Within CLASS, an observation should represent a single observing configuration, e.g. a single direction observed at a single central frequency with a single spectral resolution and in one polarization only (i.e. a single sky position, front-end and back-end combination). Each observation is given a number, named observation number at the time of creation of the CLASS data file. This number is then carried out in further manipulations.
Several version of a given observation may be stored in the same data file. Each version of a given observation thus represent different stages of the data reduction and all the versions of a given observation gives the history of the data processing. Each version of an observation is given a number (starting at 1), named version number at the time of creation of the CLASS data file. The version number increases automatically, each time the observation is modified (using WRITE). By default, only the last version of a given observation is relevant, i.e. GET reads the last version of an observation. It is possible (but not recommended) to store an observation without incrementing the version number with the UPDATE command so that you can go back to previous stages of reduction in case of big mistakes.
Provided you respect this use of the version number, data reduction can be largely automated. Failing to do this, i.e. using the same observation number for very different things at time of creation of the CLASS data file, implies that you have to remember yourself which version corresponds to which configuration.