Tracing a summary's quotations back to the page they came from

Attribution is a different job from reading

Reviewing a summarised document looks like a reading task and is actually an attribution task. The question is not whether the summary is well written; it is whether each sentence it contains is supported by the document it claims to summarise, in the place it claims to be.

In two separate windows that becomes tab-switching, and tab-switching is exactly what lets a wrong page reference survive a review. The reader checks that the sentence sounds plausible against a document they half remember, which is a test of fluency rather than of evidence.

A side-by-side viewer removes the switching cost. The summary sits next to the page it cites, and each quoted passage carries the page number where it was found. Clicking a quote moves the document rather than asking you to scroll and hope.

The result is a review that is faster and stricter at the same time, which is the only kind worth adopting.

How a quote gets matched to a page

Both the quotation and each page's extracted text are normalised before comparison: lowercased, stripped of punctuation noise, collapsed to single spaces. That way a passage split across two rendered lines still matches, and a curly apostrophe in the PDF does not defeat a straight one in the summary.

Pages are tested in order and the first match is reported. The order matters because a repeated phrase will be attributed to its earliest appearance, which is usually correct for a summary and occasionally wrong for a document that restates its thesis in every chapter.

Quotes shortened with an ellipsis fall back to their opening words, so a truncated citation can still be located. Quotes that appear nowhere are reported as not found rather than mapped to the nearest plausible page, because a guess is worse than a gap in this particular job.

Knowing the mechanism tells you what to do when it fails. A miss usually means the summary reworded slightly, or the PDF's text layer is missing, or the passage spans a page break. Each of those is a different repair and only one of them is the summary's fault.

• Quote found on a page: check that the page supports the sentence, not just the words.
• Quote not found: rewording, an empty text layer, or a genuinely invented quotation.
• Scanned PDF: no text layer, so nothing can be located until it is OCR'd.
• Passage across a page break: quote the words from a single page, or verify by eye.

The failure patterns worth naming

The paraphrase dressed as a quotation is the most consequential. Quotation marks are a claim about wording, and a summary that puts its own compression inside them has attributed invented words to a real source.

The right words on the wrong page is the second. It reads as a minor citation error and it is often a sign that the summary was assembled from a different document that shares vocabulary with this one.

The confidently absent claim is the third: a sentence with no quotation and no page reference that the reviewer simply has to check by reading. Those are the majority of a summary's sentences, and no trace will find them. A review that only checks the quoted parts has checked the easy half.

And the orphan figure: a number in the summary that appears in the document but attached to a different period, cohort or unit. Matching confirms the string exists; only reading confirms it means what the summary says.

A pass that takes twenty minutes, not two hours

Load the PDF first and give it a moment to parse, then paste the summary. Read the summary once without checking anything, purely to understand what it claims to be about. A reviewer who starts checking line one before understanding the whole will miss the structural misrepresentation.

Then work the quotes in order. For each one, jump to its page, read the surrounding paragraph and decide whether the sentence in the summary is a fair use of it. Ruling on a quote takes about fifteen seconds once the page is in front of you, which is the entire argument for the layout.

Then read for the claims with no quotes, which is where the real work is. Those are the sentences the summary asserts on its own authority, and they are where the interest lies: a compressed version of a nuanced finding, a conclusion the document explicitly declined to draw.

Finish by recording what you checked and what you did not. A summary review without a scope note reads as blanket approval, and blanket approval is the thing you will wish you had not given.

Choosing which documents deserve this

Not every PDF is worth the pass. The ones that repay it are the ones somebody will rely on without reading the original: a due diligence summary, a research brief, a policy digest, a board pack where the appendix is longer than the deck.

A useful filter is the cost of being wrong. If a wrong figure in the summary would cost money, a decision or a reputation, compare it. If it would cost a sentence, read the summary and move on.

Documents with a good text layer are fast to work through; scanned documents are slow and produce weak findings, so recognise the difference in the first minute rather than the twentieth. If the pages come up blank in the trace, the comparison is not the tool for this file.

And documents that are about to become contractual deserve their own category. Anything a client will sign, quote or attach to a proposal is worth tracing, because the summary has more reach than the original and the original is the thing that will be forgotten.

Recording scope in the handoff

A summary review without a scope note reads as full approval, and full approval over one hundred and forty pages is rarely what was given. Write down what you compared, what you skipped, and why.

Two lines are usually enough. Quotations traced on pages one to forty; figures checked against the appendix table; narrative after section six read but not traced. That is a defensible position, and it protects both the reviewer and the client.

Record the checks that failed and how they were resolved, not only the ones that passed. A finding that was corrected is evidence the process works; a finding that was waived is a decision somebody may need to defend later.

Attach the trace count. Eleven quotations traced to nine pages is a sentence that communicates effort without exaggeration, and it is the sort of detail that makes a review auditable rather than assertable.

Summaries of summaries

The most dangerous document in a review queue is a summary of a summary. Each layer compresses, each compression drops a condition, and by the third layer the caveat that made the finding conditional has been gone for two rounds.

Where the intermediate layers exist, compare the final summary against the original rather than against the middle one. The middle document is the place the distortion entered, and it will confirm its own version.

Where they do not exist, treat the summary as an unsourced claim about the source. That is a stricter standard than most briefs ask for, and it is the correct one whenever a decision will be made on the summary alone.

When the summary is the only thing anyone reads

The structural argument for this check is that summaries travel further than their sources. Board packs, briefings and forwarded emails carry the compressed version, and the reader deciding something has no access to the original and no reason to doubt a fluent paragraph.

That asymmetry is why the trace matters more than the diff for most organisations. The diff is for the editor; the trace is for the person who will never open the PDF and needs to know that somebody did.

What the viewer cannot see

It cannot judge whether the source is any good. A summary can be perfectly faithful to a poor document, and the fidelity is not a virtue if the underlying work is wrong.

It cannot read images, charts or tables inside the PDF. If a figure lives in a chart rather than in the text layer, the trace will not find it and the summary's number cannot be confirmed here.

It cannot detect omission. The single most common failure of AI summarisation is leaving out the sentence that made the finding conditional, and no comparison of present strings will find an absent one. Only reading the source will.

So the trace is a filter for the checkable failures, which leaves the reviewer's attention for the ones that need judgement. That is the same division of labour as everywhere else in the toolbox, and it is the only one that scales.