The model
drives.

The T-Rex simulator asks for jump, duck, or run. Its physics planner describes each choice as safe or unsafe and marks a preferred action. The primary condition disables the deterministic safety shield: the selected model action stands. The second condition restores the original shield, which can replace an unsafe choice or act before an answer arrives.

Survival means zero deaths over one 60-second course. A crash restarts the course after 1.5 seconds, so one seed can record multiple deaths. Agreement is the mean of the five per-seed agreement rates. Median arrival is the median of five per-seed p50 values.

Every run used the original course, seeds 0-4, 60 wall-clock seconds per seed, the labeled prompt, and six requests in flight. Each model played independently on the same client host. Its actions and latency change subsequent game states, so actual game requests are not identical between models.

This is a constrained direct-action proxy, not an open-ended software agent. The planner already supplies the safety labels. Shielded survival measures the combined model-and-planner system. Five seeds provide a small sample; these outcomes cannot establish general agent safety.

The CLM runs below are the earlier five-seed runs on the same harness. They are historical observations, not additional trials in this new evaluation.