OpenRouter Labs: Prompt to questions

The fetched page does not say. It shows a comparison panel with Answers, Requests, Time, Cost and Accuracy, and Accuracy reads "Nothing compared yet" before a run, but nothing on the page documents how accuracy is computed, against what labels, or over how many rows. So treat it as a side-by-side display of two runs rather than evidence that either side is more accurate.

It starts Jev and the chat model at the same moment on the same job, which is the right shape for a race. What the page does not document is whether the chat model is handed the rewritten questions, how many rows each preset holds, or what the comparison is scoring against. The pipeline is paste a prompt, Claude rewrites it as Jev questions, then Jev runs them on every row of a Support, Reviews or Leads set.

Both sides are costed, since the panel carries a dollar line for Jev and another for the chat model, and running requires signing in to OpenRouter. What the page does not publish is which model performs the rewrite, whether that step is billed separately, or how a run's cost is split between the two. The counters are the only cost signal it gives, so check them before running anything large.

Not from what is published here. The page offers a View code link and sits in OpenRouter's Jev recipes collection at /labs/jev, and it points outward to the typesafe/jev-1.13 model page, but it documents no endpoint, schema or request shape for the compile step itself. If you need the pattern in production, treat it as one prompt in, typed questions out, run per row, and build it on the direct model page.

You paste one prompt; Claude rewrites it as Jev questions, then Jev runs them on every row of the chosen dataset, while a chat model races the same job from the same start so the comparison is head to head.

Preset row sets are Support, Reviews and Leads. A live panel tracks answers, requests, time, cost and accuracy for both Jev and the chat model, and before a run it reads Nothing compared yet.

Running it requires signing in to OpenRouter. The page publishes no benchmark results, dataset sizes or scoring criteria, so it is a comparison harness rather than evidence about accuracy.