Source triangulation & fact-checking.

Sources are not interchangeable CORE

A research agent that returns eight sources has not done research. It has done retrieval. What turns retrieval into research is the part where a human — you — treats those eight sources as eight different kinds of evidence and weighs them differently.

Three source tiers cover most of what you will see. Learn the distinctions; they will shape every triangulation call you make.

Primary sources are the original record of the thing. A bill’s actual text. A scientific paper reporting original research. A product’s own pricing page. A court’s opinion. An interview transcript. A dataset. Primary sources are the closest you can get to the thing itself; they are not interpretations of the thing. When you can cite a primary source, you usually should — it is the most defensible evidence.

Secondary sources are competent interpretations of primary sources by a human or institution that did real work: an investigative news article, a literature review, a reputable industry report, a well-researched book. They are not the thing, but they are close readers of the thing, and they explain it. Most good secondary sources cite the primary sources they are reading. A well-written brief usually relies on a mix of primary sources (for claims you need to stand behind) and secondary sources (for context and interpretation).

Tertiary sources summarize secondary and primary sources without original analysis. Wikipedia, encyclopedia entries, textbooks, overview sites. Useful for orientation — getting you into the shape of a topic fast — but rarely the right place to end. A research brief whose footnotes are all tertiary is a brief resting on interpretations of interpretations.

The rule of thumb: orient with tertiary, work with secondary, land on primary when it matters. If a claim is load-bearing — if its being wrong would sink the brief — chase it back to a primary source. If it is background color, a secondary source is fine. If it is a single term definition, tertiary is fine.

This is not a hierarchy of trustworthiness — a careful secondary source is often more useful than a sloppy primary one, and a clear tertiary source beats a confused primary one. It is a hierarchy of closeness to the claim. The closer you are, the more the claim you make is yours to defend; the farther, the more you are borrowing someone else’s defense.

Triangulation: the three questions CORE

Once you have a set of retrieved sources sorted into tiers, triangulation is the move that turns the set into an answer. Triangulation answers three questions about every important claim.

Is this supported by more than one independent source? Independent matters. If three news articles all cite the same press release, you have one source pretending to be three. If two academic papers from the same lab report the same finding, that is weaker than two papers from two unrelated labs. Count independent sources.

Do the sources agree? If they agree, note it and move on. If they disagree, do not paper over it — the disagreement is the story. Name it in the brief. Say which sources take which position and why (often the why is methodological — they measured different things — or temporal — one is newer). A research brief that pretends disagreement does not exist is less useful than one that names it honestly.

If a claim rests on only one source, what kind of source is it? A single primary source can be enough for some claims (the product’s pricing page for what the product costs). A single secondary source is rarely enough for a contested claim. A single tertiary source is almost never enough. If a load-bearing claim rests on one source, your job is either to find another or to mark the claim as “per [Source X]” so the reader knows the weight.

The triangulation output looks something like this in your notes:

Claim: The four-day school week is associated with lower academic outcomes in middle school.

Supported by: two studies in peer-reviewed journals (Thompson 2023, Garcia & Lee 2024); one RAND brief summarizing both.

Disagreement: the Oklahoma State Department of Education report (2024) argues the effect disappears when district-level controls are added.

Weight: moderate — two independent studies support it; one methodological critique complicates it; a careful brief should name both.

That is triangulation. Not “I found some sources.” This is the claim; this is what is under it; this is where it disagrees with itself. A research agent can produce a rough draft of this. You check every line by opening the sources.

Four patterns of a fabricated citation CORE

Language models confidently produce citations that do not exist. This is well documented, happens in the Cowork tab and in every comparable research-agent tool, and is the number-one failure mode of research agents. You must assume it will happen in your own work and build the habit that catches it.

Four patterns cover most fabrications.

1. The completely invented source. A fake paper title, by a fake author, in a real-sounding journal, with a year and page numbers. Sometimes the DOI or URL will even resolve to something — rarely the claimed paper. The smell test: you cannot find the source when you search its exact title. The agent occasionally produces a quote from this non-existent paper that is rhetorically perfect for the claim, which is another tell — real sources rarely contain a single sentence that ideally supports a specific argument.

