For the same near-miss reports, safety leadership later asks for a second capability: 'once a report is tagged high severity, give the reviewer a two-sentence plain-English digest of what happened.' How should this be added to the existing design?

Adding summarization as a conditional second stage - triggered only for reports the classifier has already flagged high severity - is correct because it satisfies the new requirement (a digest, but only for the reports that need urgent attention) while preserving the working triage behavior leadership already relies on. Replacing classification with summarization entirely is wrong because it would remove the low/medium/high triage queue routing that the safety team still needs for every report, not just the high-severity ones; the new ask is additive, not a replacement. Re-running classification with longer category 'names' misunderstands the request - leadership wants a two-sentence digest of the incident's content, not a wordier label, and stretching label text is not the same task as summarizing free text. Writing the severity label into the vector index as metadata might be useful for later filtering, but it does not produce the two-sentence digest a reviewer reads, so it does not satisfy the actual ask on its own. This item tests chain design as composition: adding a second, conditionally-triggered model task onto an existing pipeline rather than replacing or misinterpreting the working stage.