predictive-coding-information-bottleneck

Pass

Audited by Gen Agent Trust Hub on Apr 21, 2026

Risk Level: SAFE
Full Analysis
  • [SAFE]: The skill provides instructional content and Python code templates for building a machine learning pipeline using standard libraries such as NumPy, SciPy, Scikit-learn, and Transformers. All operations are local to the user's environment and relate directly to the stated purpose of hallucination detection.
  • [EXTERNAL_DOWNLOADS]: The implementation examples reference pre-trained models from well-known and trusted services, specifically the NLI DeBERTa model from Hugging Face's model registry and the English core model from spaCy. These are standard dependencies for natural language processing tasks.
  • [DATA_EXPOSURE_AND_EXFILTRATION]: No evidence of hardcoded credentials, access to sensitive file paths, or unauthorized network operations was found. The skill operates on user-provided data (questions, context, and answers) to generate quality scores.
  • [COMMAND_EXECUTION]: The skill does not contain instructions for executing arbitrary shell commands, modifying system configurations, or acquiring elevated privileges.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 21, 2026, 01:36 PM