Document Extraction Runs

Produced 31 July 2026

Verifiable by you

Where this came from

Six commercial lease documents filed with the SEC by REGENXBIO Inc., including a complete lease and a four-amendment chain

A batch where four of the six documents genuinely do not contain most of the field list. That is the case an extraction run is most likely to fail dishonestly, by inferring a value from a neighbouring document rather than reporting silence.

Twelve of twelve fields correct, every cite verified, and every blank audited.

The first run scored 66.7 percent against a 95 percent threshold. Every gain after that came from a defect an audit caught rather than from tuning: a character class that could not cross a hyphen, so money inside a parenthetical was unreachable; a non-greedy capture whose shortest legal match was the empty string; a label that recurs inside an estoppel-certificate template, which produced a populated, confidently-cited, completely wrong value; a table of contents mistaken for section headings; and two rounds of blanks sitting over stated dollar amounts. The scorer itself was replaced after it passed a wrong value on a loose substring match.

Figures from the run

Documents in the batch
6
Characters read
836,000
Fields named by the buyer
12
Cells produced
72
Field agreement against hand-keyed ground truth
100%
Cites verified true, and false cites
12 true, 0 false
Blanks audited, and false blanks found
22 audited, 0 false

Checks

  • met Every populated field carries a page and section cite to the source document All 12 cites were verified mechanically: the value provably occurs inside the section named. A cite nobody checks is decoration.
  • met Field-level agreement with the labelled sample is at or above 95 percent 12 of 12 against ground truth verified by hand. Measured against one document rather than the ten-document sample a real order supplies.
  • met Every named field appears for every document, blank where the source is silent 22 blanks audited with probes phrased differently from the extractor, so the audit does not share its blind spots. A blank sitting over a stated value is the one failure a buyer cannot detect, because there is nothing in the file to check.
  • met Valid UTF-8 CSV with the field list as the header row and one row per document 6 rows, 13 columns, every field present on every row.

Limits

Bounds

Ground truth is one document, not the ten-document labelled sample the offer specifies.

Self-assessed by the party who would sell the work.

It shows the artifact can be made to the published standard. It does not show throughput at volume.

The unit · Order it