For materials literature summarization, data extraction, and report-draft generation, it would be good to first define a validation rubric. For example, one could create items such as composition match, phase-name match, temperature·pressure range match, whether controls are mentioned, and consistency with cited sentences, and then have each output checked step by step to see whether it satisfies these items.
This approach may be especially useful when the goal is to reduce answers that are “plausible but wrong.” However, if the rubric is too strict, valid interpretations may also be rejected, and as the validation steps increase, response time and computational cost may also increase.