Selected outputs from the current NULLVECTOR pipeline. These are generated results, not concept art. Some show accepted runtime systems; others are clearly labeled research prototypes.
The current nature-stage integration: live posed and damaged cell sets are batched through the continuous organism VAE while the promoted locomotion, physiology, feeding, intent, colony, society, timeline, and world-model specialists run inside the causal simulation.
The neural target field, muscle controller, and contact controller approach a material clump. An articulated arm then carries it into live feeder cells. Feet remain planted and the held object stays constrained to the hand.
Humanoids kneel and grasp. Animals bite from the ground. Plants siphon through roots. Anomalies use a phase field. Machines collect with a suspension-mounted tool.
Thrown matter has horizontal momentum, elevation, a ground-plane shadow, gravity, recoil, and landing behavior.
Different materials bounce, roll, or thud.
The complete arm detaches at the shoulder, drops its payload, twitches briefly, and settles. The intact shoulder returns to rest.
Severe neural and muscle damage leaves residual motion but prevents useful grasping.
Destroying feeder cells prevents absorption even when material reaches the body.
A learned cellular motion loop across the five organism families. Each point is a body cell.
The accepted grounded controller predicts causal muscle and contact feedback inside physical constraints.
A continuous anatomical graph VAE reconstructing held-out creature motion.
Target and reconstruction frames across families and motion phases.
The promoted runtime renders current continuous cell positions rather than selecting a sprite frame. The motion scaffold supplies the pose; the neural decoder predicts cell appearance, opacity, footprint, and sub-pixel placement.
| Breeding and symmetry | Cellular trauma | Neural cellular dynamics |
|---|---|---|
 |
 |
 |
These cover inherited chassis variation, organs and fluids, damage, healing, scarring, debris, and learned local cell-state updates.
| Living world | Map themes | Neural topology |
|---|---|---|
 |
 |
 |
The playable scaffold includes ecology, resources, organisms, construction, and settlements. Maps begin as validated semantic topology before art is rendered. Neural topology remains under evaluation and does not bypass the safety repair stage.
Teacher, rollout-aware V2, and foreground-aware V3 over one continuous untouched-world rollout. V2 removes most ghost geometry; V3 improves foreground and edge fidelity. Later frames remain visibly softer than the teacher.
| World VAE | Sparse Action-DiT |
|---|---|
 |
 |
The world VAE reconstructs held-out frames. The sparse Action-DiT prototype learned localized action edits but did not pass the stricter latent acceptance gate, so it is evidence rather than runtime authority.
This gallery stays small enough to browse on GitHub. Full corpora, checkpoints,
replay banks, failure cases, and evaluation reports remain under the local
outputs/ tree and are not included in normal Git history.