Follow-up to the quote-anchor lesson (Gemini grounded search: anchor extracted claims back to the grounded text via verbatim quotes — never paraphrase-then-match): quote-anchoring + figure verification is not sufficient. A production Gemini grounded-search pipeline (gemini-3.5-flash + googleSearch tool, temperature 0) published an impossible statistic ("US auto sales hit a record 24.10 million SAAR" — the series' all-time high is 21.71M; actual was 16.3M) that passed every gate: verbatim source_quote located in grounded_text, groundingSupports overlapping the span, and a figure-presence check on the cited page.
What actually happened, verified from recorded artifacts:
- The model invented the figure, then searched for it.
groundingMetadata.webSearchQueriescontained"vehicle sales" "24.1m" "July" "2026"— the fabricated value appears INSIDE the model's own search query. It decided the number first and fished for confirmation. This is a machine-checkable fabrication signature: any figure in the output that also appears verbatim inside a quoted webSearchQueries term, but in no fetched page, is presumptively fabricated. (Related earlier tell: OR-lists of candidate values in queries, e.g."1.1%" OR "1.2%" OR "1.3%".) - groundingSupports endorsed the fabricated span anyway. The support mapped the fabricated bullet to two chunks: one was an HTTP 202 with an empty body, the other a data-aggregator page (tradingeconomics.com) that does not contain the figure. Supports vouch for spans of GENERATED text; they are not evidence the cited page contains the claim.
- Substring figure verification is trivially satisfiable on data-table pages. The gate checked
'24.1' in page_text.replace(',', '')— after comma-stripping, any unrelated value shaped like "6,724.15" contains "24.1". A financial data table with hundreds of numbers defeats substring checks by construction. Fix: match figures as standalone numbers ((?<![\d.,])VAL(?![\d.])) AND require a headline/series-label token within a ~300-char window of the match. - A staleness gate made fabrication the default. The pipeline only re-extracted the claim from the cited page when the page was fresh; on a stale page it fell back to the model's grounded prose — so the safety mechanism (page re-extraction, which would have surfaced the real 16.3) was skipped exactly when the fabricated text shipped. Invariant to adopt: model-generated figures NEVER ship unless (a) re-extracted from a fetched page, or (b) verified present-and-label-adjacent in the fetched page. Verification must be the gate, not an optional upgrade.
Provider swap does not fix the class: Perplexity Sonar measured 37% citation-error rate (CJR 2025-03, best of eight platforms tested) with the same "real URL, fabricated claim" shape. Raw-retrieval APIs (Tavily/Exa/Brave) avoid it only because they generate no prose. For official statistics (payrolls, SAAR vehicle sales, rates), a free deterministic oracle exists: cross-check extracted figures against FRED series with a tolerance band.