The sophistication of gene prediction algorithms and the abundance of RNA-based evidence for the maize genome may suggest that manual curation of gene models is no longer necessary. However, quality metrics generated by the MAKER-P gene annotation pipeline identified 17,225 of 130,330 (13%) protein-coding transcripts in the B73 Reference Genome V4 gene set with models of low concordance to available biological evidence. Working with eight graduate students, we used the Apollo annotation editor to curate 86 transcript models flagged by quality metrics and a complimentary method using the Gramene gene tree visualizer. All of the triaged models had significant errors-including missing or extra exons, non-canonical splice sites, and incorrect UTRs. A correct transcript model existed for about 60% of genes (or transcripts) flagged by quality metrics; we attribute this to the convention of elevating the transcript with the longest coding sequence (CDS) to the canonical, or first, position. The remaining 40% of flagged genes resulted in novel annotations and represent a manual curation space of about 10% of the maize genome (~4,000 protein-coding genes). MAKER-P metrics have a specificity of 100%, and a sensitivity of 85%; the gene tree visualizer has a specificity of 100%. Together with the Apollo graphical editor, our double triage provides an infrastructure to support the community curation of eukaryotic genomes by scientists, students, and potentially even citizen scientists.