Incorporate human evaluations, ratings, and expert corrections into the evaluation loop.
Captures reviewer corrections and judgements and routes them back into evaluation sets, prompt refinement or model tuning. Turns the review that was happening anyway into a source of improvement.
Wherever human review is already mandatory, since the marginal cost of capturing the judgement is near zero.
Where reviewers lack the expertise to judge correctly, in which case the loop amplifies error rather than correcting it.
Reviewer disagreement must be tracked rather than averaged away. Consistent disagreement between qualified reviewers usually means the task is underspecified, not that one of them is wrong.
Feedback from disengaged reviewers who approve by default, which teaches the system that its current output is correct.
Every rejection of a machine-drafted reconciliation explanation captured with the reviewer's reason, building a corpus of what inadequate justification looks like in that organisation.