Assess first, route later: A multi-dimensional difficulty assessment framework for biopharmaceutical patent extraction

The batch extraction of structure-activity relationship (SAR) data from global biopharmaceutical patents is a critical bottleneck in drug discovery. The extreme heterogeneity of these documents causes "one-size-fits-all" automated pipelines to either waste resources on simple patents or fail completely on complex ones. To solve this, we shift to an "assess first, then route" paradigm. This report introduces a multi-dimensional difficulty assessment framework that decomposes extraction complexity into 16 evaluation dimensions across four core categories: document physical quality, patent metadata, bioactivity data expression, and Markush structure complexity. Emphasizing engineering pragmatism at scale, our scoring engine uses a dual-track architecture-combining fast deterministic rules with Large Language Models (LLMs) for deep semantic evaluation. By dynamically grading patents from L1 to L4, the framework enables downstream pipelines to intelligently route tasks: maximizing Straight-Through Processing (STP) for L1/L2 patents while reserving Human-in-the-Loop (HITL) resources for L3/L4 cases. Validation on over 94,000 global patents across seven patent offices demonstrates that over 53% of these heterogeneous documents can be securely routed to STP pipelines. This effectively eliminates redundant manual review overhead for straightforward cases, while preemptively preventing pipeline failures on deeply complex structural formulations.