abigailhaddad/eeoc-appellate-decisions

EEOC federal-sector appellate decisions

Every file listed on EEOC's appellate decisions page, https://www.eeoc.gov/federal-sector/appellate-decisions, downloaded as-is.

A GitHub Actions job adds newly listed files daily (code).

What's here

listing.parquet (same rows as listing.csv; ID columns stored as text, so leading zeros survive) is the index: one row per listed link. It started as a full crawl on 2026-11-01, which matched the site's own total that day; the daily job appends rows for new URLs. page and position are where the link was when first seen. EEOC's listing order shifts, so they aren't stable IDs; use url.

column	meaning
url	link as listed
file_name, folder, ext	parsed from the URL
posted_folder_date	date in a decisions/YYYY_MM_DD/ folder name. This is when EEOC posted the file, not when the decision was issued.
numbers	appeal/request numbers found in the file name
numbers_loose, numbers_loose_rule	numbers recovered from irregular names (07-2014-0029-0500.pdf, closed0520150131.r.txt), filled only when numbers is empty
first_seen	date the URL was first listed in a crawl (2026-11-01 for the original crawl)

data/files-*.parquet holds one row per download attempt. The columns are url, file_name, page, position, http_status, content_type, n_bytes, sha256, fetched_at, error, and content (raw bytes of the .txt, .pdf or .docx; null if the fetch failed). A URL that failed and was retried can show up more than once. Use the row with http_status 200.

state/uploaded.jsonl has the same rows without content, plus the shard number: a quick way to find a file without opening the shards. state/runs.jsonl has one line per daily run: window, new rows, and whether the index matched the site's total.

extracted/decisions.parquet has fields read from each file posted since 2025 by claude-haiku-5-5: one row per unique file, updated daily, with url, sha256, numbers, posted_month, and these columns:

• Case: decision_date, agency, appealing_party, date_formal_complaint
• Complainant, as stated in the decision: job_title, complainant_race_stated, complainant_sex_stated, disability_condition, representation
• Bases: basis_race, basis_sex, basis_disability, basis_age, basis_reprisal
• Result: discrimination_found, complainant_got_relief, commission_changed_result, aj_found_discrimination, agency_action_on_aj
• Issues: issue_performance_appraisal, issue_non_selection_or_promotion, theme_telework

Only columns that passed testing against an answer key are included. These are model outputs, so check anything you rely on against the decision. The answer key was made by Claude agents, not lawyers. discrimination_found means a finding exists in the case; it may predate this decision, and false includes "not reached". Definitions and checks are in extraction/README.md, and results by column in accuracy.csv.

Things to know before you use it

The number in a file name doesn't reliably identify the decision. Many numbers appear in more than one file. In a sample of those groups, most were the same decision posted more than once: byte-identical copies, txt and PDF versions, or renamed copies. Others were a decision plus its posting order, different decisions in the same case (such as an original decision and its later reconsideration, often under the identical file name), or a file name pointing to the wrong case entirely (05a40059.txt is really Request No. 05A50059). Use the case number printed in the document, and sha256 to find exact copies.

EEOC's posting dates aren't decision dates. It posted almost nothing between December 2024 and January 2026 then posted decisions issued during that gap.

Other quirks of the source, kept as-is:

• Some files have no usable number in the name. That includes truncated names like 01983-2.txt and weekly bundle PDFs that hold many decisions each.
• Some files are empty on EEOC's site (rechecked 2026-11-05). For 05a60527, 05a60499, 05a60248, 01a60482 and 0120093445, a .txt with the same name has the text. These have no copy under the same name: 0520110269.txt, 01972934_burns_-, 0120093457.r, 0120081469, 0120080695, 0120063635, 0420170003.txt, 01976694.txt, 0120161966.txt, 0120160234.txt, 01972830.txt, 0120083883.txt, 0120083780.txt.
• A couple of listing rows link to the bare site root, https://www.eeoc.gov. They're in listing.csv but weren't downloaded.
• Older decisions name complainants in full, as they do on EEOC's site.
• Many 2025-26 PDFs have an image for page 1, apparently because the real name was redacted there, so plain text extraction misses the caption. ocr_used marks them in the extracted table.

Code: https://github.com/abigailhaddad/eeoc-appellate-decisions