Aeacus CLI

aeacus reads a text file or standard input, scores it with the published Aeacus model, and prints a verdict. The first run downloads the model, tokenizer, and calibration settings (about 82 MB) from Hugging Face and caches them under ~/.cache/aeacus/.

cat draft.txt | aeacus
aeacus draft.txt
aeacus draft.txt --json

Default output is a short summary with excerpts of likely AI regions. Regions combine adjacent tokens with the same classification. --json emits one JSON object, including grouped regions with character offsets, scores, and excerpts. --porcelain emits stable tab-separated summary fields.

The document ai_score is the mean calibrated score across tokens. It is an AI-likeness heuristic, not a probability of authorship. A score at or above 0.5 is ai_leaning; below 0.5 is human_leaning. Text under 30 words is insufficient_text, since the model was trained on longer passages. Per-token regions use the model's calibrated thresholds and may be likely_ai, likely_human, or uncertain.

The CLI runs on CPU. Use --model-dir to point to a local model directory containing tokenizer.json, thresholds.json, and onnx/model_q4f16.onnx. AEACUS_CACHE_DIR overrides the default model cache location.

This package will be published to npm as aeacus-cli, exposing the command aeacus. The CLI package has no license declaration. Publish its first version once from an npm-authenticated environment, then configure npm Trusted Publishing for this repository and .github/workflows/publish-cli.yml. Later releases build a package archive and publish when a matching cli-v<version> tag is pushed.