The Agent Skills Directory

[SAFE]: The skill follows established scientific computing practices for uncertainty quantification. It leverages trusted third-party libraries such as ase (Atomic Simulation Environment) and mace-torch for handling materials data and performing machine learning inference.
[COMMAND_EXECUTION]: The skill's instructions and example documentation involve executing Python scripts and command-line utilities (e.g., mace_run_train) for data preparation, model training, and analysis. These actions are standard for the intended use case of developing and deploying machine learning models in a research environment.
[DATA_EXPOSURE_AND_EXFILTRATION]: The run_committee_inference.py script generates output in the form of JSON summaries, PNG/SVG plots, and structure files written to a local directory. There is no evidence of network-based exfiltration or unauthorized access to sensitive system files.
[INDIRECT_PROMPT_INJECTION]: The skill processes external structure files (e.g., CIF, XYZ, POSCAR) as input. While this represents a data ingestion surface, the files are parsed into scientific objects by the ase library and are not used to construct or influence agent instructions.
Ingestion points: Path to structure files provided via the --structures argument in run_committee_inference.py and read by ase.io.read.
Boundary markers: None explicitly implemented for scientific file formats within the processing script.
Capability inventory: The script writes flagged structures and summary data to the local file system using standard library functions.
Sanitization: Relies on the scientific validation and parsing logic of the specialized ase library.

ml-committee-uncertainty