The failure mode we are designing against
When a language model is asked to find evidence in a document, the most common failure is a citation that looks plausible and is partially or entirely fabricated. The quoted text approximates something the document says, but does not match the actual language. In a casual context this is acceptable. In a patent claim chart it is a credibility problem with consequences.
The cost of that failure is high enough that we treat verbatim extraction as a hard constraint rather than a quality goal. A quote that cannot be verified against the source does not appear in the chart, even if the underlying analysis is correct.
How extraction actually runs
The pipeline ingests each piece of product evidence and extracts the full text. PDFs are parsed as text where possible. Scanned PDFs and pages without selectable text fall back to OCR. Web pages and YouTube transcripts are pulled and normalized. The result is a corpus of source text indexed by document and page.
When the system is asked to find evidence for a claim limitation, it does not summarize the corpus. It surfaces candidate passages and returns them as direct quotes with their document and page attribution. The quote is meant to be the exact text the reader can find in the source.
What the reader sees
The chart row contains the quote in the same words as the source. The reader can search the original document for the quoted text and will find it. The audit trail covers every piece of evidence on every limitation, not a representative sample.
This makes the chart usable as a starting artifact for downstream work. Outside counsel can verify any cell directly. Opposing counsel cannot puncture a citation that exists, character for character, in the production. A reviewer at a target company can check the chart against the documents they themselves published.