Golden Solution

A golden solution is a fully worked, reproducible reasoning trail attached to a task, written so that a reviewer with no prior context could follow the exact same steps and independently arrive at the same conclusion. It states the assumptions being made, the formulas or rules being applied, and the logic connecting each step to the next, rather than just asserting a final answer.

A golden solution is easy to confuse with a rationale, but the two serve different purposes. A rationale justifies a choice already made: it explains, after the fact, why Response A was preferred over Response B against a rubric. A golden solution is built to be reproduced: it is written before or independent of any comparison, as the reference path someone else could re-derive the same result from, using only what is written down. One defends a judgment call; the other hands over the work.

Golden solutions matter most on tasks with objective or semi-objective answers such as math, coding, structured extraction, and multi-step logic problems. They give reviewers a way to verify a labeler's answer without re-solving the problem from scratch, and they give AI labs a ground-truth reasoning trace to train and evaluate chain-of-thought behavior against, not just a final answer to match.