Known limitations

Stated plainly, because they bound what the benchmark numbers support.

• Novel content is junctions, not genes. HelixForge calls coding loci that NAM lacks, and most of those are not RNA-supported. Treat them as candidates, not discoveries.
• ORF completeness trails the reference. Roughly 86% of HelixForge ORFs are complete versus 93% for NAM, driven by 5′-partial models. This is not parity.
• Improvement is net, not universal. Reconciliation adds far more exact transcript matches than it degrades, but a small number of models do get worse.
• Unsupported novel junctions are ambiguous. They are either unexpressed in the sampled tissues or genuine over-calls. The available evidence cannot separate the two.
• Isoform precision is the open axis. The extra isoforms HelixForge emits are largely plausible, but tightening isoform admission would raise chain-level precision.
• Requires protein evidence. Without a protein database (--protein-db) the Mikado chain does not run and the output is backstop-only, with no isoform discovery.