Jev plays ViZDoom with two decision channels

Those are the author's configured rates for two separate channels, not measured round-trip latency. The post gives no per-call timing, no cost, and no baseline, so you cannot tell whether each decision kept up with its budget or whether the harness was effectively dropping frames.

Not the screen. Jev does not take image input, so the ViZDoom harness must serialize game state into text before the model sees anything. That means the harness supplies the perception and the tactical framing, and Jev only picks among the options it is handed.

The post reports the run ended with 18 kills, 102 health, and 92 armor, then got stuck walking in a circle. Nothing about recovery, retry behaviour, or a stuck-state detector is described, so there is no documented way to tell a deliberate patrol from a loop.

None in this post. It is one informal run with a video, with no win rate, episode count, or comparison agent. Most replies are jokes about auto-aiming, which is exactly the perception question the harness quietly answers on its own.

Kevin Madura, who links a project site and opens by calling DOOM the Jev hello world, the traditional demo target for a new capability. So it is a hobby build treated as a first-look demo rather than a benchmark, which matches the informal single-run result.

The perception shortcut. One reply calls it cheating, saying the harness gave the agent auto-aiming, and another admits he thought the post was a joke before realizing it was serious. That is the point the entry raises too: the harness supplies the vision and the aiming, and Jev only picks among the options it is handed.

Navigation at five decisions per second and combat at twelve. Splitting them lets the slower, position-level decisions run at a different cadence from the faster combat calls, keeping combat responsive without paying navigation cost on every frame. The post does not explain how the channels interact or share state.