Benchmark RadarEduPluginBench
EduPluginBench is an executable benchmark for evaluating code-generation models on producing plugins that meet governed ecosystem requirements including least privilege, telemetry consent, provenance, and bounded failure. It uses 1,440 mutants and 120 clean references with staged admission levels P0-P4.
- Released
- 2026-08-01
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Compilation and functional tests do not ensure compliance with security and governance constraints. EduPluginBench provides a staged admission method targeting these gaps, enabling assessment of generated plugins in governed environments, which is critical for safe deployment.
Motivation
Code-generation models can produce executable components, but compilation and functional tests do not establish compliance with least privilege, telemetry consent, provenance, privileged-write authority, lifecycle constraints, or bounded failure.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.