Choosing a molecular model from average benchmark scores can be misleading: two methods may return different structures for the same complex, and their relative strengths can change with the input. The router is designed to make that input-specific choice easier to manage.
The launch article illustrates this with antibody-antigen complex PDB 9F91. Although several co-folding models rank strongly for the modality overall, OpenFold3 gives the strongest prediction for this particular example. The case shows why a single aggregate ranking is not enough for every molecule.
The routing system brings together Tamarind's independent model evaluations, public benchmark evidence, properties of the submitted input, and agent-assisted search. It can then choose a likely-suitable method and connect that choice to the platform's compute and result handling.
The initial router focuses on biomolecular structure prediction. The broader direction is to make model selection part of a shared scientific workflow as the number of available tools grows.