My guess would be that the PDF files you are using is somewhat different. Is it possible that a regularly scheduled PDF is slightly different during each run? I don't know, but I would think so.
Is it possible to view the origonal report you created the model with? Does it work correctly?
I appologize, I'm pretty naive when it comes to PDF, but my first guess is that your PDF input file is what changed.
I think Joey has likely identified the problem. PDF files can be generated by a number of different 'pdf writers' and each can presetn slightly different challenges. In addition the format itself can be very variable when reverting to text. That is not to say there are problems - just that the nature of the work can be variable!
I have seen situations where the 'write engine' for reports has changed from one report to the next for what otherwise appears to be a version of the same report. It can introduce variations that may lead to a different interpretation by Monarch (or indeed the convert to text feature in the PDF Reader program) to produce different output.
One possibly approach to the problem might be to develop the model so that it can deal with variable results should columns be shifted left or right.
The best way to attempt this may not be obvious and will depend on the potential differences that might appear from one version of the report to the next.
If the report will allow it the opportunity to make fields wider so that shifts are still mapped might be the simplest solution. If that does not look like it will work let us know and alternative suggestions will, I am sure, be posted.