RT-SAFE: presentation methodology

Source: the supplied Real_Time_Safety_Bench__RT_Safe_.zip manuscript archive
from RT-Safe Paper, copied 1 October 2026. Website values are extracted from its
LaTeX tables by scripts/extract-data.py.

TASK
36 routes across five urban maps in SimWorld / Unreal Engine 5.
Easy, medium, hard use nested 60%, 80%, 100% configured actor subsets.
Each model receives first-person RGB imagery with seven marked movement targets,
recent motion frames, navigation context, traffic rules, and action feedback.
16 actions: seven moves, six turns, three waits.

TIMING
Static: the simulator pauses while the agent reasons.
Real-time: the simulator advances concurrently with planning, including context
construction, the model request, response receipt, and parsing. The agent stays
in place while planning. The full response is processed before action dispatch.
The two modes also have different timing instructions in the prompts; this is a
joint intervention on timing and instructions, not a timing-only causal claim.

METRICS
Success: reached destination.
Safe success: reached destination with zero collisions, hazard interactions,
and traffic-rule violations.
Collisions: event counts per episode, including successful and failed episodes.
Active contacts are assigned to the action interval; passive to inference.
SPL: success weighted by path length relative to the reference route.
Response latency: seconds from model request to response; differs from full
inference exposure and includes the serving interface's behavior.

RESULT SCOPES
Hard real-time / static: 36 routes per model, eight models; 288 episodes per mode.
All difficulties: 108 episodes per model, equal weighting of easy/medium/hard.
Do not mix these scopes when ranking models.
Reasoning comparisons: three provider-specific effort levels on 36 hard routes.
Sol defaults to its lowest rung; the other seven default to the middle rung.
Offline training: a separate Qwen3-VL-4B study on 16 tasks from held-out maps,
with a fixed 3-second decision delay. This is not the eight-VLM protocol.

COUNTING LIMITATIONS
Repeated physical contact can generate multiple counts. Counts do not represent
unique people or injuries. Hazards use post-action overlap, not swept-path
checks. Water is a recorded flag without movement perturbation in this version.
Results describe the evaluated simulation protocol, not physical-world safety
certification. Independently rounded components may not sum exactly.

VISUAL SOURCES
Selected decision images: original 720 x 640 UE captures, with source metadata
in data/examples/case-1.json, case-6.json, and case-7.json.
Cover and social image: AI-generated concept illustrations.
Timing diagrams: explanatory schematics with illustrative positions and timing.
Film environment montage: actual UE pilot footage, condensed observations,
separate from the quantitative paper-campaign comparison.

The author list and public repository URL were not supplied as release metadata.
The downloadable manuscript retains its original anonymous submission identity.