anonymous). Rinse and repeat until the article is accepted, convert to NLM (and make any manual adjustments if necessary) then publish to the web and/or print as required.