The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates the execution of a bundled utility, scripts/calibration.py, used to process paired datasets and output agreement statistics. The script is self-contained and uses standard libraries.
[INDIRECT_PROMPT_INJECTION]: The skill handles datasets containing model responses and human labels, presenting an attack surface for indirect prompt injection.
Ingestion points: Data is loaded from pairs.jsonl, verdicts.jsonl, and labels.jsonl files (referenced in SKILL.md and scripts/calibration.py).
Boundary markers: None are defined to isolate potentially adversarial text in the model response fields during statistical analysis.
Capability inventory: The skill's capabilities are limited to reading data for statistical computation and reporting; no high-risk capabilities like network access or system modification are triggered by the data content.
Sanitization: The skill relies on standard data structures without specific filtering for natural language injection patterns.
[SAFE]: The skill and its referenced script focus on transparent statistical analysis. No evidence of obfuscation, exfiltration, or unauthorized privilege acquisition was found.

align-human