LLM Summarizers Skip the Identification Step

LLM meeting summarizers often invent decisions and action items that never occurred, confidently presenting fabricated content to fit a predefined format. This failure is analogous to a classic error in causal inference where a conclusion is drawn without first establishing that the data can actually support the claim. A more rigorous approach forces the summarizer to treat every output as a claim that must be categorized as directly observed, inferred, or a recommendation, and linked to specific evidence. This system uses a constrained audit stage that can only weaken or delete unsupported claims, preventing it from inventing content to fill gaps. As a result, summaries accurately reflect a meeting's substance by leaving sections empty or noting insufficient evidence, making the lack of a decision as informative as its presence.