Nonfiction fact-checking often fails for a structural reason, not a diligence reason. The writer reads the manuscript carefully, checks the sentences that look risky, and moves on. Then a chapter gets reorganized, a number gets updated, and the careful pass no longer describes the book that exists. The check was real; the system did not survive the revision.
The fix is to stop treating fact-checking as a reading task and start treating it as a ledger. A ledger has a unit, a status, and a link to the text. When the text changes, the ledger tells you what to re-check instead of asking you to remember.
1. Start with a claim inventory, not a page
The unit of fact-checking is the claim, not the chapter or the sentence. A claim is any assertion the argument leans on: a number, a date, a causal statement, a comparison, a quotation, or a characterization of what a source says. If a reader could reasonably ask “says who?”, it is a claim.
Before you check anything, enumerate the claims. For each one, record four things:
- The claim — stated in one sentence, in the words the manuscript uses.
- The source or sources — the specific document, dataset, interview, or page.
- The evidence status — not just whether a source exists, but what it does: supports, weakly supports, disputes, contradicts, or only mentions.
- The manuscript anchor — the passage the claim supports, precise enough to re-find after an edit.
A claim you cannot enumerate is a claim you cannot check. This is the first place a manuscript loses its thread: the writer remembers the argument but not every assertion the argument depends on.
2. Triage by claim type
Different claims fail in different ways. Sorting them by type tells you what kind of check each one needs.
Numbers
Check four things: the figure, the year, the population, and the denominator. A percentage without a base is not a fact. A number that traveled from a study to a review article to a press release is a number that may have changed meaning at each step.
Causal claims
“X causes Y” needs a source that actually tested the link. A source that mentions both X and Y is not enough. This is a common substitution error in nonfiction: a real source is attached to a causal claim it never supported.
Comparisons
A comparison needs a source for each side and a consistent baseline. “More than” compared to what, measured how, in which population? If the two sides come from different studies with different methods, say so rather than presenting the comparison as clean.
Universals
“All,” “every,” “never,” “always,” and “proves” are the hardest claims to support. One counterexample defeats them, and one source rarely covers every case. Often the honest fix is to narrow the wording, not to find a bigger source.
Attributions
If the manuscript says someone argued, found, or claimed something, name the person or document and the date. An unnamed attribution is not a source; it is a rumor with a citation-shaped hole.
3. Check the source, not just the citation
A citation is a pointer. It is not evidence that the pointer supports the sentence. The check is always the same question: does this source support this claim as worded?
That question has a few reliable failure modes:
- Adjacent support. The source supports a nearby claim, not the one on the page.
- Stale support. The source was correct when you read it, but it has been corrected, retracted, or superseded.
- Secondary support. The source is a summary of a study, and the summary overstates what the study found.
- Quote drift. The quotation is accurate in isolation but changes meaning outside its context.
None of these are caught by checking that a citation exists. They are caught by opening the source and reading the passage the claim depends on.
4. Record the decision, then re-run the pass
For each claim, record the status and the reason. “Supported” without a note is not a decision; it is a mood. A useful record says what you checked and what you concluded, so the next reader does not have to repeat the work.
Then accept that the pass is not permanent. Structural edits break anchors. A paragraph moves, a quotation gets trimmed, and the link between the claim and its source goes stale. A ledger with anchors can tell you which links broke; a mental note cannot.
This is also where cross-chapter contradictions surface. Two chapters can make incompatible claims from different sources, and neither chapter looks wrong on its own. A whole-manuscript view is the only place that conflict becomes visible.
5. A five-claim spot check
If you want to see the method before you build the whole ledger, take five claims from your draft and check them with the free source-gap check. Paste one claim per line and it returns which claims name no source, which numbers and causal claims are unsupported, and what to attach next.
Be clear about what that check is: a gap check, not a fact-check. It finds missing or weak source links. It does not verify that a source is true, current, or correctly quoted. The human check is still the human check.
6. The bar is defensible, not certified
No tool can certify that a nonfiction book is true. What you can do is make every load-bearing claim traceable to a source a reader could check, record what that source does for the claim, and keep the link alive as the draft changes. That is a defensible manuscript. It is also the difference between a fact-check that survives revision and one that quietly expires.
The Claim Ledger is built for that job: every claim with its sources, evidence status, and manuscript anchor, plus the whole-manuscript exceptions that a single read misses. If you are weighing BellerWriters against Scrivener, Ulysses, ProWritingAid, or Grammarly, the named comparison says where each tool leads and what we are building.