Skip to main content
Two evaluation systems measure whether the engine — and the agents built on it — actually work. Both grade against evidence the agent cannot see, so scores can’t be gamed.

​
Structural-biology environments

Each task is a resettable workspace: an agent gets input files and a prompt, and must produce named artifacts — a structure model, a contact table, a rendered image. Those artifacts are graded deterministically against a hidden reference, and the reward is a weighted composite of geometric measurements (lDDT, TM-score, DockQ, clash, contact F1, map fit) versus a per-task threshold. Held-out validation metrics penalize overfitting to the evidence the agent can see. Tool use is policy-gated rather than scored, and full trajectories are logged for replay and audit.

​
Homology-recall benchmarks

For the search engine itself, curated truth sets derived from SCOPe/CATH structure classification and Pfam/InterPro domain families label query–target pairs as positive, negative, or ignore. Runs are scored on biological recall (positive_recall_at_k) against false-positive ceilings, across strata like remote homologs, low-complexity, transmembrane, and shuffled decoys. Tool outputs (MMseqs2, DIAMOND, BLAST) are kept as comparison baselines only — never as ground-truth labels.