The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and evaluate external content (legal outputs), which is a common vector for indirect prompt injection. However, the skill operates within a restricted environment (allowed-tools: [Read]) and performs no external network operations or sensitive file writes. The multi-layered verification approach (objective and subjective) serves as a defensive mechanism against such injections by cross-referencing claims against sources.
Ingestion points: Input text provided for evaluation (legal output to be scored).
Boundary markers: None explicitly mentioned in the instructions for the input data stream.
Capability inventory: Restricted to the Read tool and integration with local companion verification skills.
Sanitization: No specific filtering of external content is described, but the analysis is purely diagnostic and returned as a report.
[SAFE]: The skill implements robust safety logic, including 'anchors' (objective failures like incorrect citations that block positive scores) and 'monotonicity' (requirement for new evidence to raise scores in revisions), which mitigate the risk of over-reliance on hallucinated or low-quality AI output.
[SAFE]: External references to established research projects (Fudan University's DISC Lab) and open-source patterns are used appropriately for attribution and methodology. These references do not involve remote code execution or untrusted dependency installation.
[SAFE]: The execution environment is tightly scoped with allowed-tools: [Read], data-residency: local, and pii-egress: none, providing strong protections against data exfiltration.

ocena-outputu-pl