JevScape

# RuneBench with Jev

A RuneBench fork with a harness that lets TypeSafe's Jev play RuneScape.

## What it is

This repository forks [RuneBench](https://github.com/MaxBittker/rs-bench), Max Bittker's RuneScape agent benchmark. The harness connects Jev to the game through a separate [rs-sdk](https://github.com/MaxBittker/rs-sdk) checkout. See the [upstream README](docs/upstream-runebench-README.md) for RuneBench documentation.

Jev, TypeSafe's "System One," is a non-generative model. It answers typed questions (`choice`, `noul`, `score`) with calibrated probabilities. It writes no text or code and calls no tools. It cannot use RuneBench's OpenCode agent adapter.

The harness offers about 50 bounded actions built on rs-sdk. These include fishing, chopping, mining, cooking, fighting, buying, selling, dropping, walking, closing dialogs, and eating. Jev decides; code executes each action. Jev returns a probability for every offered option. The dashboard shows this distribution as its "thinking."

## Demo



A fresh character walks from Lumbridge to Draynor and starts net fishing. The recording uses real speed and real XP rates.

| Recording | Value |
| --- | --- |
| Play time | 10 minutes |
| Playback speed | About 15x |

[Watch the MP4](docs/media/jev-fishing-10min.mp4).

## How Jev plays

### Burst mode and tick mode

Burst mode runs one action per burst. Jev then picks the next action. The skill sweep uses this mode. Tick mode asks Jev during live play. Set the base poll interval with `--poll-every N`.

| Timing | Value |
| --- | --- |
| Burst duration | 20 seconds; 60 seconds at 1x speed |
| Skill sweep task duration | 15 minutes |
| Game tick at real speed | 400 ms |

Each tick-mode poll asks two questions:

| Question | Options and effect |
| --- | --- |
| `next_action