Why PDF Modification Checks Read the File, Not the Page

This article is a snapshot – content was accurate as of February 2026 (code examples tested against the API as of February 2026). The product evolves actively; specific counts, examples, and detection rules may have changed since publication – see the changelog for the current state.
You can look at a PDF for an hour and learn nothing about whether it was edited. The page renders the same way whether it came straight from the issuer’s system or passed through an editor on the way to you. That is the starting point for how HTPBE? is designed: the question ‘was this file changed after it was made?’ is answered from the file itself, not from how it looks.
This article explains the reasoning behind that choice, what kinds of evidence a file can carry, and what the method cannot tell you.
The rendered page is the wrong place to look
A PDF is a container. What you see on screen is the last step of a process that starts with how the file was built. Two files can look identical to the eye and still differ in how they were produced, by which software, and how many times they were handled afterwards.
Reading only the rendered page, whether by eye or by OCR, discards exactly that information. Reading only the visible metadata fields is not much better: those fields are easy to edit and easy to leave behind by accident. A check that trusts a single layer inherits that layer’s weaknesses.
So HTPBE? reads the file itself, at several levels, and reports the specific findings it turns up.
What a file can carry
HTPBE? reports findings under named checks, grouped by the kind of evidence they rest on. The full public list is on the How It Works page. In outline:
- Dates and timestamps. Dates that disagree with each other, or that could not have happened.
- Digital signatures. Content changed after signing, or a signature stripped afterwards.
- Revision history. Signs that the document was saved again after its creation.
- Generator identity. Signs that the file’s stated origin was altered, or that an editor or online conversion service handled it after creation.
- Document assembly. Pages that come from different sources or tooling inside one document.
- Page content. Edits to what is drawn on the page, such as covered or overlaid values.
- Images. Embedded images that show signs of being edited and re-saved.
- Overall construction. Whether the file looks like an original or like one rebuilt afterwards.
Not every finding carries the same weight. A legitimate document can be signed, converted for archiving, merged, stamped by a print system or optimised for size, and each of those can leave signs of later handling. That is why HTPBE? reports a verdict together with the named findings behind it, so you can read the result in context. There is no numeric score.
Why one file is enough
Comparison-based approaches need two versions: the original and the suspect copy. In practice you rarely have the original. If you did, you would not need a check.
Because the evidence lives inside the file, HTPBE? works from a single upload. It does not ask the sender for anything and does not need a reference copy. The longer version of this argument is in PDF forensics without the original file.
Three outcomes, not two
A result is Intact, Modified or Inconclusive.
Inconclusive is not a polite way of saying ‘probably fine’. Some documents cannot show an untouched origin at all. A file produced by widely available software, a scan or a screenshot looks the same whether it is genuine or recreated, so the check refuses to call it intact. Reporting inconclusive in those cases is the honest answer, and it is better than a green tick that a forger could earn with the same free tool. The FAQ explains why an official document can show inconclusive.
What the method cannot do
A modification check reads how a file was made, not whether what it says is true.
- A document fabricated from scratch is not a modified document. If the file was created fresh with false content, it may carry no sign of later editing, and the check cannot call it fake.
- Scans and screenshots can never be called intact. They can still be reported as modified when the file shows signs of later editing, but a clean-looking one stays inconclusive.
- The check does not read the content for sense. Whether the amounts, names and dates printed on the page match reality is a different question.
- A result is evidence about the file, not a ruling on intent. A modified verdict means the file was changed after it was made; it does not say why.
For anything high-stakes, treat the result as one input. Confirm the document with the party that issued it, check any digital signature, and keep control over how the file reached you.
Two statements that look the same
A lender receives two bank statements for the same month, from two applicants. Both show the bank’s logo, matching fonts and a balance that adds up. Nothing on either page gives a reviewer a reason to stop.
- Statement A was downloaded from the bank’s portal and forwarded unchanged. The result is intact: no sign of later handling.
- Statement B was downloaded, opened in an editor, saved again and sent. The page looks the same, but the file carries evidence of editing after the bank produced it. The result is modified, with the findings that support it named in the response.
The reviewer looking at the two pages sees one document twice. The reviewer reading the two files sees two different histories. That gap is the whole argument for checking the file instead of the page.
What to ask of any file-level check
If you are comparing ways to check PDFs, these questions separate a method that reads the file from one that reads the page with extra steps:
- Does it work from one file, with no original and no cooperation from the sender?
- Does it return named findings you can read, or a single number you must trust?
- Does it have an honest third outcome for files that cannot show an untouched origin, or does it call everything green or red?
- Does it say what it cannot do, such as judging whether the content is true?
- Does the same file get the same answer every time?
A tool that answers yes to all five is making a narrower claim than ‘fraud detection’, and a narrower claim is easier to rely on. The wider picture of how structural and content approaches divide the work is in structural vs content-based document forgery detection, and the reasoning for running a tool against your own known files is in build vs buy for PDF tamper detection.
Where to go next
If you want to see the kinds of findings in practice, the anatomy of a modified PDF walks through what a result looks like. Two related posts go deeper on specific areas: PDF xref table forensics and common PDF editing tools and the traces they leave.
To run the same check from your own systems, the PDF integrity check API returns the verdict and the named findings in a single REST call. Pricing lists the plans and the pay-per-check credits.