Evaluate it without deploying it.
The output of this runtime is not only a safer trajectory. It is a record: which actions were available, which were refused, and on what evidence — generated by operation rather than written afterwards.
Shadow mode
The floor records its verdict even when it is switched off. Run it disabled against a recorded mission and it produces a counterfactual trace: every action it would have refused, with the evidence it would have cited, while changing nothing about what actually happened.
That is the honest first engagement. It costs nothing operationally, it cannot affect an outcome, and it answers the only question that matters before trusting a safety layer: what would this have done to my mission?
Two numbers fall out immediately, and they are the two a programme office should ask for: how often would it have refused something the crew did anyway and the mission was fine (false restriction), and how often would it have refused something that preceded a bad outcome (justified restriction).
The safety-case artifact
Assurance for autonomy is usually a document written about a system. This is a record produced by the system, in operation, with the same determinism that makes it replayable.
- Per-decision. Not a summary. Each admitted and refused action, at the moment it was decided.
- Evidence-linked. Each refusal carries the rule and the events that earned it — 30 rules expanded into 203 chained evidence tiles in the measured case.
- Replayable. Deterministic in, deterministic out. The same recording produces the same verdicts, which is what lets an assessor check rather than take your word.
- Counterfactual. Retained even when the floor is inert, so the layer can be evaluated for as long as you like before it is ever given authority.
What it looked like in the drive case
In the CARLA demonstration an adversarial controller drives toward a collision. The floor takes the vehicle, holds a ~9.7 m gap, and logs approximately 400 refusals across 100–490 m at 38–40 km/h. Hits: 0. With the floor off, the same controller crashes — and the log still records what the floor would have refused on the way in.