# Jev Tutorial baseline protocol

Status: protocol published; no result has been published by Jev Tutorial.

This is a small synthetic smoke test, not a production benchmark and not evidence of general Jev accuracy. Its purpose is to make the first independent run inspectable before a larger human-labelled dataset is prepared.

## Task

For each support ticket, ask two independent questions in one request:

1. `route`: choose `billing`, `technical`, or `account`.
2. `urgent`: estimate whether the ticket requires attention now.

The expected labels were written with the synthetic fixtures. They are not customer data.

## Required disclosure

Publish all of the following together or do not call the result a benchmark:

- run date, runtime, region, SDK or HTTP client version;
- returned model identifier and request ids when available;
- unedited input cases and question definitions;
- every raw response, HTTP failure, retry, and exclusion;
- median and p95 latency, route exact-match accuracy, urgent Brier score;
- total input tokens and provider cost when the response exposes enough data;
- limitations, including the synthetic and very small dataset.

## Run

```bash
export TYPESAFE_API_KEY="your-server-side-key"
node run.mjs
```

The script writes timestamped JSONL into the current directory. It never prints the API key. Review the raw file before calculating or publishing conclusions.

## Promotion to E3

This smoke test alone does not qualify a project for Jev Tutorial's E3 evidence level. E3 requires a published environment, inputs, outputs, results, limitations, and a reproducible project-specific procedure.