Project configuration¶

Most scans need only flags. Use --config when a project needs reusable settings, custom labels, or a parser written by you or an agent:

jevotron preview data.csv --config jev_config.py
jevotron scan data.csv --config jev_config.py --warnings-only

A small local config¶

Save this as jev_config.py, with rules.md next to it:

from pathlib import Path
from jevotron import Config
from jevotron.parsers import CSV

HERE = Path(__file__).parent
config = Config(
    parser=CSV(id_column="ident", fields=["/name", "/iso_country"]),
    guidance=(HERE / "rules.md").read_text(encoding="utf-8"),
    threshold=0.7,
)

Config files are explicitly loaded, ordinary Python. They export one Config object named config. No plugin registration is required. Install any custom parser dependencies in the same environment as the CLI.

Override settings for a run¶

jevotron scan data.csv --config jev_config.py \
  --field /iso_country --threshold 0.9 --guidance-file stricter-rules.md

CLI guidance, exemplars, model, threshold, field selection, and ID selection override the corresponding config or parser behavior. Relative CLI paths use your working directory; config-relative paths should use __file__ as above.

For built-in file and database parsers, --id-column replaces the configured ID column before parsing and validation, without modifying the reusable config. An arbitrary custom parser must first yield a chunk successfully; the CLI then replaces its reporting ID. The flag cannot bypass errors inside custom parser code.

Use OpenAI Decisions¶

The OpenAI Decisions API offers the same question types as Jev. Pick a gpt-* model and jevotron sends each request there instead:

export OPENAI_API_KEY="your-api-key"
jevotron scan spiked.csv --model gpt-6-luna --guidance "Check airport locations."

config = Config(model="gpt-6-luna", guidance="Check airport locations.")

Each field is still one choice question. The entry, guidance, and exemplars go in the single-turn text input. Question keys become question names, so answers route back to the same fields. Cache, batching, and output formats are unchanged. The model name is part of every cache key, so OpenAI and Jev assessments never mix.

jevotron preview shows the cached (Jev-shaped) request, and for OpenAI models it also shows the translated wire_request that is actually sent.

Caveats:

Pricing differs: OpenAI charges $0.10 per million input tokens and Jev charges $0.042, with no output charge from either. Token estimates and --batch-tokens budgets are based on Jev's tokenizer and context limits.
If the API declines a question (a refusal answer), the scan stops and the refused request's answers are not cached. Entries assessed in earlier requests stay cached; others in the same batch do not. If an oversized request was split, parts that were answered before the refusal may already be cached. The error names the refused entry and field. A rerun usually hits the same refusal, so exclude or edit that entry to continue.
jevotron sends text only. Image input is not used.
Scores from different models are not directly comparable. Re-check thresholds on a labelled sample before switching.

Customize labels¶

from jevotron import Config

config = Config(
    labels=["PASS", "REVIEW"],
    criteria={"PASS": "Consistent with the guidance", "REVIEW": "Likely incorrect"},
    anomaly_label="REVIEW",
)

anomaly_label selects the probability used for scoring and warnings. Then run the same preview and scan commands with this config.

For custom formats, see the advanced parser contract.