Literature Synthesis

Synthesis is not summary. A summary lists what each source says; a synthesis produces a structure those sources collectively imply, including where they disagree and what none of them answer.

Build a claim-evidence map, not a reading list

1. Extract claims, not papers: decompose each source into discrete claims with their evidence type (theory, simulation, observational, RCT, anecdote) and strength.
2. Cluster by claim, then see which sources support, contradict, or qualify each. The unit of synthesis is the claim, agreed/disputed across sources.
3. Weight by evidence quality and independence: ten papers citing one original study is one data point, not ten (citation cascades manufacture false consensus).

Surface the structure of disagreement

• When sources conflict, don't average them — diagnose why: different populations, definitions, time periods, methods, or incentives? The disagreement is often more informative than either claim.
• Separate settled (broad independent agreement), contested (credible sources disagree), and unknown (no one has tested it). Most fields have far more "unknown" than people assume.
• Track the provenance and recency of each claim; fields move, and the most-cited result is often the oldest, not the most correct.

Adversarial validation

• Publication / file-drawer bias: the literature over-represents positive, novel, surprising results. The absence of a finding may mean it failed to publish, not that it's false.
• Citation laundering: a speculative claim in an early paper becomes "established" through repeated citation without re-verification. Trace to the primary source.
• Definitional drift: are all sources measuring the same construct, or the same word for different things?
• Conflicts of interest / incentives: who funded it, what does the author gain from the conclusion?
• What would the strongest dissenting expert say, and is that view represented or filtered out?

Non-obvious risks

• Consensus among non-independent sources is not evidence; it's correlation of error.
• The framing of the survey determines its conclusion; choosing the inclusion criteria is choosing the answer.
• Recency bias and seminal-paper bias pull in opposite directions; name which one you're indexing on.

Output structure

Deliver: a claim-evidence matrix (claim × sources × support/contradict × evidence strength), an explicit settled/contested/unknown partition, the diagnosis of each major disagreement, the gaps no source addresses, and the synthesized position with its confidence and what would change it.