synthesis drift

Synthesis drift describes a systematic failure mode in LLM pipelines where the correct source material is retrieved but the generated output misrepresents it—paraphrasing, fabricating, or mis-citing direct quotes at the presentation layer rather than the retrieval layer. Measured failure rates are consistently alarming: across multiple studies, between roughly 29% and 56% of synthesized quotes could not be located verbatim in the original source, even when the underlying retrieval was sound. Critically, this drift cannot be caught by prompting the model to verify its own output, because the same synthesis process that produces a hallucinated quote will hallucinate its confirmation, making self-checking a category error. The implication is that detecting synthesis drift requires external, deterministic output-side verification—such as grep re-checks against the source—rather than relying on retrieval improvements alone, since the two failure types demand fundamentally different fixes.