Question 1

Is the 11-0 win a real result or a rigged demo?

Accepted Answer

It is a real recorded run and the README is careful about what it claims: one demonstration, not a benchmark or win-rate claim. The table gives the specifics, 124 decisions, 0 late actions, 269.031 seconds elapsed, median and maximum API latency of 325 ms and 889 ms, model jev-1.13.0, and a calculated cost of $0.005218836, with the full verification notes in docs/VERIFIED_RUN.md.

Question 2

Why does it need Chrome started with remote debugging?

Accepted Answer

The controller drives the game through ordinary mouse and keyboard input over the Chrome DevTools Protocol, using a dependency-free CDP client rather than an automation library. You launch Chrome yourself with --remote-debugging-port=9222 and a dedicated --user-data-dir, leave that window open, and npm run play attaches to it, navigates the current page to RALLY and shows decisions in a fixed overlay.

Question 3

What does a full match actually cost to run?

Accepted Answer

The recorded Club run used 124,258 input tokens and 8,631 output tokens for a calculated cost of $0.005218836. The checked-in rate is $0.042 per million input tokens with output tokens free, but the README says pricing can change and to verify against the current model documentation before putting cost estimates into production. A Jev tick logs its token usage so you can compute it per run.

Question 4

Does Jev see the screen, or only numbers?

Accepted Answer

Only numbers. There are no screenshots anywhere in the loop: game telemetry becomes structured JSON, Jev answers typed Choice questions, the response is validated, and only then does code send normal mouse and keyboard input. That is the reusable boundary the project is demonstrating, so the same pattern can be pointed at another game or interactive task.

Question 5

What are the requirements to run it locally?

Accepted Answer

Node 22 or newer, Google Chrome and a TypeSafe API key from console.typesafe.ai. There are no npm dependencies to install. Copy .env.example to .env, put the key in it and chmod 600 on macOS or Linux, or export TYPESAFE_API_KEY instead. TYPESAFE_MODEL is optional and defaults to jev-latest. You then start Chrome yourself with --remote-debugging-port=9222 and a dedicated --user-data-dir and leave that window open while npm run play attaches to it.

Question 6

What does the run artifact actually record?

Accepted Answer

Every run writes an ignored artifacts/run-<timestamp>.jsonl with the exact model state and typed questions, the returned choices, probabilities, confidence and model version, latency, retry count and token usage, whether each time-sensitive action actually executed, and the point outcomes, final score, elapsed time and calculated cost. You can then run npm run summarize against that file. Recording a share-ready MP4 additionally needs playwright-cli and FFmpeg.

Question 7

What are the timing and network constraints?

Accepted Answer

A return crosses the table in roughly 0.85 to 0.94 seconds, so real-time success depends partly on end-to-end latency. The verified run reported median and maximum API latency of 325 ms and 889 ms across 124 decisions with zero late actions, which is why the README stresses that this is one demonstration and not a benchmark or win-rate claim. For slow or variable networks there is a --pause mode that freezes physics only during inference, plus --level and --max-seconds options.

Jev Play Ping Pong

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community