First, the format must be portable between machines. By ``portable,'' we mean simply that one should be able to read the file, not necessarily make sense of its contents (that falls under ``understandability,'' discussed below).
Ideally, portability would require that the dataset be written in terms of some explicitly portable binary data representation, such as the eXternal Data Representation (XDR) promoted by Sun Microsystems, Inc. Data could then be moved from machine to machine without any need for conversion.
Even if one relaxes this condition and permits native binary representations, though, portability also requires that a standard format cannot rely on special file types or file record formats which are unique to a single type of machine. That is, the concepts behind data structures must be uniform across machines, even if the implementation of those concepts is permitted to vary from machine to machine. For example, all machines used in scientific data processing have integer and floating-point data types, however differently they may be implemented. Few machines, on the other hand, have special representations for irrational numbers; one should then exclude irrational numbers from the data types permitted in the file.
In fact, because of the wide variation in the capabilities of machine operating systems (from personal computers to supercomputers), portability considerations require that any data file format specification be limited to the information written into a file, and not concern itself with how files are named or how they are arranged and manipulated within a data system. The format should not even depend on the concept of ``file;'' it should be possible to implement it using any sort of bit stream.