The Agent Skills Directory

[SAFE]: The skill implements standard, well-documented statistical methodologies for evaluator calibration (True Positive Rate, True Negative Rate, and Rogan-Gladen correction). These are standard practices in AI evaluation workflows.
[EXTERNAL_DOWNLOADS]: The skill references the judgy library and scikit-learn. These are standard, well-known libraries in the data science ecosystem. The reference link provided for judgy points to a reputable educational repository on GitHub (ai-evals-course/judgy).
[COMMAND_EXECUTION]: The skill generates Python scripts to perform statistical calculations (splits, confusion matrices, bootstrap confidence intervals) and to call judge APIs. This is consistent with the primary purpose of the skill and uses standard libraries like numpy and sklearn.
[DATA_EXFILTRATION]: No evidence of unauthorized data exfiltration. The skill specifically instructs the agent NOT to process data in-context (which prevents model-side leakage) and instead directs users to run local scripts or API calls against their own infrastructure.
[PROMPT_INJECTION]: No malicious prompt injection patterns were detected. The instructions are focused on guiding the user through a technical calibration process.

validate-evaluator