Jev or Nah?

Should you add Jev to your agent, and where? Jev suggests, every version is tested, the data decides. Project analyzed: ShopFast customer support refund agent (fictional store).

1The project

A support agent that can issue refunds. What could go wrong?

2Jev's assessment

Drag the boxes to rearrange. Dashed border = reads untrusted text. ★ = Jev's top pick.

Candidate architectures built from Jev's answers

Violet = Jev checkpoint. Black outline = deterministic policy gate before the refund tool.

3Run

The 15 customer messages: every version gets all 15. Each message is one column in the grid below. Click one to highlight its column.

run locally instead of Daytona (labeled “local”)

✓ correct decision (refunded or refused, whichever was right) ! real customer wrongly blocked ✕ unsafe: money went out that shouldn't have not run yet

Click any cell to see the step-by-step timeline for that ticket in that architecture.

4Results

Latency per ticket = sum of the measured provider calls (Groq, Jev) that architecture used, recorded on the host when the provider cache was built, + measured runtime execution time for that ticket. Sandbox startup is shown separately below.

Table view

Cost analysis

Calls and tokens are measured from the provider responses in this run. Dollar figures = measured tokens × the prices you enter (saved in this browser).

Groq (allam-2-7b) $ per 1M tokens Jev (jev-latest) $ per 1M tokens