Benchmarks
How AI models and agents perform on regulated aviation documents
Each benchmark puts models and agents on the same task, the same files and the same brief, and scores the output against the sources. The data set, the rules, the rubric and the limits are published with every edition, and every page carries its version and date.
Working paper v1.0 · · 9 min read
RDB-1: Regulated Document Benchmark
Do you have to trade data security for quality? RDB-1 scores models that run inside an aviation company's own perimeter against cloud models and a general agent tool on one task: a technical report a domain expert would sign, written from a consolidated data set of 96,180 spreadsheet cells, 8,561 PDF pages and 872 review comments from a completed regulatory assessment. Working paper v1.0, September 2026.
Read the benchmark
Want a model or agent scored?
Runs are performed on the unchanged data set and published with the same rubric. Write to [email protected].