Jevotron: Multiple Jev integrations from the command line
8 hours ago
- jevotron is a command-line tool for one-shot anomaly detection that provides field-level anomaly scores and a focused review queue.
- It includes a cache that makes subsequent runs cheaper by reusing assessments for unchanged input.
- Users start by scanning a file with guidance, like 'Check airport locations.', after setting an API key from the TypeSafe dashboard.
- The workflow has three steps: preview (inspect data and model requests), scan (assess entries with field scores), and review (sort warnings, export CSV or JSONL).
- jevotron supports multiple formats out of the box, including CSV, YAML, JSON, TOML, text, OBO, and gzip files, with options like --field and --guidance.
- Results are saved in SQLite, allowing resumption of failed runs or reordering without re-assessment.
- Reports include field probabilities, entry scores, source locations, and assessment dates, with a warning threshold based on the highest field anomaly probability.
- Example use cases include finding injected errors in airport CSV data, applying rules to YAML inventory, scoring OBO measurement units, and classifying agent traces.
- In a pilot test on agent traces, jevotron matched 79.8% of step-quality labels, with 89.7% precision and 70.3% recall for harmful steps.