Preprint reports more faithful AI extraction of hardware specifications

Researchers report that a staged language-model pipeline extracted hardware specifications with 98.8% property faithfulness, compared with 76.1% for a one-shot baseline, in a preprint evaluating 20 hardware datasheets. The pipeline extracts properties from document chunks, organizes them into hierarchical graphs, and iteratively refines the results. It preserves links to source material and executes validated graph updates through deterministic local operations, providing specifications for downstream hardware design and verification. The pipeline fully covered 83.6% of faithful baseline properties; the baseline fully covered only 10.5% of the pipeline's properties. Those coverage measures compare the graphs with each other rather than establish completeness against exhaustive ground truth.