Verify Claims in a Document: 7 Complete Checks to Run

A four hundred word post and a forty page report are not the same job, and the difference is not length. It is that a long document repeats itself, cites itself, contradicts itself and carries numbers in tables that nobody reads twice. It takes forty minutes to verify claims in a

A four hundred word post and a forty page report are not the same job, and the difference is not length. It is that a long document repeats itself, cites itself, contradicts itself and carries numbers in tables that nobody reads twice. It takes forty minutes to verify claims in a document of ten pages, and the first twenty are the summary and the tables.

To verify claims in a document of that size, seven checks do the work. Four of them are the same checks you would run on a page. Three exist only because the document is long, and those three are where long drafts fail.

This page is written for the report, the proposal, the white paper and the thesis chapter. If you are checking a web page instead, verify AI content before publishing is the shorter routine. Where two sections disagree, that is exactly the moment to verify claims in a document against the source rather than against each other.

[One long document, seven checks, and the three that only long documents need.]

Seven checks applied to a long document with three marked as length specific
Figure 1: Four checks from the page routine, three that only long drafts need.

Why long documents fail differently

Three structural reasons, and none of them is carelessness.

Repetition without a single source of truth. A figure appears in the summary, the section, the table and the conclusion. Four copies, one source, and a correction that touches three of them. The cheapest way to verify claims in a document is one pass with a ledger, not three passes with a pen.

Internal reference. A report cites its own earlier page: as shown in section two. If section two changes, the sentence on page nine becomes false without being edited. Anyone who has to verify claims in a document for a client will keep the report folder as part of the delivery.

Density in tables. A table of thirty rows holds thirty claims in two hundred words. Nobody proofreads a table, and a table is where a wrong number survives longest. A figure that appears three times is a single claim, so verify claims in a document once and fix every copy.

Add a fourth reason that is specific to documents rather than pages. A document is usually handed to someone else when it is finished, so the checking happens under a deadline, and the last section is always checked least. Because the ledger holds the quote, you can verify claims in a document six months later without reopening every source.

Check 1: count the claims before you read

Start with arithmetic, not reading. Extract every checkable claim and count them.

A claim carries a number, a price, a date, a percentage, a quote, a named person or a ranking word. A forty page report typically holds between two hundred and six hundred. Long documents fail at the joins, so verify claims in a document at each heading change before reading the prose.

Pass condition: a count exists, and the count is plausible for the length. If a forty page document returns twenty claims, the extraction stopped early, and a run that stops early produces a report that is accurate about a fraction of the document. The appendix is where corrections go to hide, which is one more reason to verify claims in a document from the back.

To verify claims in a document at this size by hand, the count is also a plan. Six hundred claims at thirty seconds each is five hours, which is why the next six checks prioritise rather than sample.

Check 2: find the repeats and give them one source

The second check is the one that saves the most time, because it turns six hundred claims into a much smaller number of facts. When the owner of a report leaves, the next person can still verify claims in a document using the record that was left behind.

Group identical claims. The price in the summary, the section and the table is one fact with three copies. Bind the fact once, and then check that every copy carries the same wording, the same currency and the same period. A limitations page costs twenty minutes and changes how a reader judges everything you did not verify claims in a document about.

Pass condition: every repeated claim traces to one ledger entry, and every copy is identical. A copy that differs by a word is a separate claim and needs its own entry, because a near copy is how a general statement becomes a specific one during revision. Two reviewers should not both verify claims in a document independently unless the document is going to a regulator.

This is also where most document errors hide. Not in the numbers themselves, but in the copies drifting apart, one of them edited at eleven at night while the others were not. The table is where money is decided, so verify claims in a document there before spending an hour on the introduction.

[One fact, four copies, and the drift between them.]

One fact with four copies in a document, showing where they drifted apart
Figure 2: One fact, four copies. The error is in the differences.

Check 3: treat every table as a claim farm

A table is the densest claim surface in any document, and the least reviewed.

Three habits make tables checkable.

One row, one claim. Bind each row separately, with the row label copied. A row label is the subject of that claim, and dropping it is how a figure gets attached to the wrong product in a comparison table. If you only have ten minutes, verify claims in a document for prices, percentages and dates, in that order.

Check the totals by hand. A total is a claim about arithmetic, and a table can be internally consistent and still wrong because a row was edited after the total was computed. A claim written as an estimate with a date is easier to verify claims in a document against than a claim stated as fact.

Check the units column. Millions versus thousands is the most expensive single character in a report, and it is invisible to a reader who is scanning. Print the report before you sign the delivery, because the client will ask how you verify claims in a document of that length.

Pass condition: every row has a source, every total recomputes, and every unit is stated in the table rather than in the paragraph above it. Where a source has been withdrawn, the record still shows when you did verify claims in a document and what it said then.

A table that passes those three is a table you can defend. A table that fails any one of them will be challenged by the first reader who adds up the column. A proposal with one wrong figure costs more than the afternoon it takes to verify claims in a document of forty pages.

Check 4: follow the internal references

Long documents refer to themselves, and internal references rot silently.

