[alganet]

What I described happens in the raw offline model too. Those don't have post-inference heuristics such as those you described, implying the bias is baked in the training data or fine tuning steps.