Internal evaluation is currently based on cost values, which are usually derived from duration via vehicles costs.per_hour factors.

But think of a situation where a user is providing custom costs matrices to model the cost of "reaching out to a given zone", typically when taking subcontracting costs into account. Then all costs for locations from zone A to zone B will be identical, and all intra-zone costs will be zero. In this situation comparing based on cost only means all intra-zone orderings in routes have the same cost so we can provide an arbitrary (bad duration-wise) ordering for zones.

So we should compare Eval instances based first on cost, then on durations in case of draw, which makes sense anyway.