Jevotron: Multiple Jev integrations from the command line

One-shot anomaly detection on the command line¶ Give jevotron a file and a little guidance. Get field-level anomaly scores, a focused review queue, and a cache that makes the next run cheaper. Run your first scan Explore the commands Install agent skills $ jevotron scan airports.csv --guidance "Check airport locations." First, set your API key Get a key from the TypeSafe dashboard and set it in your shell: export TYPESAFE_API_KEY="your-api-key" Already set? You're ready to scan. preview needs no key. API key setup → · See the actual airport results → 01 / PREVIEW See what goes in¶ Inspect chunks, field paths, and exact model requests before making an API call. Preview a file → 02 / SCAN Assess every entry¶ Jev scores selected fields together. Each entry gets the same guidance and optional examples. Choose your fields → 03 / REVIEW Start with the warnings¶ Sort suspicious entries, export CSV, or pipe JSONL into your existing shell workflow. Build a review queue → Small setup. Useful defaults.¶ CSV, YAML, JSON, TOML, text, OBO, and more work out of the box, including gzip files. The format reference covers defaults and format-specific options. Select fields with --field, add a sentence with --guidance, and run. Longer instructions can come from --guidance-file. A local Python config is available when a project needs custom parsing or reusable settings. Unchanged input reuses its assessment. SQLite saves each successful result as it arrives. Reorder a file, change a reporting threshold, or resume a failed run without reassessing unchanged entries. A review aid with visible evidence. Reports retain each field's probabilities, the entry score, source location, and assessment date. The warning score is the highest field anomaly probability; you choose the threshold. Try a complete example¶ Agent traces: a measured pilot¶ On a small, length-filtered sample of 24 public traces, jt matched 130 of 163 step-quality labels (79.8%). Harmful-step precision was 89.7%, with 70.3% recall. The example uses original messages and tool definitions, with human labels withheld from the model. Try the trace example → · Read the full analysis and limitations →