Strategic experiment: executable game worlds as verifiable spatial data engine

Finding

Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models (arXiv:2608.25518, submitted 2026-08-26) argues that executable game scenes provide denser and more objective spatial reward than fuzzy perceptual proxies. Collision, physics, navigability, bounded playability, and developer acceptance can jointly supervise long-horizon spatial trajectories.

RuV mapping

WorldGraph, RuField, RuView, Dream Machine, MetaHarness, RuVector, and embodied/spatial Cognitum workloads.

Hypothesis

Use procedurally generated executable scenes as a synthetic verification layer for spatial reasoning components before transferring to RF/LiDAR/real-world data. The engine is an evaluator, not a ground-truth replacement for physics or sensing.

Reversible experiment

Generate a small corpus of versioned scenes with explicit object state, collision meshes, navigation graph, occlusion, sensor poses, and deterministic engine checks. Compare a spatial policy trained/evaluated on static image/video proxies against one receiving engine-verifiable trajectories. Then test both on held-out real RuView/RuField trajectories.

Falsification

Reject if gains disappear on real trajectories, engine-generated scenes induce shortcut learning, or the executable reward improves simulator metrics while degrading RF/LiDAR transfer.

Acceptance criteria

At least 10% relative improvement on two held-out real spatial metrics with no worse uncertainty calibration and complete scene/trajectory provenance. No production promotion from simulator-only gains.

Reference: https://arxiv.org/abs/2608.25518

The issue names no files, tests, or entry points; use the proposed WorldGraph, RuField, RuView, Dream Machine, MetaHarness, and RuVector mapping as the initial research map. Done means a reversible scene and trajectory experiment reports the stated held-out real-trajectory metrics, uncertainty calibration, provenance, and simulator-only promotion safeguard.