Measure Jev against LLMs on real workflows, weekly

01 · Role in the system

What Jev does here

The series asks which steps of a real workflow are decisions, which are writing, and what happens to latency and bill when a frontier model stops making the decisions. Week one, inbound lead triage: the same leads go to Jev and to Haiku 4.5 and Sonnet 5, judged against the annotated ground truth - route accuracy 90 percent for Jev against 78.3 for Sonnet 5 and 53.3 for unsplit Haiku, with zero of sixteen funded buyers mis-binned (Sonnet lost one, Haiku seven), at a p50 of 366 milliseconds against 2.6 seconds, and the author reports about four cents per thousand leads against dollars. Sonnet then writes the emails, because Jev structurally cannot. Week two applies the same measurement to deal-risk assessment. Every week ships its code, and the numbers come from the committed runs rather than estimates.

02 · Control boundary

Where Jev sits

Same workflow, same fixtures, multiple decision engines; typed questions to Jev against LLM baselines, measured on accuracy, latency, and cost from committed runs.

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

• Week one covers 60 annotated leads and week two deal risk; the series is early and the numbers are author-measured from committed runs.
• Shares its repository name with an unrelated Jev project; this record covers only deepdave98's measured series.
• Cost comparisons quote the author's token accounting through OpenRouter.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• deepdave98 ↗Community · github · public · checked 2026-09-25