# Permit Document Enrichment — Release 2 QA report

Published: 2026-07-29
Index snapshot: 2026-07-27

## Scope and QA tiers

- Official Virginia DEQ permit documents represented: 199
- Fully second-reviewed Release 1 documents: 35
- Additional machine-extracted documents manually second-reviewed: 5
- Machine-extracted documents pending a second review: 159
- Equipment-group rows: 873
- Operating-scope rows: 405
- Emission-limit rows: 4176
- Direct official document URL and one-based PDF page on every content row: 100%

Release 2 is a complete-corpus, tiered-QA release. It does not represent all
199 documents as manually verified. Document-level `qa_status`,
row-level extraction notes, confidence and manual corrections remain public.

## Coverage

- Permittee printed in document: 179/199
- Facility address: 168/199
- At least one stated equipment count: 189/199
- At least one operating or emission limit: 192/199
- Project-calculated combined capacity totals: 0
- Published rows derived from actual operating or emissions values: 0

## Manual second-review sample

Five machine-extracted documents were compared again with the official PDF:
53197-2, 11541-3, 21527-15, 74242-2 and 74130-2. Two required manual
corrections. Permit 11541-3 had OCR-collapsed equipment and emission tables.
Permit 74242-2 had a merged header association error and a subsequent
measured-VOC-dependent formula that was excluded. The corrected rows and
manual review log are included in this release.

## Permit-document formats observed

1. Modern text PDFs with ruled equipment and emissions tables.
2. Modern text PDFs with merged or multi-row table headers.
3. Older text PDFs whose tables collapse into reading-order lines.
4. OCR-heavy historical permits requiring visual transcription.
5. Amended or superseding permits with history language on authorization pages.

## Scaling recommendation

- Scale with automated extraction and page provenance: direct document URL,
  document date, regional office, permittee candidate, address candidate,
  equipment count candidate, and literal numeric limit candidates.
- Require manual review: merged-header equipment associations, control state,
  per-unit versus combined scope, operating-hours basis, OCR-heavy tables, and
  superseding-document history.
- Do not publish as derived fields: calculated combined capacity or emissions,
  beneficial owner inferred from facility name, construction or operating
  status, actual hours or emissions, compliance, violation, exposure, harm, or
  causation.

## Publication boundary

Every value must be read with its `qa_status`, extraction note and cited page.
Machine-extracted pending rows are leads for verification, not a substitute for
the official document. Blank means absent, ambiguous, conflicting or not
reliably extractable; it does not mean zero.