This column shows the coding-potential score generated by the txCdsPredict program. This program weighs a variety of evidence, including the presence of a Kozak consensus sequence at the start codon, the length of the ORF, the presence of upstream ORFs, homology in other species, and nonsense-mediated decay. In general a score over 1000 almost certainly indicates that the transcript codes for a protein, and scores over 800 indicate approximately a 90% chance that a protein is produced.
The txCdsPredict program is part of the set of programs that are used to create the UCSC Genes track, summarized here. The source code, including the txCdsPredict program, is available by following the instructions here.