Source & summary comparison

A model asked to summarise rarely invents a whole paragraph. It rounds 18 percent to 20, drops the qualifier that made a sentence true, merges two findings into one claim, or attaches a quotation to the wrong speaker. Each of those is a small edit that survives a fluent read, because the summary reads better than the source it came from.

Two views catch it. The diff shows what moved between the two documents, sentence by sentence. The trace below it ignores phrasing entirely and asks one question of every figure, quotation and link: does this string appear anywhere in the source? What it cannot find, it lists.

It catches invented numbers, rewritten quotations presented as verbatim, and links the source never contained — the failures that are cheapest to prove and most damaging to publish. It is deliberately mechanical, because a mechanical check cannot be talked out of a finding.

It misses everything subtler. A number can be present in the source and still be attached to the wrong year. A paragraph can quote correctly and draw the opposite conclusion. A summary can omit the sentence that made the finding conditional. Treat a clean trace as permission to read the diff closely, not as an accuracy certificate.

Paste the source first, then the summary, and work the flags before reading for style. Each item you resolve is either a correction to the summary or a line in the record explaining why the phrasing is legitimate. Keep both documents next to the final text: the comparison is the evidence that the summary was checked against something.