REVISION TABLES -- output/revision/tables

Conventions
letter_id: stable source-record identifier; application_id: normalized primary
application; category codes follow notes/phase3_codebook.md. Counts are integers.
share, difference, ci_low and ci_high are fractions except in the explicitly named
primary_categories_with_intervals table, where percent and interval bounds use 0--100.
Date strings are ISO dates. Zero-denominator transition rows are blank in proportion
CSV files and dashes in the manuscript. The original panel dictionary is supplied at
data/final/panel_codebook.csv. No missing submission reference implies original status.

strict_crl_analysis_corrected_secondary.csv
Unit: one of 443 strict CRLs. Original panel columns plus deficiency_categories_original
and review_sequence_id. deficiency_categories applies the three documented secondary additions.
Primary labels are unchanged. Category lists are pipe-delimited.

category_overlap.csv
Unit: category. letters: denominator 443; primary_n: primary assignments; secondary_n:
nonprimary assignments; any_n: primary plus secondary; any_share: any_n/443.

cooccurrence.csv
Unit: category pair. Row category and category-named columns give numbers of letters
with both categories. Diagonal is the any-category count. Corrected secondary coding.

consecutive_pairs.csv
Unit: adjacent linked observed pair within one review sequence (109 pairs). pair_sequence starts at 1;
first_letter_id/second_letter_id and first_date/second_date identify both records;
first/second are primary categories. Order follows date, then stable letter_id.

transition_counts.csv / transition_row_proportions.csv
Unit: originating category. first_category identifies the row; each category-named
column identifies the next observed category. Counts or fractions of originating pairs.

recurrence_sensitivity.csv
Unit: sample/definition/category combination. sample selects all pairs, first pair per
application, or exclusion of same-date pairs. scheme is separate or merged. n is the
originating-pair denominator; repeated counts equal first and second categories;
share=repeated/n. Wilson ci_low/ci_high are reference intervals; all-pair intervals in
the CSV assume independence and are not used for inference in the paper. First-pair
intervals avoid repeated pair contributions. Manufacturing merges facility and CMC only.

two_year_rates.csv
Unit: category grouping in the eligible 250-application cohort. n: applications;
events: another released strict CRL observed within 730 days of the first; share=events/n;
ci_low/ci_high: Wilson reference intervals.

two_year_differences.csv
Unit: comparison. difference: first named group's share minus comparison share;
ci_low/ci_high: Newcombe Wilson-score interval without continuity correction.

facility_by_period_type.csv
Unit: issuance period by NDA/BLA type. n: letters; facility_n: facility-primary letters;
facility_share=facility_n/n. Empty type/period cells are absent, not estimated.

primary_categories_with_intervals.csv
Unit: cohort/category. n: letters in cohort; count: primary assignments; percent:
100*count/n; ci_low/ci_high: Wilson bounds in percentage units. Cohorts overlap.

same_date_records.csv
Unit: strict letter in an application/date group with more than one source record.
The two retained same-day BLA 761303 records concern different original submissions;
they do not form a consecutive pair. The duplicate BLA 761215 record was removed.

existing_human_labels.csv / existing_human_disagreements.csv
Unit: existing validation-packet letter (50 total; nine disagreements). packet_number
maps directly to the Review worksheet in the completed workbook and the original
packet-builder's row order. human_primary_category and model_primary_category are
unaltered saved labels. No new review or adjudication is implied.

Other files
manufacturing_clinical_overlap.csv
Unit: sample/clinical-definition/group. Manufacturing is any facility_inspection or
cmc category, primary or secondary. Broad clinical includes clinical_efficacy,
clinical_safety, trial_design_conduct, and bioequivalence_bioavailability; Core clinical
excludes the last category. Groups are the four mutually exclusive combinations.
letters is the sample denominator; count is the group count; share=count/letters.
Unflagged classifications restrict to classification_status=ok. Other categories may
co-occur. No clinical code does not establish clinical adequacy.

manufacturing_clinical_letter_flags.csv
Unit: strict letter, identified by letter_id and application_id. classification_status
retains the original flag. has_manufacturing, has_broad_clinical, has_core_clinical are
0/1 indicators using corrected primary-plus-secondary categories as defined above.

analysis_summary.json: revision counts and estimands.
validation_report.json: workbook mapping, numerical checks, and provenance limits.

APPLICATION-LEVEL TABLES
output/revision/application-analysis/first_observed_crl_applications.csv:
  One earliest strict CRL per application, ordered by date then letter_id.
  Fields: application_id, letter_id, letter_date, primary_deficiency_category,
  classification_status. N=333; first observed from 2024 onward: N=128.
application_primary_categories.csv:
  category, label; full/recent_n: category counts; full/recent_denominator:
  applications in the cohort; full/recent_share: count divided by denominator.
application_facility_sensitivity.csv:
  cohort: Full or 2024+; sample: All, Unflagged, Exclude OCR, Exclude redaction;
  applications: denominator; facility_n: facility-primary count;
  facility_share: facility_n/applications. Flags are from the first letter.
Rebuilt by code/19_application_summary.py during full offline reproduction.

NOVEMBER 18 ADDITIONS
review_sequence_exclusions.csv: excluded letter IDs and reasons.
review_sequence_id separates BLA 761303 Original 1 and Original 2. pair_sequence
orders valid linked pairs within application; the earliest valid pair is 1.
Two-year event flags are recalculated after exclusions and are blank when
follow-up is shorter than 730 days.

two_year_by_application_type.csv: NDA/BLA-specific facility_events, facility_n,
facility_share, other_events, other_n, other_share, difference, ci_low, ci_high.
Shares and difference/interval bounds use fractions.

application_all_categories.csv: counts of every primary or secondary category
in first letters, with full/recent_n, denominator and share fields.
first_observed_crl_all_codes.csv: one first letter per application with original
primary category and corrected pipe-delimited all_categories.

facility-recurrence-check/pair_review.csv: 28 first-pair facility recurrences;
facility_continuity distinguishes identified common plants, narrative support,
and indeterminate cases. Evidence excerpts and assessment notes accompany
flags for additional facilities and explicit unproven resolution. A zero
confirmation flag means not established by available text, not proved absent.
Review method is LLM-assisted, not independent human validation.
