Lacuna Index — We read between the lines. Forensic analysis · Pure evidence · Accountable insight.
◂ JOURNAL
·LacunaIndex Team·5 min read

The Quote That Isn't There: How a Deterministic Audit and Source Snapshots Test Whether a Cited Sentence Actually Exists

A quotation mark promises that these exact words appear in that exact document, and AI-assisted research can break the promise invisibly; here is how LacunaIndex checks it, and where the checking stops.

Every diligence memo has a quote in it that someone remembers differently from the source. A chief executive "committed" to a margin target on the third-quarter call; the memo says so in quotation marks; and three weeks later a portfolio manager asks where, exactly, that sentence appears. Nobody can find it. Either the wording was paraphrased, the quote was attached to the wrong document, or the document changed after it was read.

LacunaIndex treats a quotation as a claim that has to pass checks, not as decoration. This post walks through what those checks are, what they catch, and, just as important, what they do not yet prove.

Why quotes deserve their own audit

Research produced with the help of artificial intelligence has a particular failure pattern. The prose is fluent, the attribution looks plausible, and the quotation marks give the sentence an air of authority. But a quotation mark is a promise: these exact words appear in that exact document. A language model can break that promise without any visible sign.

The LacunaIndex report payload carries verbatim quotes in several places: the evidence behind each execution dimension, the evidence behind each named capital allocation investment, the best quote and its source for each member of the executive leadership team, and the older leadership executive rows. The quote auditor walks every one of those fields and runs the same series of checks on each.

Five structural checks

The auditor is a deterministic function. It makes no call to an AI model and no network request, and it is recomputed each time a report is rendered, so a rerun immediately reflects the latest quote set. Each quote is tested for:

  1. Missing attribution. A quote with no source at all.
  2. Placeholder text. Filler such as "N/A", "Unknown", "TBD", "Not disclosed" or a bare dash sitting where a quotation should be.
  3. Suspiciously short text. A "quote" under 12 characters is flagged, because fragments that short cannot anchor a claim.
  4. Unlisted source. The attribution has to match a document the report actually lists as consulted, either by web address or by overlapping words in the title. A quote credited to a document the report never read is a warning sign.
  5. Duplicate verbatim text. The same sentence appearing under two different sources. Either one attribution is wrong or the sentence was recycled.

From "looks right" to "is in the document"

Structural checks catch sloppy evidence. They cannot catch a confident, well-formed, fabricated quote. For that, LacunaIndex adds a second layer: source snapshots.

When a source web address is fetched, the platform stores a cleaned copy of the text, together with a fingerprint of the content (a SHA-256 hash, a standard way of detecting any change to a file), the fetch time and the fetch status. The web address is normalised first, so trivially different spellings of the same address collapse to one record.

If a snapshot exists for a quote's source, the auditor checks that the quote text actually appears inside the stored copy. The comparison is forgiving about typography only: it ignores capitalisation, collapses runs of whitespace, and treats curly and straight quotation marks, and different dash characters, as equal. It is not forgiving about words. It also refuses to "verify" anything shorter than eight characters, since a tiny string will appear in almost any document by chance.

The result for each quote is one of three states: verified (found in the snapshot), not in snapshot (a snapshot exists and the words are not in it), or not yet checkable (no snapshot exists for that source).

What counts as a failure, and what is only a warning

The auditor separates problems by severity:

  • Failing: no attribution, placeholder text, or a quote that is absent from its own source snapshot.
  • Weak: too short, source not on the consulted list, or duplicate text. These deserve a look but do not by themselves prove a problem.
  • Clean: passes every check.

The summary reports the total number of quotes, how many are clean, weak and failing, and how many were proven against a snapshot. For an analyst, that last number is the one to read first. A report with twelve clean quotes and zero verified ones has passed formatting checks, not truth checks.

How this connects to source quality

Verification tells you a sentence exists in a document. It does not tell you how much that document should count. LacunaIndex also tags evidence by source type and applies fixed weights: filings with the Securities and Exchange Commission and audited financials carry full weight (1.00), investor days 0.85, earnings calls 0.70, press releases 0.55, media interviews 0.35 and marketing material 0.20. The weights are written in code, not left to a prompt, so the same source always counts the same way.

Put together, a quote is judged on two independent questions: is it really in the document? and how much is that kind of document worth?

What the auditor does not do

Honesty about limits matters more here than anywhere. The code comments describe the current auditor as a first version of immutable evidence. Three limits follow:

  • A quote with no snapshot cannot be verified, only structurally checked. "Not yet checkable" is not a pass.
  • Matching is on text, not meaning. A real sentence quoted out of context will verify.
  • A verified quote proves the words exist in the stored copy of the page. It says nothing about whether the page itself is reliable, which is what source weighting is for.

A practical workflow for the analyst

  1. Open the quote audit for the company and read the verified count before the clean count.
  2. Treat every failing row as unusable until you have opened the source yourself.
  3. For weak rows, check the attribution. Duplicate text under two sources usually means one of them is wrong.
  4. For quotes that matter to your thesis and are not yet checkable, pull the source document and search for the wording by hand before it goes into a memo.
  5. Weigh what survives by source type: a verified line from a filing is stronger than a verified line from a press release.

The goal is modest and useful: when someone asks where a sentence came from, the answer is already on the page.

Filed under: quote-audit · source-snapshots · source-hierarchy · evidence-verification · methodology · diligence-workflow