2. The real source, fabricated quote. The paper exists. The author exists. The year and journal are correct. The quotation the agent produces is not in the paper. This one is especially insidious because the reader who checks the top of the citation and does not read the paper itself will see a real title-author-year combination and trust the quote. Defense: open the source and find the quote. If you cannot find it, it is not in there.

3. The real source, misrepresented finding. The paper exists. The citation is correct. But the claim the brief attributes to it is not what the paper says — or is what the paper explicitly argues against. This is the hardest fabrication to catch because it requires reading the source carefully enough to see what it actually claims. Defense: read the abstract and the conclusion at minimum.

4. The URL that does not exist, or the URL that is not what was claimed. The agent cites https://example.gov/report-2024.pdf✓ as supporting a claim. The URL 404s. Or the URL loads a page that is not about the claimed topic at all. Defense: open every URL. This is cheap and fast.

Pattern 1 is catchable by search. Pattern 4 is catchable by clicking. Patterns 2 and 3 require opening and reading — which is, again, why the rule is you do not cite what you have not opened.

The open-before-cite checklist CORE

This is the operational drill for Module 4 and the heart of the module’s safety norm. For every source that will appear in a final research output, run the following four-step check. On every source. Every time.

1. Open it. Actually open the URL in a browser, or actually read the file. If the URL does not load, or the file does not exist, the source is dropped from the brief. (Do not “keep looking” for the right URL — the agent may have invented the whole thing.)

2. Confirm the source is what it says it is. Is the author real? Is the publication real? Is the date real? A thirty-second sanity check: search the author’s name, see if the publication exists and is reputable, check the page is actually on the publication’s real site and not a typo-squat.

3. Find the claim the brief attributes to this source. Skim the source for the specific sentence or finding the brief is using it for. If the claim is not in the source, the source is dropped. If the claim is partially in the source but the brief is overstating it, rewrite the brief’s sentence to match what the source actually says.

4. Mark the source in your notes. In your source list (the folder you are saving sources to, per the recipe below), mark each source as opened-and-confirmed, opened-and-dropped, or not-yet-opened. A source that is not-yet-opened at the time of shipping does not get cited. Full stop.

This is a short, bounded investment per research output. It is the only investment in this module that cannot be delegated to the agent. And it is the one that decides whether your work is trustworthy or is plausible-looking fiction.

The fact-check memo CORE

Your capstone entry 2 is a fact-check memo on your own topic. A fact-check memo is a short, source-anchored document that answers one bounded question: “Is claim X, as stated, actually true?”

Fact-check memos are useful in a way that full briefs are not: they force you to take one claim seriously, open every source that bears on it, and deliver a verdict rather than a summary. They also demonstrate the open-before-cite habit at its sharpest, because the whole point of the document is whether the evidence holds.

The shape:

1. The claim under check, quoted or paraphrased in one sentence, with its source (where you first heard it).
2. Verdict, in one word: true, partially true, unclear, false, or unsupported.
3. Evidence, 300–600 words, in which you list every source you checked (by tier), describe what each one says about the claim, triangulate, and explain your verdict.
4. Confidence, one line: low, medium, or high. Say why.
5. What would change your mind. One sentence: what piece of evidence, if it appeared, would move you off the verdict.

That last line is the integrity line. It forces you to name the thing you would need to see to change your answer. If you cannot name anything, your verdict is religious, not research.

Pick one specific contested claim inside your Module 4 topic. Not the topic itself — one claim inside it. For the solar-panels topic it might be “Ohio utilities will buy back excess generation at the full retail rate.” For the curriculum topic it might be “Beast Academy covers the 7th-grade math standards by the end of its sequence.” For the Reconstruction paper it might be “The Civil Rights Cases of 1883 were decided 8–1.” Claims with a yes / no / partially answer make the best fact-check memos.

Run the checklist on every source. Drop sources the checklist fails. Write the memo against the sources that remain.