The worst tragedy of this is the security implications of subtly different parsers. As your application surface increases, you're likely to mix languages (and thus different parsers), which means that the same input data will produce different output data depending on whether your parser replaces, truncates, ignores, or otherwise attempts to automatically "fix up" the data. A carefully crafted document could exploit this to trick your data storage layer into storing truncated data that elevates privileges or sets zero cost, while your access control layer that ignores or replaces the data is perfectly happy to let the bad document pass by.
And here's something else to keep you up at night: Just think of how many unintentional land mines lurk in your serialized data, waiting to blow up spectacularly (or even worse, silently) as soon as you attempt to change implementation technologies!
This is why I've been so anal about consistent decoder behavior in Concise Encoding https://github.com/kstenerud/concise-encoding/blob/master/ce...