# Jev Flow product tour — English

**Language:** English
**Narration:** two-speaker dialogue in natural American English, generated with Gemini 3.8 Flash Lite TTS.
**Voices:** Alex — Charon (masculine-presenting voice); Maya — Kore (feminine-presenting voice)

## Texto narrado

**Alex:** What if AI could add judgment without taking control away from your software?

**Maya:** That is the idea behind Jev Flow: clear inputs, typed judgments, and routes the application can inspect.

**Alex:** Deterministic AI does not promise a perfect model. It means the workflow rules are explicit, validated, and owned by code.

**Maya:** Jev answers a bounded question with a Choice, a probability, or a score. Code validates the shape before any branch runs.

**Alex:** And Laya? It offers an optional local inference path for compatible models when a team prefers to run its own checkpoint.

**Maya:** In the Flow Studio, connect triggers, Jev judgments, conditions, and logs on one visual canvas.

**Alex:** This support-triage walkthrough uses a labeled fixture. It demonstrates a route without presenting fixture data as a provider call.

**Maya:** The Compendium generates and certifies 388,080 configurations for validity and unique identifiers. That is catalog coverage, not model evaluations.

**Alex:** Now the Battle Arena sends the same synthetic incident live to TypeSafe Jev and the selected paid OpenRouter model.

**Maya:** This captured run returned five typed Jev judgments in 740 milliseconds; GPT-4.1 Mini answered in 1,638 milliseconds. One run is a demonstration, not a benchmark.

**Alex:** The games make rules tangible. Tic-tac-toe uses a local minimax opponent, so its move needs no model or API call.

**Maya:** In Chess, local rules reject illegal moves. In Blast Garden and City, inspect how each state changes as you play.

**Alex:** The Self-Driving Sim sends structured lane, speed, and obstacle state to a decision step. This demo uses no camera vision or OpenCV.

**Maya:** Inspect the trace, change a flow, and test it with your own synthetic cases. Jev Flow makes model-assisted workflows visible and auditable.

## Notas de produção (não narradas)

- The Studio support-triage view is a labeled simulated fixture. It is not a live provider answer.
- The Compendium figure of 388,080 represents generated configurations checked for validity and unique identifiers, not model evaluations.
- The Arena footage is one recorded live comparison: TypeSafe Jev returned five typed judgments; the selected paid OpenRouter model was OpenAI GPT-4.1 Mini. The captured latencies were 740 ms for Jev and 1,638 ms for GPT-4.1 Mini. This is a demonstration, not a benchmark.
- Tic-tac-toe uses local minimax by default, and chess legality is enforced by local code.
- The Self-Driving Sim generates structured obstacle and lane state; this demo does not use a camera or OpenCV.
- Speech: Google Gemini 3.8 Flash Lite TTS through OpenRouter. Captions are burned into the video; a separate WebVTT file is provided.