CORRECT is the scaling step of XDS. It also re-refines the geometry of the experiment (cell parameters, crystal orientation, distance, beam direction and spindle direction) and gives all relevant statistics. The main outputs of this step are XDS_ASCII.HKL (reflection file) and CORRECT.LP (logfile), but there is a number of other files that are also produced for diagnostic purposes.
An estimate for the overall quality of an experimental setup
A single number that depends on the overall quality of an experimental setup (beam, crystal, spindle, detector, cryo, software, ...) is the upper limit of I/sigma(I) for any reflection in your dataset - even if your crystal is great, all reflections are bound to be worse than that.
This number is called I/Sigma(I)asymptotic ([Diederichs, Acta Cryst. (2010). D66, 733-740 http://dx.doi.org/10.1107/S0907444910014836])
What is that number? Scaling procedures (like the ones used in XDS, SCALA, SCALEPACK and DSCALEAVERAGE) scale (or rather, inflate) the variances of individual observations such that they match the experimental spread of symmetry-related observations. To this end, two contributions to the variance v(I) of a reflection are considered: the first component is random error, and the other component is systematic error. The two values a and b appearing in the variance-scaling formula v(I)=a*(v0(I)+b*I2) are printed out by CORRECT. a scales the random error component, a*b scales the systematic error component. For strong and well-measured reflections, the variance is dominated by the systematic error a*b* I2 that is introduced by any beam /spindle / detector /cryo or other instability or malfunction. For weak reflections, a*v0(I), the variance from counting statistics, dominates.
Versions of XDS since May 10, 2010 print out I/Sigma(I)asymptotic = 1/√(a*b) as "ISa". ISa is the I/sigma of an infinitely strong reflection. If there were no systematic error, ISa would be infinite. In the presence of systematic error, ISa is finite and is the upper limit of I/sigma of any observation in your dataset.
As you can see from the formula, low values of a and b are good in the sense that a high upper limit of I/sigma(I) results. If e.g. the crystal is badly split or broken, or reflections are too close on the detector, or the data reduction is not good (wrong parameters), then the values of a and b are elevated.
If your crystal is good, then a and b will reflect the quality of the other components of the experimental setup (e.g. beamline stability).
ISa is well suited to judge the quality of the experimental setup, because its value does not depend on random error, whereas the low-resolution Rmeas does, and is thus influenced by crystal size and exposure. If you see a high value of the low-resolution Rmeas, you don't know if it is high because the crystal diffracted weakly, or because the beamline was broken. Conversely, a low value of ISa indicates that something is broken, no matter how small the crystal is or how weakly it was exposed.
In practice, both crystal quality and beamline quality limit the value of I/Sigma(I)asymptotic. A good crystal (even with elevated mosaicity and medium resolution) should give a high value on a good beamline.
I have seen values around 15-20 for good crystals that allowed me to solve a MAD structure, but that required high multiplicity of observations. Values around 30 allowed me to solve a sulfur-SAD structure at medium resolution (diffraction to 2.3 A, anomalous signal to 3 A). I have also seen a value around 40 for Z. Dauter's 0.98A Proteinase K (2ID8) sulfur-SAD data from the APS/22-ID beamline, and recently even higher values were obtained at the SLS, beamline X06SA, with a Pilatus detector.
On the other hand, I have sometimes obtained values less than 10 with good test crystals, clearly indicating strong systematic errors. It is always good to discuss this with the people who are responsible for the beamline. They might know what is broken, or might be able to find out what went wrong.
A low I/Sigma(I)asymptotic may be compensated by high multiplicity, at the expense of radiation damage. Conversely, high multiplicity is not needed to solve a structure, if the data have a high I/Sigma(I)asymptotic. For molecular replacement and refinement, a high value of I/Sigma(I)asymptotic is not strictly needed (but the maps are better with better data, and the model R values lower!).