See what Zils decides.
Zils beside TypeSafe JEV in every demo. Same inputs, real recorded answers.
Zils reaches the armor. JEV dies in the doorway.
Behind this game
Test seed 100001 is a selected highlight: the lowest-numbered held-out start where trained Zils collected the green armor and JEV died. 96 of the 100 starts qualified; none went the other way. Both runs began from the same frame and the same observation, and the recorded hashes match. The footage is rendered from the native game recordings with labels added, not a recreated win.
Both models received the same text: health, the names and screen positions of visible enemies and the armor, the previous observation and the last action. Neither saw pixels, the map, or anything off-screen. Thinking paused the game, so playback is game time at 35 tics per second and omits inference waits.
Zils: h2o-lightning-4b-v1.2.3 with one adapter trained on 2,048 decisions from a scripted reference program. JEV: a live call to TypeSafe’s API, version jev-1.13.0, with no task training.
Full test details
Correct means it collected the armor and reached the exit. 100 held-out starts on one map, every replay audited. One selected game is shown above.
Training improved Zils from 0/100 to 97/100 on this test.
| Model | Correct | Score |
|---|---|---|
| ZilsTrained H2O Lightning 4B adapter | 97/100 | 97.0% |
| TypeSafe JEVLive API run · jev-1.13.0 | 0/100 | 0.0% |
| Zils before trainingUntrained H2O base | 0/100 | 0.0% |
| Scripted referenceThe program Zils learned from | 100/100 | 100.0% |
2026-10-13: 100 held-out starts of ViZDoom 1.3.0’s unmodified Deadly Corridor scenario, difficulty 5, with a 2,100-tic limit. Success means collecting the armor and reaching the map’s exit. Trained Zils: 97 successes, 2 timeouts, 1 death. JEV: 99 deaths and 1 technical failure, counted separately. The untrained H2O base died in all 100. The scripted reference program Zils learned from succeeded in all 100.
Training, development and test seed lists were disjoint, but the map is shared, so this measures robustness to the scenario’s randomness, not play on unseen levels. The structured observations and custom action interface make it different from standard pixel-based ViZDoom scores.
The paired success difference is +97 points, with a descriptive 95% seed-bootstrap interval of 93 to 100 points; Zils’s Wilson 95% interval is 91.5% to 99.0%. The native replay audit verified 399 completed episodes and matched starting frames for all 100 seeds. JEV’s one technical failure, seed 100085, was an operator-interrupted episode and is not counted as a death.
Full benchmark reportAll 100 starts (CSV)Results (JSON)
Each example uses its task-specific Zils adapter. These studies do not establish production reliability. The research adapters shown here have not been deployed.