Current LLM summarizers produce confident, structured outputs that often include claims unsupported by the source material. These systems invent sections, infer from ambiguous statements, and pattern-match from their training data without verification. The fundamental issue is that summarizers skip the identification step—determining what the actual transcript can support—before generating claims, resulting in outputs that appear legitimate but may be fabrications.

The proposed solution requires each claim to declare its support category: observed, inferred, or recommendation. This approach uses a three-stage pipeline with strict constraints. The first extracts facts conservatively, the second synthesizes them into labeled claims, and the audit stage can only weaken or remove unsupported claims, not improve them. The system literally inserts placeholders where evidence is insufficient, making the limitations visible to the reader.

This design acknowledges that LLMs will sometimes make unsupported claims but prevents them from being presented as facts. By limiting reviewers to only weakening claims rather than improving them, the system tolerates excessive caution but avoids inventing unsupported conclusions. The resulting summaries accurately reflect what the meeting actually contained rather than what an algorithm thinks a meeting should contain.