If I am an accident oracle -- say I know with certainty where, say, 50% of accidents will occur (the other 50% are truly stochastic with no structural component at all) -- and the city believes me, then surely when I tell them an accident will occur at Intersection X, they will take measures that prevent accidents at Intersection X.
But this also means any measure of my ability to predict accidents is confounded, because the equilibrium behaviour would be that I never predict an accident and plenty of other accidents occur. Moreover, because of all the other confounders, it's actually unclear whether we should expect accidents to go down or up or stay the same or what, and to the extent year to year variability was already quite high the problem will be even larger, so even high level numbers before and after hiring my services aren't easily interpretable.
Which is fine, you just need to convince decision makers of this particular inferential fallacy and then hopefully they keep listening even though the KPI is wrong. Except what if the type of accidents change and I'm no longer an accident oracle? Then the mitigation efforts are wasted, and also the money they pay me.
One solution would be to basically engage in some kind of RCT where some of my predictions are held out for assessing my model while others are acted on and where the strength of the prior about my effect decays. Good luck telling voters and lawyers that justification, though.
Fun stuff.