Quote Attribution Verifier

LLMs are notorious for inventing plausible-sounding quotes, and human editors miss the subtle ones (single word changed, source slightly misremembered). Verifying after drafting is cheaper than preventing at generation time and scales to long documents. The tight verbatim-first + paraphrase-fallback logic matches how fact-checkers actually work: exact match is strong evidence, semantic match is a weaker signal that still warrants a flag.

1) Regex-extract every sentence between double quotes or within blockquote markdown. 2) For each quote, run a verbatim substring search across every source in the corpus — if found, mark verified. 3) For the remainder, compute embedding similarity against every source sentence; top-1 above threshold becomes a 'paraphrase candidate' (still flagged). 4) For the remainder (no verbatim, no embedding match), mark 'unsupported'. 5) Output a report with line-by-line status and suggested fixes: the nearest real quote from the source, or a recommendation to remove the quote. 6) Optional: rewrite unsupported quotes using an LLM, constrained to only use phrases from the source.