What a survivability witness actually proves

Connectivity is not enough

In Gradient Gators, a path to the goal can still be unusable if hazards cause the agent to run out of health. Before a scenario enters a controlled experiment, it helps to know that at least one candidate trajectory can reach the goal alive.

I implemented a survivability witness system that considers shortest-path and damage-aware candidates, adds bounded waiting variants, and replays those candidates through the environment's mechanics.

Replay makes the evidence concrete

A geometrically valid route is a proposal. Replaying it under the same rules used by the environment checks whether that particular trajectory survives. This keeps the claim tied to concrete behavior instead of assuming that graph connectivity implies a feasible episode.

The boundary matters: a successful witness demonstrates one feasible trajectory. Failure to find a witness in this bounded search does not establish that a scenario is impossible. The procedure is also not a proof that the chosen trajectory is optimal.

A foundation for repeatable comparisons

Witnesses fit into a larger pipeline of frozen scenario collections and deterministic schedules. Stable scenario identities and controlled ordering let us inspect what each training run was given.

That infrastructure supports the upcoming reward-design comparison. It does not replace evaluating the trained policies themselves. Keeping those responsibilities separate makes the eventual conclusions easier to interpret and reproduce.