The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions involve processing human-labeled data and production traces, which introduces a potential surface for indirect prompt injection if the evaluated content contains malicious instructions.
Ingestion points: Data is ingested through variables like labeled_data, human_labels, and prod_eval_labels in SKILL.md.
Boundary markers: There are no specified delimiters or 'ignore instructions' markers for the data being processed by the judge model.
Capability inventory: Performs statistical evaluation and bootstrap simulations using standard libraries like scikit-learn and numpy.
Sanitization: The provided methodology does not specify any sanitization or filtering for the evaluated traces.
[EXTERNAL_DOWNLOADS]: The skill recommends the installation and use of a specialized statistical package for bias correction.
Evidence: The instruction pip install judgy is provided in Step 8 of SKILL.md for success rate estimation.

validate-evaluator