jevotron
CLI that scans tabular, structured and text files or databases and asks Jev which fields look anomalous, building a review queue.
Overview
jevotron (alias jt) finds the fields worth a second look in CSV, structured and text files, SQLite and DuckDB. You preview the exact request for free, then scan every entry with optional guidance and exemplars; each chunk is assessed independently and Jev labels fields as normal or anomalous with a confidence. Results can be filtered and ordered with local DuckDB SQL expressions and exported as a focused review queue. A SQLite cache reuses assessments across file versions, and token-aware packing is available for large guidance documents.
It is aimed at data curators and researchers who need to spot suspicious values at scale. The README includes a pilot classifying agent traces: on 24 public traces jt matched 130 of 163 published step-quality labels (79.8%), with 89.7% precision and 70.3% recall for harmful steps, noted as a small, length-filtered sample. A spiked airports CSV is provided for trying it.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Is this field value normal or an anomaly given the guidance? | NORMAL · ANOMALY |
| Choice | What quality label applies to this agent-trace step? (custom classification)(inferred) | — |
How it works
- Preview the request for a file and field (free, no key)
- Scan entries with optional guidance and exemplars
- Jev assigns a label and confidence per field
- Filter and sort with SQL expressions and a minimum anomaly score
- Export the review queue; cached assessments are reused later
Strengths
- Free preview shows exactly what will be sent
- Works on many file and database formats
- Local SQL filtering and caching
- Reproducible pilot analysis included
Caveats
- Pilot results come from a small sample
- Every scanned entry is a paid API call
- Anomaly score depends on guidance quality
Getting started
uvx jevotron --help; then export TYPESAFE_API_KEY and run jevotron scan spiked.csv --guidance "Check airport locations."At a glance
- Type
- CLI
- Status
- beta
- Language
- Python
- Runtime
- CLI (Python 3.12+)
- Has code
- yes
- License
- BSD-3-Clause
Use cases and integrations
How curated lists describe it
CLI-first field-level anomaly detection for structured files and text, powered by Jev.
hellogumbo/awesome-jevCLI for field-level anomaly detection in tabular, structured and text files, SQLite and DuckDB, flagging the fields worth a second look with Jev. BSD-3-Clause.
MrJev/awesome-jev
Frequently asked questions
What does Jev decide?
Whether each field value is normal or anomalous given your guidance, with a confidence used to sort the review queue.
Can I see the request before paying?
Yes, jevotron preview shows the exact request without sending it.
Which inputs are supported?
Tabular, structured and text files plus SQLite and DuckDB databases.
Related projects
dbt package whose macros run Jev Choice, Noul and Score over SQL values in DuckDB and ClickHouse models.
ProjectbetaSQL (Jinja)DuckDB extension that asks Jev about every row in SQL and returns answers as ENUM, DOUBLE or STRUCT columns.
PluginbetaC++dbt package and Terraform module adding jev_noul, jev_choice, jev_score and jev_ask SQL functions to Snowflake.
LibrarybetaSQLPostgreSQL extension with jev(), jev_prob, jev_choice and jev_score to filter, rank and classify rows by plain-language conditions.
PluginbetaCDuckDB extension that labels rows in CSV, Parquet or DuckDB tables with Jev, reported at about ten seconds per thousand rows.
DiscussionprototypeSQLite and Node library adding Jev judgments to SQL for matching, evidence picking, ranking, review queues and semantic CI checks.
ProjectbetaJavaScript