What is breaking the tie in that AI ranking?
A ranking can show three decimal places even when the deciding signal measures something other than what you want to pick.
Imagine a tool that ranks short video clips so you can pick the funniest one. Two clips come back at 7.420 and 7.390. Before trusting the order, ask what contributed the 0.030 difference. If both clips got the same score for visible laughter and the rest came from changes in the audio, a louder reaction can come first even when the quieter clip has the better setup and payoff.
That split showed up in one clip-ranking test. The model’s laughter judgments gave the selected clips the same starting score. Smaller differences in the audio signals then did most of the work of ordering them. That was enough to sort the list. It left the question of which clip was funniest open.
If you build a ranking like this, show the signals separately. A label such as “visible laughter” and a separate audio score tell an editor why two clips came out in that order. Then try a pair you would rank differently, such as a quiet payoff and a loud reaction with the setup missing. If the tool keeps putting the loud one first, rounding the score may make the list easier to read. The same signal will still choose what sits at the top.