TRACE is a five-step way to check research before anyone acts on it: Target, Retrieve, Assess, Cross-check, Export. You tie the question to a decision, get the actual source documents, judge what each one can prove, compare them claim by claim, and hand over a short memo that says what is supported, what is still open, and who decides next.
TRACE comes from The Verified Research Loop by Len P. van der Hof, available in English and Dutch. You do not need the book to use it. Everything below works with a notepad and the links you already have.
Why would you need a method for this?
Because a research summary can look finished long before anyone knows what it proves. AI assistants and search engines make it cheap to collect ten links and write a confident paragraph. The paragraph reads well. Nobody checks whether the ten links say what the paragraph says.
Imagine a memo that says a competitor has launched a new feature for all paying customers. It cites the competitor’s announcement and two news articles. Every link works. But both articles repeat the announcement, and the announcement describes an invitation-only preview.
Three documents, one origin, and a claim broader than any of them. TRACE is built to catch that gap before it turns into a roadmap change, a hire, or an investor update.
What are the five steps?
Each step answers one question and hands one piece of work to the next.
| Step | The question it answers | What you hand on | Go back or stop when |
|---|---|---|---|
| Target | Which decision is this for, and what exactly must we find out? | A short brief: decision, owner, deadline, the precise claim, what is out of scope | Nobody can say which decision the research serves |
| Retrieve | Which sources did we actually obtain? | A source list: title, date or version, and how much of each you could read | You only have search snippets or an AI summary |
| Assess | What can this source prove about this claim? | A note per source: its role, its method, its limits for this claim | The source covers a different group, place, or date |
| Cross-check | What does each source add, contradict, or repeat? | A claim ledger: one row per source per claim, with a status | Several sources turn out to share one origin |
| Export | What may the decision owner do now? | A dated memo: supported, inferred, open, owner, next check | A gap that matters has no owner |
In plain words:
Target. Turn a topic into the question a decision depends on. “Tell me about Competitor X” is a topic. “Before Friday’s roadmap review: is Competitor X’s feature available to all paying customers in the EU?” is a target. Write down who decides, by when, and what would make you stop.
Retrieve. Get the source itself, not a description of it. A search snippet, an abstract, or an AI summary is a lead. Record the title, the date or version, and whether you read the full text or only part of it.
Assess. Ask what this particular source can show about this particular claim. The competitor’s announcement proves what the competitor announced. It does not prove who can use the feature today.
Cross-check. Give each source its own row against the claim and mark whether it supports the claim, contradicts it, adds a missing piece, or only gives background. Then find out where each source got its information.
Export. Write a short, dated memo that keeps three things apart: what the sources support, what you infer, and what you recommend. Every open question gets an owner and a date to check again.
TRACE is a loop, not a conveyor belt. If Assess shows the question was vague, you go back to Target. If Cross-check finds that two sources share one origin, you go back to Retrieve for an independent one, or you record the dependence honestly.
Which sources count most?
TRACE sorts sources into four roles. The role tells you what to ask next. It never tells you the answer.
- Primary: close to the event, measurement, rule, or original analysis. A company filing, a dataset, a contract, a court judgment, a system log.
- Secondary: built from other sources. A news report, a review article, an analyst report.
- Advocacy: made by a party that wants something from the reader. A vendor case study, a campaign report.
- Grey: relevant material outside those three, such as a forum post, a working paper, or an unattributed slide. Grey does not mean anonymous or false. It means the label alone tells you little, so the content decides.
The trap is treating the role as a score. A real case shows why. In May 2020, The Lancet published a large study of hydroxychloroquine and chloroquine in COVID-19 patients, said to draw on data from 671 hospitals on six continents. According to STAT, its findings led to the pause of some international clinical trials of the drug.
In June 2020, three of the authors retracted it. Independent reviewers could not get the full dataset from the data company, Surgisphere, and the authors wrote that they could “no longer vouch for the veracity of the primary data sources.” A peer-reviewed primary study in a leading journal, and the claim could not bear weight. The retraction notice records the failed data access. It does not, on its own, prove the data were invented, and TRACE would record exactly that limit.