Every sentence that says as shown above, as discussed in section three, or see the appendix is a claim about the document, and it needs a check that no external source can supply. Teams that verify claims in a document before every release stop describing corrections as a surprise.

Pass condition: every internal reference points at content that still says what the sentence claims it says. Ten minutes with a printed copy and a pen does this faster than any tool, because the check is navigation rather than comparison. The exit code answers whether to publish, and only the report tells you what to verify claims in a document about next.

The failure mode is subtle and common. Section two is revised to narrow a claim. Page nine still says the broader version, and now the document contradicts itself in a way that a reviewer reading top to bottom will not notice, because both sentences are correct in isolation. Reading the claims list is faster than reading the document, which is the reason to verify claims in a document this way at all.

Check 5: check the document’s own numbers

Reports contain claims about themselves: number of respondents, number of sources reviewed, number of months covered, number of cases.

These are the claims a reader can most easily verify, and the ones most often written from memory, because they feel like metadata rather than content. The summary page makes the largest promises, so verify claims in a document there before reading any argument.

Pass condition: every self-referential figure is derived from a list in the document rather than typed. If the report says forty two interviews, there is a list of forty two, or the number is a claim with nothing behind it. Where a number appears without a year, that absence is the first thing to verify claims in a document about.

This is the same rule that applies to a product page claiming a test count. If the count is not printed by the process, it is a number someone remembered. A reviewer who signs the delivery should also be the person who helped verify claims in a document at least once.

Check 6: quotes and citations, separately

A quote and a citation fail in different ways, so check them separately.

A quote fails when the words are not exact, when the speaker is wrong, or when the quotation is a summary wearing quotation marks. A citation fails when the source exists but does not say the thing, when the page number is wrong, or when the link is dead. Before the run, list the claims you expect to fail, then compare that list with what you needed to verify claims in a document for.

Pass condition: every quote is exact against the source page, and every citation resolves to a page that supports the sentence it is attached to. Two different checks, two different lists. A document of forty pages with three owners needs one ledger, or you will verify claims in a document three times.

In a document with forty citations, this check finds something roughly every time. The most common finding is not a fabricated source, it is a real source attached to a sentence it does not quite support, which is harder to notice and just as damaging. Where a figure comes from a vendor, the vendor page is the source, which makes it the first thing to verify claims in a document against.

[Quotes and citations as two separate lists with different failures.]

Two check lists, one for quotes and one for citations, with separate failure modes
Figure 3: Two lists, because they fail in different ways.

Check 7: the handover proof

The last check exists because a document usually leaves your hands.

Pass condition: someone who was not involved can open the folder, read a claim, read its source and quote, and reach the same verdict without asking you anything. If that is not true, the run produced paperwork. The check takes minutes per claim, and the total is small next to a correction issued at month end. A claim about a competitor is the one readers check first, so treat it as urgent when you verify claims in a document for a launch.

The folder holds the report, the proof file with a hash and a source card per bound claim. That hash confirms nothing in the file changed after the run finished. It does not prove who ran it, and any tool claiming otherwise is overclaiming. A claim that survives the check twice is still checked, because sources move and the second run is cheaper. Where the source is a screenshot, the screenshot is not the source, and the page behind it is what to verify claims in a document against.

For a document, the folder should also hold the version checked, named exactly, because a report has versions and a verdict about version four is not a verdict about version five. Deliver the report with the document, and the reader can see exactly what you did verify claims in a document for.

[A versioned folder as the handover artefact for a long document.]

A versioned evidence folder prepared for handover with a long document
Figure 4: The version name is part of the proof.

What changes at ten thousand words

Four things, and each one changes the process rather than the checks.

Prioritise by surface, not by page. Check every table, every summary paragraph and every number in the conclusion before checking body prose. The surfaces readers act on are the surfaces worth the first hour. Where a claim depends on a date, the date is part of the claim and needs its own entry in the record. The wording of a claim matters: a fair summary needs no repair, and a loose one needs a rewrite before you verify claims in a document.

Check the first and last two pages twice. A reader’s attention concentrates there, and so does a reviewer’s, which is why errors survive in the middle. A short note naming the two claims you could not check is worth more than a long report claiming everything was.

Keep one claim list for the whole document. A single list is what makes the repeats visible. Per section lists hide the drift that check two exists to find.

Freeze the file before the run. To verify claims in a document that is still being edited is to check a moving target, and the resulting verdict is about a version that no longer exists.

The order matters at any length: extract, count, group, then judge. Judging first is the habit that produces a five hour run with no verdicts at the end of it.

The automated version

By hand, seven checks on a forty page report are a day. With a prepared ledger, the checking is minutes and the human time goes into the claims that failed.

node bin/saakshi.js scan report.md
node bin/saakshi.js prove report.md --ledger ledger.json
node bin/saakshi.js pack report.md

Exit 0 publishes, 3 sends a claim to a human, 2 blocks, 1 means the engine could not run. The engine is at SOVEL SAAKSHI, and its limits are published before purchase rather than in a support thread.

