Several parameters can be used as convergence criteria. The crystallographic R factor is a reasonable quantity to monitor.
In warpNtrace mode, look at the 'Connectivity index' after each autobuilding. In general, if the program stopped with a connectivity index of above 0.8 there is room for further improvement, the refinement has not converged. Submit additional cycles. A value for connectivity index of above 0.95 is a very good sign for convergence, though it can often stay between 0.9-0.95 if you have disordered parts. Also keep an eyes on the number of residues actually built and the number of chains. Ideally the number of chains should be one (for a single polypeptide protein) and the number of residues should be as close as possible to the actual number in the protein - this will rarely be the case during initial stages autobuilding, but towards the end of the run something is going wrong if you only have loads of short chain fragments.
What to do if the R factor stays at the values around 30%:
If for example after 5 or 10 cycles, R dropped to 28-34% and stayed there for the next 10 cycles without any tendency to drop further, you may be in trouble. Increase your high resolution data would be an effective method -haha. Try to change from Fast to Slow protocol or opposite, try to introduce phase restraints, change the REFMAC parameters, panic, cry, etc.! We are working on more sensible suggestions all the time! Your feedback is needed and greatly appreciated!
Version 6.0 should converge several times faster for 2 Å and better data and can handle poorer input. Appraocing the 2.5 Å regime, the routines are unlikely to produce a full model, mistracing is possibe - especially for these cases, feedback would be appreciated.