WORK / CHARGERUN-JEV · 2026-09

Letting Jev play my own game — how the hand-off changes the score

A controlled experiment in which TypeSafe's Jev decision model played Charge Run, my one-key dodging game. Over five rounds I changed how the game state was handed over, and compared Jev with random play and my own hand-written policy on the same 20 courses. What raised the score was the hand-off and code-side safety; under these conditions, Jev's own choices did not show a measurable benefit.

Type: Self-initiated
Role: Experiment design, build, analysis
Assets used: My own game, Jev, OpenJev
Delivered as: Four-way comparison video (38 s, no audio)

§ 01 — Problem

Can a game that demands sub-second decisions be handed to an LLM-based decision model? And if it can, is the gain coming from the model's judgment or from the information and scaffolding I give it? Without separating the two, "it worked when I let the AI play" is not evidence of anything.

§ 02 — Decision

I fixed the interface: on every landing, hand over the current state and ask for one of three moves — wait, small jump, big jump. As baselines I used players that ignore the state (always big-jump, random) and my own hand-written policy, and ran everyone on the same 20 courses, comparing medians and per-course wins. Each round changed only one thing about the hand-off, and for all five rounds the pass criterion was written down before running (the final lookahead round was a score-maximizing attempt run outside that protocol). The steps were: raw position and velocity → a code-computed "does this move collide?" → code removes colliding moves first → keep only moves that survive the next 0.5 seconds.

Bar chart of median scores by round. Jev goes from 31 in R1 to 285 in R2, 434 in R5 and 1863.5 in the lookahead round. The same scaffolding without Jev scores 532 in R5 and 1698.5 in the lookahead round. Always big jump scores 153.5–167 per round; my hand-written policy 599.5–1125.5. — FIG.02 — Score by round (median of 20 courses per round). Each round used different courses, so cross-round comparison is approximate

FIG.02 — Score by round (median of 20 courses per round). Each round used different courses, so cross-round comparison is approximate

§ 03 — Learning

The hand-off moved the score. Given collisions as a count, the model preferred the moves that collided; given a true/false "collides", it avoided them, and the median went up five-fold (56 → 285). Removing colliding moves in code first was the first setup that beat the always-big-jump baseline, but random choice among the remaining moves did just as well as Jev (11 wins, 9 losses). Adding lookahead raised the score from 726 to 1863 — yet the rule-only policy without Jev reached 1698. Writing "prefer the big jump among safe moves" into the prompt made Jev's scores identical to the always-big-jump player on 14 of 20 courses. Total Jev usage cost was $0.53. Next time I would size the number of courses needed to detect the model's own contribution before running, and handle real-time latency (two frames of delay cut the score to about a quarter) by queueing moves ahead.

Read the article →

Letting Jev play my own game — how the hand-off changes the score