jevotron

Find the fields worth a second look. A CLI for field-level anomaly detection in tabular, structured, and text files, SQLite, and DuckDB, powered by Jev.

Preview the input, scan every entry, and export a focused review queue. Each chunk is independent and sees the same optional guidance and exemplars. SQLite caching reuses assessments across file versions.

For large guidance documents, opt into token-aware packing with --batch-size auto. Batching controls and limits explain when it saves tokens and how to compare scores with per-entry requests.

Example: classify agent traces. In a 24-trace public-data pilot, jt matched 130 of 163 published step-quality labels (79.8%). Harmful-step precision was 89.7%, with 70.3% recall. This was a small, length-filtered sample. Try the example · Read the results and reproduce the analysis.

Install and run

Requires Python 3.12+.

uvx jevotron --help          # run it without installing anything
uv tool install jevotron     # or keep it on your path
pip install jevotron         # or install it the usual way

Try it on the example file, no API key and no checkout needed:

curl -sO https://cmungall.github.io/jevotron/downloads/airports/spiked.csv

# Preview is free: it shows the exact request without sending it.
jevotron preview spiked.csv --id-column ident --field /iso_country --limit 1

# Get a key from https://console.typesafe.ai/ and set it in this shell.
export TYPESAFE_API_KEY="your-api-key"

jevotron scan spiked.csv --guidance "Check airport locations."

jt is a short alias for jevotron; both accept the same commands and options.

Show all entries with anomalous fields, highest confidence first, as YAML:

jt scan spiked.csv -a -s -O yaml
jt scan spiked.csv -f /iso_country -t 0.8 --sort-score -O yaml

Filter and order results with SQL expressions, including custom classifications:

jt scan spiked.csv --where "label = 'NORMAL'" -O yaml
jt scan spiked.csv --rows fields --where "label = 'ANOMALY'" \
  --order-by "confidence DESC" -O yaml

The default view retains whole entries; --rows fields filters and orders individual field assessments. SQL expressions are evaluated locally with DuckDB.

Short options include -l (limit), -f (field), -O (output format), and -t (minimum anomaly score). -a selects anomalies and -s sorts by confidence. See review & export for filtering and score semantics.

Database files are detected by their headers, even with an unfamiliar extension. No export or database configuration is needed; DuckDB support is included in the installation:

jt tables inventory.db                          # List tables, columns, and keys
jt preview inventory.db                         # Preview rows, no API key
jt scan warehouse.duckdb --table products        # Select a table
jt scan inventory.db                            # Scan all user tables

Databases are opened read-only. Each row is an entry, each column is a field, and primary keys supply table-qualified IDs. See the database guide for examples and selection rules.

API key: Live scans require TYPESAFE_API_KEY from the TypeSafe dashboard. An existing environment variable is used automatically; preview and fully cached runs need no key. Full setup instructions.

OpenAI Decisions: --model gpt-6-luna (or Config(model="gpt-6-luna")) sends the same choice questions to the OpenAI Decisions API instead, using OPENAI_API_KEY. Caching, batching, and output are unchanged. Details and caveats.

For development, use uv sync and uv run jevotron .... The CLI uses Typer, with grouped help and shell completion:

jevotron scan --help
jevotron formats
jevotron formats csv
jevotron --install-completion

Agent skills

Give your agent a large data file and a review question. The jt-triage skill teaches it to preview a few records, run a focused pilot, inspect a small queue of candidates, refine the guidance, and expand the scan. Full results stay on disk; the agent reads the cases that need reasoning. Cached assessments make repeat runs cheaper when the requests are unchanged.

npx skills add cmungall/jevotron --skill jt-triage

The skill installer supports multiple agents; it installs instructions and a local shortlist helper. Install the CLI separately with uv tool install jevotron and set TYPESAFE_API_KEY for live scans.

Try: “Use jt to find inconsistent country assignments in this large airport file. Pilot the rules, investigate the strongest cases, then scan the full file and give me a review queue with source locations.”

Skill catalog · Installation and workflow guide · Skill instructions

Documentation

Read the published documentation.

A Material for MkDocs site with light/dark modes, search, downloadable examples, and a CLI reference generated from the actual commands:

uv run --group docs mkdocs serve

Open http://127.0.0.1:8000. Build static files with uv run --group docs mkdocs build --strict. Changes on main are published to GitHub Pages by the documentation workflow.

• First scan
• Files and field selection
• SQLite and DuckDB
• File format reference — CSV/TSV, JSON/JSONL, YAML, TOML, text/Markdown, OBO, FASTA, GMT, and gzip; defaults and parser options
• Guidance and examples
• Review and export
• Caching across versions
• Shell pipelines
• Project configuration

Examples

Example	Includes
Airports	Public-domain CSV, baseline and spiked versions, provenance, known changes
Inventory	Synthetic YAML with written rules and a selected exemplar
Units	Synthetic OBO with repeated tags and a custom local config
Agent traces	Classify original public JSONL traces and individual steps; compare with published labels

Measured runs, sampling decisions, and reproduction scripts live in analysis/. The agent-trace pilot includes all predictions, baselines, and inspection of successes and misses.

Flags handle common tasks. For custom formats or reusable settings, pass --config jev_config.py; a parser is an ordinary Python function yielding chunks. The advanced docs cover that contract.

Development

just wraps the checks. just check is exactly what CI runs, so a green local run means a green pull request.

just install      # sync the dev and docs dependency groups
just check        # lint, tests, strict docs build, package build
just              # list every recipe

Individual steps are available too: just test, just doctest, just lint, just fix, just docs, just serve, just build. just mypy runs the type checker, which is not yet clean and so is not part of just check.

Unit tests use fake responses and remove the API key; they make no paid calls.