What is jev-curate? What is Jev?

Core terms

jev-curate

An open source Rust binary that streams dataset rows from Parquet or JSONL, scores each row with Jev, and writes accepted and rejected rows to two output files. Repo: AkashPriyadarshii/jev-curate.

TypeSafe Jev

The System One decision model behind the tool, served at https://api.typesafe.ai/v1/systemone. "System One" means the model answers typed questions fast and cheaply rather than generating prose. Pricing: 0.042 USD per million input tokens, output unmetered, roughly 250k tokens per second.

Noul

Jev's calibrated probability answer. A Noul of 0.05 for "has_circular_reasoning" reads as a 5% chance the row contains circular logic. Presets set thresholds on these values to decide keep or reject.

Score

Jev's ordinal rating, 1 to 5, with a separate confidence value. The reasoning-math preset uses a reasoning depth score to keep steps that entail the answer.

Choice

Jev's pick from a fixed option list. Useful for routing decisions on a row, such as dataset family, error class, or whether to pass a row to a heavier eval.

Speculative fan-out

The client sends one row as state and all preset questions in a single request. One round trip per row instead of one per question. A 12.2x reduction in request cost at the batching layer.

Token bucket / rate limit

The client's adaptive limiter matching TypeSafe's 1,200 requests per minute. It refills 20 tokens per second, which is why a single node sustains about 20 rows per second. See benchmarks for the measured number.

Preset

A bundled evaluation rubric. The three shipped presets are reasoning-math, anti-sycophancy and code-correctness. Each asks Jev the same questions for every row.

Dry run

Runs the pipeline without contacting the API. Reports the keep and reject counts you would get, so a bad threshold costs zero tokens. Run --dry-run before anything else.

Parquet / JSONL

The two supported I/O formats. Parquet for columnar datasets, JSONL for line-oriented ones. Rows stream in and out; the binary does not load the dataset into memory.

clean.jsonl / rejected.jsonl

The two outputs in the output folder. Accepted rows land in clean.jsonl, rejected rows with their verdicts land in rejected.jsonl, so a human reviewer can audit every drop.