Research & Synthesis Patterns

Six tabs open, four of them saying slightly different things, and a deadline in an hour. That's the actual shape of most research work — not "find the answer," but "reconcile several partial, sometimes-conflicting answers into one you can defend if someone pushes back on it." AI is fast at gathering; it is not automatically good at reconciling, and a synthesis that silently averages away a real disagreement is worse than no synthesis at all, because it looks trustworthy while being wrong.

This lesson gives you five named patterns for turning scattered sources into something you can stand behind. Each pattern has an explicit failure mode, because a synthesis pattern that quietly breaks is more dangerous than one that obviously doesn't apply — you keep trusting output that's silently gone wrong.

Pattern 1: Source-Triangulation

The pattern: Never conclude from one source. Require at least three, and check them against each other before accepting a claim as established.

I have three sources on [claim]. Before treating this as confirmed,
check: do any of these three cite a common upstream origin (the same
original report, the same press release, the same primary study)?
If so, this is one source wearing three bylines, not three independent
confirmations. Flag that explicitly.

Source 1: [...]
Source 2: [...]
Source 3: [...]

When it works: The three sources are genuinely independent — different organizations, different methodologies, arriving at the claim from different angles.

When it fails: All three sources trace back to the same original report, press release, or viral post. This looks like triangulation on the surface — three separate articles — but it's actually one unverified claim that's been repeated, not confirmed.

Pattern 2: Steelman-Both-Sides

The pattern: Before concluding on a genuinely contested question, build the strongest honest version of each competing position — not the weakest, easiest-to-dismiss version of the side you disagree with.

There are two competing claims about [topic]. Before I conclude,
build the STRONGEST version of each — the best evidence, the most
charitable interpretation, the case a smart, informed advocate would
actually make. Do not build a strawman for either side.

Position A: [...]
Position B: [...]

When it works: The disagreement is real — two credible sources or camps genuinely differ, and understanding the strongest form of each sharpens your eventual conclusion.

When it fails: One side has no credible support at all. Steelmanning a claim that isn't actually contested by anyone credible manufactures a "debate" that doesn't exist and hands the weak claim false legitimacy just by giving it equal argumentative weight.

Pattern 3: Claim-Evidence Pairing

The pattern: Every claim in the synthesis carries its source, explicitly, inline — not a bibliography at the bottom, a direct link between what's asserted and what supports it.

Write the synthesis so every factual claim is immediately followed by
its source in brackets, like this: "X grew 12% in Q3 [Source: Company
10-Q, filed Nov 2026]." If a claim has no traceable source, do not
include it — flag it as unsupported instead.

When it works: You have the full source list and can verify each pairing yourself before publishing — the discipline catches you if the model paraphrases a claim so far from the original that the citation no longer actually supports it.

When it fails: Sources have been summarized or paraphrased so many times upstream that a claim can no longer be traced back to anything verifiable. Pairing a claim with a citation that doesn't actually say what you're claiming it says is a fabricated citation with extra steps — always spot-check the pairing against the original.

Pattern 4: Contradiction Flag

The pattern: When sources genuinely disagree, surface the disagreement explicitly instead of averaging it into a single smoothed-over number or claim.

If any of these sources genuinely conflict on a fact (not a difference
in scope, time period, or definition — an actual factual disagreement),
flag it explicitly: "Sources conflict: [Source A] reports X, [Source B]
reports Y. This is a genuine disagreement, not resolved by averaging."
Do not blend conflicting numbers into a single average without flagging
the conflict first.

When it works: Sources genuinely conflict on the same fact, for the same scope and time period — a real disagreement worth surfacing to whoever is making a decision based on the synthesis.

When it fails: The apparent conflict is actually a scope or definition mismatch — one source reports a global figure, another a regional one; one reports year-to-date, another a single quarter. Flagging that as a "contradiction" (or, worse, averaging the two numbers together) hides the real nuance, which is that both are true statements about different things.

Pattern 5: Synthesis Grid

The pattern: Force every source into the same comparison structure — the same rows or dimensions — so you're comparing like to like instead of letting each source frame itself on its own terms.

Build a comparison grid with these sources as columns and these
dimensions as rows: [dimension 1, dimension 2, dimension 3]. If a
source does not address a dimension, mark the cell "not addressed" —
do not infer or fill it in.

When it works: The sources genuinely address comparable dimensions — same metric, same unit, same rough scope — and the grid makes differences visible that prose would bury.

When it fails: Sources use incompatible categories or metrics — one measures revenue, another measures user count, a third measures sentiment. Forcing three incompatible measurements into one grid manufactures a comparability that doesn't exist and can make apples look like they're competing with oranges.

Build It: Synthesis Validator

A synthesis pattern that lives only as a mental checklist degrades the first time you're rushed. A function that refuses to let an un-cited claim through is a pattern you can actually rely on.

Bottom Line

Five patterns, one shared risk: a synthesis that looks trustworthy while quietly smoothing over the parts that don't fit. Triangulate sources, but check for a shared upstream origin first. Steelman both sides, but only when both sides actually have a case. Pair every claim with its evidence. Flag genuine contradictions instead of averaging them away. Grid sources only when they're actually comparable. Applied together, these patterns turn six open tabs into one answer you can defend under scrutiny — not just one that reads clean. Next: analysis and data reasoning — the patterns for stress-testing a conclusion before you act on it.