AI Eval Corporation

Independent AI test and evaluation. We measure how AI systems actually perform: accuracy, calibration, robustness, and fitness for the mission. We also publish open, offline models and benchmarks that anyone can check.

• Certified Service-Disabled Veteran-Owned Small Business (SDVOSB) · CAGE 10WJ5
• Web: aieval.org · Contact: scott@aieval.org

Leadership: Scott Mastin, President, previously led test and evaluation on Project Maven.

Released

• cadence-flash: an open-source ModernBERT-large decision model for military message type and precedence triage, with a human in the loop. Runs offline, calibrated, Apache-2.0. Type accuracy 0.985, precedence 0.957, Flash recall 0.995 (0.989 on short messages). Code: GitHub.
• cadence-flash-dataset: 13,500 synthetic military messages plus a 500-item public test slice, labeled by message type and precedence and built only from US Army doctrine approved for public release (Distribution A). CC-BY-4.0.

Results from our own test sets are labeled AI Eval curated benchmark.