Two limits matter more for documents than for pages. Extraction is deterministic over Markdown, so a scanned PDF has to be converted first, and a figure inside an image stays invisible. And the hash proves integrity rather than identity, which matters when a document carries four authors.

[The run, the refusal, and the folder left behind for a long document.]

A document run ending in a refusal with the evidence folder it leaves
Figure 5: One refusal is worth more than a day of reading.

The report reader who is not you

A document is read by people who did not write it, and their first question is always the same: where did this come from.

Two things make that question cheap to answer.

A source card per figure, not per section. A section level citation is an invitation to hunt.

A one page summary of what was checked and what was not. A report that says which five figures could not be verified is more trustworthy than one that implies everything was, because readers already suspect that nothing is perfect. Naming the gaps buys credibility for the rest.

That is the part of verify claims in a document that has nothing to do with software. It is the choice to be specific about what you did not check, and it is the reason a short limitations paragraph belongs in every long document.

[The claims checked and not checked, as a one page summary.]

A one page summary listing claims checked and claims left unverified
Figure 6: Naming the gaps buys credibility for everything else.

Working with a document you inherited

Most long documents are not written by one person, and the one you are checking now was probably written by someone who has left.

Three adjustments make an inherited document checkable.

Do not trust the reference list. It is a list of what the previous author read, not a list of what supports the sentences. Check the citations against the claims, not against the bibliography.

Assume the numbers were correct once. An inherited figure is usually the right figure for a different year, a different edition or a different sample. Look for the version rather than the value, because the value is what makes the sentence look fine.

Read the limitations section first. If there was one, it tells you which claims the author already knew were weak, and those are where the first hour goes. If there was none, that absence is itself a finding worth reporting.

The awkward part of an inherited document is that corrections may contradict a claim the client has already repeated elsewhere. When that happens, the honest move is a short note that names the change and the date, not a silent edit. A silent edit is how the same wrong figure survives into the next document, which is exactly the pattern this page exists to break.

The one page that makes the rest defensible

Every long document benefits from a single page at the front, listing what was checked and what was not, with the date of the check.

It costs twenty minutes and it changes how the document is read. A reader who sees a limitations page stops hunting for what you hid, and starts trusting the parts you claimed. That is a better outcome than a document that implies perfection, because no reader believes perfection and every reader suspects its absence.

The page also protects you over time. Two years later, when a figure turns out to be stale, the page shows that it was correct on the date of checking, which is a different and much better conversation than the one about a number nobody checked at all.

The audit described below was carried out on 21 November 2026 on a forty one citation report.

SOVEL SAAKSHI keeps the ledger and writes the folder, and the document work above is what it was built for.

To verify claims in a document you need the quote, not the link, and any audit that tries to verify claims in a document without quotes is an audit of links.

Where sources are numbered, a citation verification tool checks the number as well as the quote, which is the failure mode a spell check never sees.

Questions about how to verify claims in a document

Does this work for a proposal as well as a report?

Yes, and proposals gain more from it. A proposal carries prices, dates, scope figures and claims about past work, and every one of those is a commitment the client can hold you to later.

What about a document written by four authors?

Then the folder matters more, because the verdict has to name who decided. An offline hash proves the file was not altered, and a register names who ran the check. Between the two you have traceability without pretending to have identity.

Can it read a PDF or a scanned document?

Deterministic extraction works on Markdown text. A PDF has to be converted to text first, and a scanned page needs OCR, which introduces its own errors. A figure inside an image is not visible at all, so it has to be transcribed by hand.

How do I handle a document with tables in images?

Transcribe the table into text before checking, and keep the image name in the ledger note. A table that exists only as a picture is uncheckable by anyone, including the next author.

What if the same fact appears twelve times?

That is the normal case in a long document. Bind it once and check that all twelve copies read identically, which is faster than twelve separate checks and catches the drift that twelve separate checks would miss.

Does this replace a proofreader or an editor?

No. It replaces the mechanical part of reviewing a long document and removes the excuse for skipping it. An editor reads for argument and order. These checks read for evidence, and a document needs both.

What does it cost?

Pricing is ₹2,499 Personal, ₹6,999 Pro and ₹14,999 Agency, one payment, with GST added. Perpetual for the v0.x line, offline, no account and no activation. For a document practice, Agency is the edition that covers client deliverables.

Where do I start with a document I have already delivered?

Take the last report you sent. Extract the claims, count them, and check the five largest numbers in it. That hour tells you whether the document practice exists or whether it was a hope, and it is more useful than reading this page twice.

For the page level routine, verify AI content before publishing is the shorter version. For the editorial method behind both, how to fact check AI generated content covers seven checks in order. And if the question is why any of this is now expected rather than optional, the answer is dated: Google’s guidance on generative AI content asks for a manual review before publishing, and the spam policies describe what happens to sites that publish unsupported claims at scale.

One sentence to end on. A long document cannot verify itself, and the seven checks above exist because the failure that matters is never in the section you were paying attention to when you wrote it.

Connected Systems & Architecture

This publication connects directly to the formal SOVEL software registry, production engines, and architectural documentation: