ViZDoom Level 3 · recorded evidence

76 decisions. 24 seconds. One corridor survived.

Watch Jev fight through six monsters and reach the armor, then inspect the action, response time, probability and tactical state behind the recorded decisions.

Seed 42: goal reached · 6 kills · 52 health · no fallbackRecorded 2026-09-20 · silent · 24 s gameplay · 27 s video

41/50held-out goals reached

260 msmedian episode mean latency

346 msmedian episode p95 latency

$0.00179mean cost per final run

0fallbacks in 50 final runs

What “fast” means: Jev makes a high-level tactical choice in roughly a quarter second. It does not control every rendered frame. Turns last about two Doom tics, movement eight, and shots twelve; ViZDoom executes those motor details while Jev chooses the next tactic.

What Jev is deciding

The game wrapper builds a context-dependent action mask. Jev sees structured tactical state and selects one allowed action with a probability distribution.

turn_left / turn_right

Align the view with the locked monster, or turn toward the armor’s goal bearing after combat.

shoot

Fire only when the locked monster is within 15 pixels of the crosshair. The public trace captured 27 shoot choices.

forward

Advance when no monster is locked and the goal is within 8 degrees of the current heading. The trace captured 22 advances.

strafe / backward

Reposition around obstacles or retreat from a close threat. These remained available in relevant states but were not selected in the captured snapshots.

Recorded decision trace

The application made 76 decisions. The 250 ms page recorder retained 73 distinct action snapshots; decisions 27, 31 and 45 completed between polls. Their timings remain in the run totals, but their actions are deliberately not guessed.

#	At	Decision	Latency	Selected p	Game state	Why this action was available
Loading recorded trace…

How to read the latency claim

Measured boundary

The timer starts immediately before the Vercel AI Gateway evaluation and stops when Jev returns its choice. Retry waits are included. Game rendering and action execution are outside this number.

Held-out aggregation

For every episode, the runner records mean and p95 decision latency. The published 260 ms and 346 ms figures are the medians of those episode-level values across 50 untouched final seeds.

Safety boundary

Jev chooses among context-appropriate actions. The wrapper rejects an unavailable choice and records a deterministic fallback. No fallback occurred in the final 50 Level 3 runs.

Real limitation

The loop waits for each remote answer. This demonstrates rapid sequential judgment over structured state; it is not raw-pixel perception or an asynchronous 60 Hz shooter agent.