Charity Impact Media: where the money goes

Live: https://theguysaccount.github.io/charity-impact-media/ · Data (CC0): latest release · Screens: desktop and phone

Sourced spending breakdowns of every US 501(c)(3) public charity with $10 million or more in revenue (about 25,000), built from each charity's own IRS Form 990 filings and refreshed as new filings post:

• Where $100 of spending went: programs, management and general, fundraising (Part IX).
• Where $100 of revenue came from: contributions, government grants, program service revenue, investment income (Part VIII).
• The 990 scorecard: the same twelve measures for every charity, each next to the median for its cause. No grades.
• Highest reported compensation, by role, with pay from related organizations and benefits (Part VII). People are never named.
• Programs in the charity's own words, governance answers, the largest outside contractors, up to seven years of trends, similar charities, and private-foundation funders via WHOFUNDS.

Every figure shows its filing year and links to the filing it came from. Every page has a public correction link and a right-of-reply for the charity. There are no ratings, ads, accounts or tracking cookies.

How it runs (no human in the loop)

What	Where	When
Fetch new filings, rebuild, test, guard, publish a data release	.github/workflows/refresh.yml	weekly; rebuilds only when an index, the BMF or the data code changed
Build the site from the latest release and deploy to GitHub Pages; IndexNow	.github/workflows/deploy.yml	after each refresh, on site changes, when new reply issues appear
Live site check, reply sync; weekly live source self-test	.github/workflows/health.yml	daily; self-test Sundays
Tests	.github/workflows/test.yml	every push

Each job fails safe: nothing is published unless every step passes, and the live site keeps the last good build. Scheduled workflows keep themselves enabled (scripts/keepalive.sh). Details, alerting and the optional keys (custom domain, support link, analytics, Search Console): OPERATIONS.md.

Layout

pipeline/                 Python 3.9+, standard library only (Pillow optional, for share cards)
  sources.py              fingerprint the public inputs, download them, guard a rebuild
  select_filings.py       covered charities (IRS BMF + data-lake indexes) and every Form 990 they filed since 2019
  fetch_filings.py        fetch + parse into the cache: data lake over S3, IRS-only filings by HTTP range request
  irszip.py, inflate64.py read single members of the IRS monthly zips (incl. Deflate64) without downloading them
  parse990.py             Form 990 XML -> compact record; never reads personal names
  build_data.py           pick each year's return (amended wins), compute measures, medians, ranks
  render.py, pages.py     HTML templates (charity pages; home, causes, states, A-Z, guides, method, corrections)
  build_site.py           write dist/: pages, data.json, search index, sitemap, feed, CSV, IndexNow lists
  og_cards.py             1200x630 share cards ("Of every $100 spent")
  crosslinks.py           WHOFUNDS cross-links; replies.py: right-of-reply and correction issues by EIN
  indexnow.py, healthcheck.py
content/guides/           the guide backlog, with live-data placeholders
site/static/              app.css, app.js (search, share, compare, embed, chart tooltips, opt-in analytics), favicon
tests/                    unit, privacy, content-rule and build tests; test_live.py checks the live sources

Run it locally

Inputs are public and anonymous; no keys.

python3 pipeline/sources.py check && python3 pipeline/sources.py download     # indexes + BMF, ~2.5 GB
python3 pipeline/select_filings.py                                            # ~25,000 charities, ~150,000 filings
python3 pipeline/fetch_filings.py --workers 48                                # ~30 min cold, resumable
python3 pipeline/build_data.py
CIM_BASE_URL=http://localhost:8765/charity-impact-media python3 pipeline/build_site.py
mkdir -p /tmp/serve && ln -sfn "$PWD/dist" /tmp/serve/charity-impact-media && python3 -m http.server 8765 --directory /tmp/serve

Test

python3 -m unittest discover -s tests -v              # parser, privacy, measures, Deflate64, guard, full site build + content rules
CIM_LIVE=1 python3 -m unittest tests.test_live -v     # parse the newest filing from each live source
node scripts/screens.mjs <site url>                   # screenshots at 1280px and a real 390px mobile viewport -> docs/screens/

Content rules (enforced in code and tests)

• Figures are copied from the filing or are ratios of two lines of the same filing, with the part and line named.
• Every figure links to its filing; the test fails on any uncited dollar amount or percentage.
• No rating language and no status colors; medians are labelled as context.
• Names of people never reach the cache or the site: roles only, business contractors only.
• Every page has a correction link and a right-of-reply link.

Sources

IRS Form 990 e-file data from the GivingTuesday 990 Data Lake and the IRS Form 990 series downloads; the IRS Exempt Organizations Business Master File. Filings are public under 26 U.S.C. § 6104. Filing pages link to ProPublica Nonprofit Explorer.

Licence

CC0 1.0 for data and code. See LICENSE.