Existing meeting-summary evaluations mainly ask whether a summary resembles a reference or is consistent with the source. They do not directly show which information roles needed for decision understanding are retained or dropped.

We propose SIEVE, a method that evaluates meeting summaries from seven information-role perspectives. For each role, SIEVE separately measures Presence, whether the role appears in the summary, and Fact Check, whether the retained content is free of contradictions with the source. This produces a role-level retention profile, allowing summarizers and length limits to be compared by what they preserve and lose rather than by a single aggregate score.

We apply SIEVE to 2,250 summaries from 50 MeetingBank municipal-meeting segments, generated by three models under 2-, 3-, and 6-sentence limits. The role-level profiles show that summaries more reliably retain topic, outcome, and action-detail information, while more often dropping background and concerns that shaped the discussion, as well as unresolved items left open afterward. Reference-similarity metrics widely used to evaluate MeetingBank summaries (ROUGE-L, BERTScore) correlate weakly or negatively with SIEVE Presence. Consistent with this, the dataset-provided reference summaries against which similarity is computed contain topic and action-detail information but few of the other roles. Reference similarity and the role retention that SIEVE measures are different perspectives on summary quality. SIEVE therefore supports auditing, selecting, and designing meeting summarizers based on which decision-relevant information survives compression.