Jevotron: Multiple Jev integrations from the command line

Jevotron is a command-line tool for one-shot anomaly detection that scores individual fields in structured files and builds a focused review queue. Provide a file and brief guidance, preview model requests without an API key, then run scans that call a remote TypeSafe API (API key set via environment variable) to produce field-level anomaly probabilities. It supports CSV, YAML, JSON, TOML, OBO, plain text and gzipped inputs, lets you select fields or provide longer guidance via a file, and offers a local Python project config for custom parsing or reuseable settings. A built-in SQLite cache records successful results so unchanged entries are not reassessed, making reruns or reordered files cheap; a warning score equals the highest field anomaly probability and thresholds drive what appears in the review queue.

Outputs retain each field’s probability, an overall entry score, source location and assessment date; results can be sorted, exported as CSV, or piped as JSONL into shell workflows. Examples include an airports CSV workflow that surfaces injected country errors and an agent-trace pilot on 24 traces where step-quality labeling matched 130 of 163 labels (79.8%) with harmful-step precision 89.7% and recall 70.3%. The tool emphasizes quick, auditable evidence for reviewers and integrates with existing shell and Python pipelines for repeatable anomaly detection.