DETECTSecurity & safety
Jev Evaluation
adhish@xadhish
This project evaluates Jev, a system from @typesafeai (System One AI via RLCD), across various use cases including AML structuring, ICU QTc drug interactions, SOC IMDS credential theft, and Tinygrad GPU thread predication.
An entire ecosystem of defensive code - regex cleaners, markdown strippers, and JSON retry loops - exists because we use text generation models where software simply needs a typed semantic judgment.
I evaluated @typesafeai's Jev (System One AI via RLCD) across:
• AML structuring (0% FP on $9.9k payroll)
• ICU QTc drug interactions (100% recall)
• SOC IMDS credential theft
• Tinygrad GPU thread predication
Link to the full post on the 4 failure modes and the $0.000032/decision economics in the thread.