Because those sources vary widely in terms of metadata quality (author, date...etc. records), some of the converted articles look a bit messy. That's the next Hermes version's job: to get a bit wiser in terms of recovering the metadata, and in terms of robustly recovering the internal document references/citations.
Many of the original articles are missing from the converted lot: a few of them were using drawing packages, and Hermes doesn't support drawing yet, a few of them weren't converted because of an unexperienced script which filters the original lot and doesn't do an intelligent renaming of files, and about 5 of them weren't converted because Hermes needs some (small) fixes.
The total number of articles tried is 1005 (751 articles of the month 2005/3, and 254 of 2005/4); the success ratio today is 0.438.
The good news is that this version tries to push/pop parentheses at the table cell border, or around the align marker in (ams)math structures, so that the presentation-MathML result is well-formed with less need of human intervention.