Jev Experiments — mittal-parth

A repo of two TypeSafe System One loops. Dino-Jev drives Chrome's offline dinosaur at chrome://dino/ with a decision-log inspector, and Krunker-Jev is a Doom-shaped experiment in a local Krunker-style arena, not a live krunker.io aimbot.

Code snapshots structured state — for Dino-Jev, Runner.getInstance() exposes pose, speed and obstacles — and sends one TypeSafe request per tick with an action Choice, speculative jump/duck Nouls and an urgency Score. The canvas stays in Chrome; Jev never sees pixels.

No key runs a clearly labeled in-page heuristic collision bot. Setting JEV_API_KEY or TYPESAFE_API_KEY makes the loop default to jev; the README's commands include a heuristic run and a Jev run side by side so you can compare.

Create a venv, activate it, pip install -e '.[dev]', copy .env.example to .env, then run python -m dino_jev play or python -m dino_jev serve. The command launches Chromium itself on chrome://dino/ in fullscreen arcade mode; the inspector is at http://127.0.0.1:8766.

--speed-cap 9 (default) keeps maxSpeed just above the pterodactyl spawn of 8.5 so Jev's ~80 ms tick can still commit; --speed-cap none is full acceleration. --lead-frames 8 (default) sets how many frames early the runner jumps; lower values jump later. --windowed runs a normal window.

In dino_jev/policy.py, jump/duck gates prevent illegal moves: high pterodactyls cannot be jumped into and cacti cannot be ducked. The action Choice is only consumed when confidence is at least 0.28; below that the last action is kept, and urgency is displayed without moving the body.

Krunker-Jev is the Doom-shaped experiment applying the same decomposition — structured combat state in, typed actions out, Python owning physics, vision, hitscan and the inspector. The README excerpt truncates before its details, so this entry only covers what the visible text states.