GeneBench
A construct, its constraints, and an independent check of the outcome.
Public benchmark results have not been published. The earlier leaderboard contained interface test data, not measured model results, and has been removed. It should not be used to compare models or assess Helicase.
What an evaluation needs to show
Defined tasks, held-out constructs, and explicit constraints.
Task-appropriate baselines with the same tools and data access.
Final sequence artefacts and checks independent of the model’s own assertions.
Failures, unresolved findings, latency, and cost alongside successful outcomes.
Computational correctness does not establish experimental function. Helicase model results and GeneLoop workflow results will need to be distinguished.