Encounter-set construction is the program
Most detect-and-avoid programs we are asked to verify have an encounter-set library that was built around what was easy to script, not around what the airworthiness authority needs to see. We almost always rebuild the library at the start of the engagement.
The pattern that produces the bad library is straightforward. The team builds a parameterized encounter generator early, finds a set of parameters that produces a wide variety of geometries, and runs the simulation-of-record across the parameter sweep. The resulting library has tens of thousands of encounters and looks comprehensive. It is not. The parameter sweep reflects the convenience of the generator, not the geometry distribution the platform will actually see in operation.
The library that survives review is anchored in the operational design domain. We start with the airspace classes the platform is going to operate in, the traffic mix in those classes (drawn from the responsible flight-information region's published statistics), the vertical profile of that traffic, and the approach-departure geometries at the airports the platform will operate near. The encounter set is then built to span that distribution, with deliberate over-sampling of the geometries that are operationally relevant and rare.
The investment that pays back the most is in the sampling itself. A library of 1,000 encounters that reflects the operational geometry distribution is more defensible than a library of 50,000 encounters that reflects the parameter sweep of a convenient generator. The smaller library also runs faster, which means the iteration loop on detect-and-avoid algorithm changes is shorter. Both of those matter.
The follow-on point is on the encounter-classification engine that consumes the library output. Our practice is to bin encounters by their post-encounter outcome (well-clear maintained, well-clear violated by N seconds, near mid-air collision threshold, etc.) and to track the distribution of outcomes over time as the algorithm changes. The trend is the program-management artifact. The single number is not.