Su Tung.
Sign in

← Blog

21 August 2026

Quotes are copied, not written

Checking that an authority was verified says nothing about the words inside the quotation marks. Here is the check that reads them, and the defect that bought it.

Every citation check we had until recently was authority-level. It asked one question: does the authority named beside this quotation carry a verification row?

Once Cap. 132 had one, every mention of Cap. 132 in that answer passed — whatever words sat in the quotation marks next to it.

What that misses

From one of our own evaluation runs, scored clean by the citation audit, the reference module and the quality gate alike.

The verified passage in the record was:

"The Authority may specify the purposes for which any civic centre, or any part thereof, may be used." — Cap. 132, s. 105N

The delivered answer said:

Cap. 132, s. 105H(3): "The Authority may specify the purposes for which any museum, or any part thereof, may be used."

A different section number, one different word, and the whole conclusion of the answer — that this was the operative power to hire out museum space — rested on that sentence.

Section 105H(3) may well say exactly that. The words did appear somewhere the run had read. That is the point: it was an unchecked quotation, not necessarily a false one, and nothing in the system had looked. Had the wording been carried across unchanged from the civic-centre provision, every check would still have passed it.

A lawyer reading that answer sees quotation marks and a section number. There is no way to tell the two cases apart without opening the chapter.

The rule now

A quoted passage in an answer may be a shorter part of a verified quote. It may not be longer than one.

That asymmetry is the whole rule. Showing less than was checked is still verbatim. Showing more means words nobody checked are inside the quotation marks — and once the answer is written, nothing downstream re-reads the source to find out which words those were.

The failure that bought it looked like this:

"…may fix fees for the use of any facility, and may do so retrospectively"

The bold half was never in the statute. It passed a containment check that accepted any span containing a verified quote, and it shipped inside quotation marks.

If a proposition needs wider language than was verified, the answer paraphrases it without quotation marks, or verifies a wider passage. Quotation marks are reserved for text that was copied.

What it measures

An offline instrument now takes every quoted passage in a delivered answer, attributes it to the citation the answer wrote it against, and asks where those exact words came from: a verified quote, a document the run read but never verified, or nowhere it can find.

Measured on 2026-08-21 over 284 stored answers:

quoted passagesmatched a verified quote
the older prompt-driven path12941 (31.8%)
the current pipeline913864 (94.6%)

The gap is the mechanism, not the model. In the pipeline the answer's only citation vocabulary is placeholders standing for verified passages, so quoting the evidence is the path of least resistance. When the model is merely told to quote accurately, it does so about a third of the time.

The same check now runs as a pass or fail on every evaluation: every quoted passage must be found in the document it cites. It also catches quotations with no citation near them at all — a passage nothing can be resolved against fails for that reason and says so.

What is still yours to do

A verbatim quote is the right words in the right place. It is not the right authority for your point, and a passage quoted accurately can still be quoted out of context — a proposition the court went on to reject two paragraphs later reads perfectly well in isolation.

Read the paragraph around it. That is why the pinpoint is there.

Quotes are copied, not written — Su Tung