how does cuffcompare choose which transcript to put in combined.gtf file?

[d f]

When cuffcompare is run with multiple input gtf files from different sequencing expts, how does it decide which transcript to be the "representative" transcript to list in the combined.gtf file? If multiple transcripts have the same intron structure but are not identical, which one does it choose to put in combined.gtf? It does not always choose the longest one as the "representative" transcript (which is what I thought it would do).

For example, from these two input gtf files,

contig00177     Cufflinks       transcript      230     711     1000    .       .       gene_id "CUFF.68933"; transcript_id "CUFF.68933.1"; FPKM "107.2929288684"; frac "1.000000"; conf_lo "86.576469"; conf_hi "128.009389"; cov "2.269504";
contig00177     Cufflinks       exon    230     711     1000    .       .       gene_id "CUFF.68933"; transcript_id "CUFF.68933.1"; exon_number "1"; FPKM "107.2929288684"; frac "1.000000"; conf_lo "86.576469"; conf_hi "128.009389"; cov "2.269504";

and

contig00177     Cufflinks       transcript      230     1047    1000    .       .       gene_id "CUFF.71009"; transcript_id "CUFF.71009.1"; FPKM "86.1874620509"; frac "1.000000"; conf_lo "67.620022"; conf_hi "104.754903"; cov "2.055336";
contig00177     Cufflinks       exon    230     1047    1000    .       .       gene_id "CUFF.71009"; transcript_id "CUFF.71009.1"; exon_number "1"; FPKM "86.1874620509"; frac "1.000000"; conf_lo "67.620022"; conf_hi "104.754903"; cov "2.055336";

the transcript listed in the combined.gtf file is the first and shorter transcript:

contig00177     Cufflinks       exon    230     711     .       ^@      .       gene_id "XLOC_000001"; transcript_id "TCONS_00000001"; exon_number "1"; oId "CUFF.68933.1"; class_code ".";

I am running version 0.9.2 of cufflinks/cuffcompare, and am running it without a "reference" annotation GTF.

Incidentally, my goal here is to create a transcriptome for a mostly unannotated, novel genome. I have RNA-seq data from two different sequencing runs. I ran TopHat (with a couple of parameter sets) and Cufflinks on the reads to predict transcripts. I am now using cuffcompare to consolidate the results of the transcript predictions from the two sequencing/TopHat runs to create the most "comprehensive" transcriptome.