Food. Water. Rest.
Three needs, remembered resource sites, and shelters. Recorded on Linux · 3,600 accelerated ticks · seed 27.
Synthetic agent with learned goal choices and a scripted route planner. Terrain is known; resource sites are discovered within three cells. This is separate from the fly-connectome experiments.
Loading recording…
B · bugF · foodW · waterS · shelterHeavy border · rememberedFaded · regrowing
Blue · energyGreen · waterOrange · fatigueScale 0–100
Alive
Current goal: explore
- Food eaten
- 0
- Water taken
- 0
- Shelter sleep ticks
- 0
- Learned contexts
- 0
- Remembered sites
- 0
Same-world comparison · 20 seeds
| Policy | Survived | Mean food | Mean water | Mean shelter sleep |
|---|---|---|---|---|
| Learner | 20 / 20 | 109.2 | 86.8 | 858.4 ticks |
| Learning off | 20 / 20 | 102.2 | 76.6 | 439.3 ticks |
| Site memory off | 20 / 20 | 64.3 | 48.4 | 815.1 ticks |
| Scripted priorities | 20 / 20 | 60.6 | 52.7 | 262.4 ticks |
All four policies survived all 20 held-out worlds. This setup does not demonstrate a survival benefit from learning. Resource counts describe behavior; collecting more is not automatically better.
Seeds 101–120, 3,600 ticks each, initially zero learned values. Site-memory off retains learned values and exploration visit ages. Physiology, rewards and high-level goals are engineered. No day/night, predators or biological neurons. These are accelerated simulations, not another wall-clock hour.
Evidence & reproduction
Run with Node 22.23.4: node run.mjs NEW_OUTPUT_DIRECTORY beside the model. A separate save/resume check matched the entire uninterrupted state. Recorded world is stopped; playback does not control Linux.