Community · E2 · artifact verified
Annotate tabular data with confidence-routed review
jevdata batch-annotates jsonl/csv rows with Jev noul, choice and score questions - concurrent, resumable, metered - and routes low-confidence rows to a review file; entity-matching benchmarks show a Jev-first Opus cascade hits F1 93.7 at 19% of all-Opus cost.
01 · Role in the system
What Jev does here
The pipeline takes state fields from each jsonl/csv row, asks all of a row's questions in one batched request with concurrency, backoff and a SQLite cache keyed by model, config, state and question, then flattens answers into columns; rows whose confidence falls below the threshold go to a separate review file instead of getting a forced label. The benchmarks are the second half: on Abt-Buy entity matching (5,397 candidates), Jev-only reached F1 90.2 for $0.12 in 1.8 minutes versus DeepSeek's 91.0 for $0.64 in 10.9 - a gap the author reports as not significant - while escalating Jev's unsure 12% to Opus reached F1 93.7 for $2.92 against a stratified $15.4 all-Opus estimate. Packaging three predicates per request cut a 3,949-sentence run from 117 to 78 seconds, and third-party routes doubled latency (0.59s vs 0.31s), which is why real tasks default to the official API.
02 · Control boundary
Where Jev sits
Code owns IO and policy; Jev only answers typed questions. One batched multi-question request per row, a SQLite-resumable cache, concurrency with backoff, and a confidence threshold that splits unsure rows into a review file for a human or a stronger model - a judgment-first cascade that escalates only on the uncertain band.
Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.
03 · Known limits
What this evidence does not prove
- Independent project, not affiliated with TypeSafe; 0 stars, first pushed 2026-10-14.
- Benchmarks are author-run on public datasets; third-party routes (OpenCode/OpenRouter) serve only as baselines and are rate-limited.
- The database/ DuckDB function experiments duplicate the mechanism of the existing duckdb-jev directory case; this record covers the annotation pipeline and benchmarks.
- clef_server.py serves a local /v1/systemone-compatible endpoint (Cloudflare Clef-flash) for experiments only; the pipeline defaults to the official API for real tasks.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.
- ElyasEbrahim ↗Community · github · public · checked 2026-10-14