The current manuscript is paper/main_v2.pdf, compiled
from paper/main_v2.tex.
pip install -e ".[dev]"Layer 1 runs locally. Layers 2 and 3 use remote LLM APIs through LiteLLM. Provider credentials must be supplied through local environment or LiteLLM configuration and must not be committed to this repository.
skillsieve scan ./path/to/skill/
skillsieve scan ./path/to/skill/ --max-layer 1
skillsieve scan-batch ./skills-dir/ --output results.jsonresults/labeled_dataset_final.json
contains the benchmark index and labels used by the paper. It includes the
source author/skill path, label provenance, and the flag indicating whether a
record belongs to the 390-skill evaluation subset.
Copies of third-party skill packages are not redistributed. This avoids republishing upstream code, credentials embedded in public samples, and files with incompatible or unclear licenses. The labels apply to the archived versions used in the study and are not claims about an author or a package's current version.
Raw LLM responses, intermediate checkpoints, internal audits, superseded result files, and draft manuscripts are intentionally excluded. Aggregate results are reported in the paper.
skillsieve/ Core detection pipeline
experiments/ Paper experiment source code
results/ Minimal public benchmark labels
paper/ LaTeX source, required figures, and current PDF
tests/ Unit tests
bot/ Feishu deployment described in the paper
pytest -q testsMIT