LOCAL REFRESH -- November 18, 2026
Stable archive filename retained; replace the earlier public file before distribution.

- Exclude duplicate BLA 761215 record and NDA 204017 incomplete-response acknowledgement.
- Separate the BLA 761303 original-submission review sequences.
- Corrected sample: 443 letters, 333 applications, 109 pairs across 81 applications.
- Two-year rate: 24/84 facility versus 24/166 other; difference 14.1 points.
- Add NDA/BLA two-year comparisons and all-category first-letter distributions.
- Add LLM-assisted source-text review of 28 facility recurrences and evidence trail.
- Update manuscript wording, tables, and figures. Stored primary labels unchanged.

HISTORICAL RELEASE NOTES BELOW (counts refer to the earlier release)

RELEASE v20260807_r20261103 -- November 3, 2026

The data vintage, stored primary classifications, and headline estimates are unchanged.
The previous release remains identifiable as v20260807_r20261102.

REPRODUCTION
- Rebuilt CSV, TeX, and JSON text is compared after CRLF/LF normalization.
  All other content differences remain failures. Archive integrity and frozen-input
  hashes still require exact bytes. Both summary JSON files are now compared too.
- Regression tests cover line endings, substantive changes, missing files, and
  integrity failures. Run python code/test_reproduction.py.

PAPER AND DOCUMENTATION
- The strict-CRL sample begins May 15, 2009. The first period is 2009-2014 (N=28).
- Drugs@FDA matching uses 322/445 study letters (72.4%) and 44/158 recent letters
  (27.8%), rather than denominators from the broader source-record collection.
- Any uncompleted-inspection mention occurs in 32/260 facility letters (12.3%),
  including five recent letters. The mutually exclusive 5b classification covers
  20 letters, including two recent ones; adverse findings take precedence in 12
  mixed letters. Both measures are now distinguished explicitly.
- Armstrong NDA205920 (2014-05-22) is trial-design/conduct primary, with facility
  inspection secondary. Its quoted inspection finding remains a 5a example.
- MediWound BLA761192 (2021-06-25) replaces Kashiv as the 5b example. Kashiv's
  2021-08-02 letter has both uncompleted and adverse inspection findings and is 5a.
- The RedHill quote restores 'proposed'; its discussion acknowledges the additional
  facility finding instead of treating the letter as an exclusively CMC example.
- All five same-problem CMC-to-facility pairs already cite facility inspection.
- The 30-letter extraction review has year groups 3/14/13, includes four excluded
  predecessor actions, and is distinct from the 50-letter classification-validation
  sample, which contains no predecessor actions.
- Table A4 reports row percentages to one decimal place, preserving 12.5 and 37.5.
- The standard validation stage now checks these facts against frozen records,
  classifications and adjudications, including the exact words of the worked quotes.

RELEASE VERIFICATION
- Reproduced the prior archive's failure on Ubuntu Linux: seven generated TeX
  tables and two summary JSON files differed only in CRLF/LF line endings.
- The corrected full data reconstruction passes on Windows (Python 3.12.10) and
  Ubuntu Linux (Python 3.12.3): all 38 compared outputs match. Both manuscript PDFs
  were compiled and visually checked on Windows. A native macOS run was not performed.