Two sources that agree: is that enough?
No. This is the most common shortcut in research checks, and TRACE rejects it on purpose. What matters is whether the sources reached their conclusion by different routes: different data, different people, different methods. Three articles quoting one press release are one source with three web addresses.
TRACE uses four plain status words for each claim:
- Corroborated: supported by origins that are distinct enough for this claim.
- Partial: some support, but a gap, a limit, or a shared origin remains.
- Unresolved: not settled within the scope you set.
- Conflicting: serious evidence points both ways. Keep both sides visible.
The reverse is also true. One primary record can support a narrow claim about itself, such as what a contract says. If only one source exists, write that down. Do not go hunting for a token second link to make the row look complete.
What does the finished memo look like?
Short. It opens with the decision and its current status. It gives the answer while keeping evidence, inference, and recommendation apart. It links each important claim to its row in the claim ledger, lists what is still open and who disagrees, and ends with actions, owners, and the event that would reopen it.
The decision status uses a fixed set of words: Proceed, Conditional proceed, Pilot, Defer, Reject, Specialist review, Accepted risk, or HOLD. HOLD means the action waits until a named person resolves the gap or formally accepts the risk. Accepting a risk changes what you do. It does not make the evidence any stronger.
A worked example: did the competitor really launch?
This is a hypothetical, about half an hour of work.
- Target. Decision: do we pull our own feature forward at Friday’s roadmap review? Claim to test: “Competitor X’s feature is available to all paying EU customers this week.” Owner: the head of product.
- Retrieve. The competitor’s announcement, its help-center page on availability, two news articles, and one forum thread. All read in full, all dated.
- Assess. The announcement is the company’s own record of what it announced. The help-center page is the company’s own record of who has access. The two articles are secondary. The forum thread is grey: one user saying they have access.
- Cross-check. Both articles say “launched”, and both quote the announcement. The announcement says “invite-only preview”. The help center shows a waitlist. One invited forum user fits a preview. Status for general availability: Unresolved. The narrower claim, “a preview was announced”, is supported by the company’s own record.
- Export. The memo reads:
Supported: Competitor X announced an invite-only preview (announcement and help center, both dated). Not supported: general availability; the two “launched” articles repeat the announcement. Decision: HOLD on pulling our feature forward. Owner: head of product. Recheck: when the help-center waitlist disappears, or on 15 October.
The memo is narrower than the original claim. That is the point. It is also the version a colleague can check in two minutes.
What TRACE is not
- Not a two-source rule. Origins count, not links.
- Not a credibility score. Primary, Secondary, Advocacy, and Grey are roles. A primary source can be wrong, out of date, or beside the point.
- Not RAG. RAG (retrieval-augmented generation) is software that fetches text for an AI model before it answers. It can help with Retrieve. It does not judge what the fetched text proves.
- Not a guarantee. The book says plainly that TRACE has not been shown to make research more accurate or faster. It makes the evidence behind a decision visible, so someone else can check it. Finishing the five steps proves the work was done, not that the conclusion is true.
Try this today (15 minutes)
- Pick one sentence someone will act on this week, in a memo, a deck, or a chat thread.
- Rewrite it as a precise claim: who, what, where, and as of when.
- Open the source behind it. The source itself, not the snippet or the AI summary.
- Label its role: Primary, Secondary, Advocacy, or Grey.
- Check whether the other links behind the sentence share its origin.
- Write one line under the sentence: supported, partial, or unresolved, plus a name and a date to recheck.
If step 3 fails, you have found the most useful thing TRACE can show you.
The templates for each step (research brief, source list, claim ledger, decision memo) are in The Verified Research Loop. The TRACE framework page has the one-page summary.
Cite this:What is TRACE? Five steps to check research before you act on it.Len P. van der Hof. https://lenvanderhof.com/en/blog/what-is-trace/ ·