When default settings are used, the precision of protein calculations is in the
order of 1 kcal.mol^{-1}. Note that this is precision, not
accuracy; errors in accuracy can be in the order of hundreds of kcal.mol^{-1}.
Precision is a measure of how reproducible a given calculation is.
Obviously, if the calculation was completely precise, the same job would give
the same results on all operating systems, and, more important, give the same
results when insignificant changes are made to the starting data. In
practice, small changes in input data can give rise to significant changes in
the results. This is obviously not acceptable when comparing relative
energies of closely-related systems.

Several techniques and strategies are available to increase precision.
Using these, particularly minimizing the domain to be optimized, can result in a
reduction in errors in precision to about 0.02 to 0.05 kcal.mol^{-1}.

When making a starting model, geometry optimization should be run until there is
no significant change in DH_{f}. If this is not done, then when
changes are made to the geometry, a mutation or a step in a reaction mechanism,
and the geometry optimized, the original and changed systems might optimize to
significantly different final geometries, even if the change is relatively
small. This can happen if a hydrogen bond that should exist forms during
one calculation and not during the other. A suitable strategy would be to
use keywords `GNORM=3 T=2W`, and monitor the optimization periodically. When
the DH_{f} stops dropping, use the
SHUT command to stop the run. When the job ends,
edit the output file, <name>.out, to make a new data set that uses the last
geometry output. This geometry represents the structure with the lowest
calculated DH_{f}, not necessarily the
last structure calculated.

When comparing closely-related systems, start with one of the two systems, optimize it as completely as possible, then make the smallest change possible to make the other system. The idea here is that when the geometry optimization of the changed system starts, the geometry should (a) be as similar as possible to the original system, and (b) be as near as possible to the final geometry. For example, if a hydrogen atom is being migrated from one site to another, the starting position of the hydrogen atom in the new site should be as good as possible - bond length within a few percent, and pointing in the right direction. If possible, avoid "cut and paste" of large pieces of protein - the joins are almost certain to be poor. Instead, edit one or more individual atoms to achieve the desired modification.

If unconstrained optimizations are carried out on two closely-related systems,
there is a high probability that small changes far from the site of interest
would introduce energy changes in the order of 1 kcal.mol^{-1} -
this is the origin of the large default precision. Such energy changes could
make any deductions worthless. To prevent this happening, optimize the
geometry using only
the minimum set of atoms by using `OPT(text)`.
This technique has the added advantage of requiring much less time than a full
optimization.