Add July 2026 refresh changelog to dataset card

README.md

@@ -95,3 +95,42 @@ Database build code: MIT. Underlying data: public domain (U.S. government record
 ## GitHub
 
 Full source code, build instructions, and example queries: [github.com/ian-nason/fec-database](https://github.com/ian-nason/fec-database)
+
+
+# Changelog
+
+## 2026-07-06 — Full refresh + data-quality audit
+
+Rebuilt from current FEC bulk downloads (cycles 2004-2026) and repaired
+after an independent SQL-verified audit.
+
+**Data changes**
+- 347,171,337 rows across 10 tables (previous published build: 339.5M).
+- Cycle 2020 individual contributions complete at 69,352,160 rows.
+- `independent_expenditures`: 74,030 superseded amendment versions removed —
+  only the latest version of each filing is kept (2022 IE total dropped from
+  $46.9B to $33.6B; remaining inflation is upstream prank filings, see
+  caveats).
+
+**Fixes**
+- Standalone-table dates now parse (previously 100% NULL):
+  `independent_expenditures` (DD-MON-YY), `communication_costs` (YYYYMMDD),
+  `electioneering_communications` (DD-MON-YY, amounts now numeric).
+- All four views rebuilt with correct money semantics: conduit rows
+  (`TRANSACTION_TP = '24T'`, the ActBlue/WinRed double count worth 14-22% of
+  every cycle since 2018) excluded; refunds (20Y/21Y/22Y, stored as positive
+  amounts) netted out; memo rows excluded.
+- `v_candidate_totals` aggregates per principal campaign committee, so
+  candidates sharing a committee (Biden/Harris 2024) no longer fan out into
+  double-counted rows.
+- `v_pac_to_candidate` is direct contributions only (24K/24Z) — it
+  previously counted independent expenditures *against* a candidate as
+  support.
+
+**Known caveats** (see the README's "How to sum money correctly")
+- Upstream prank filings remain in `independent_expenditures` (e.g. a $10B
+  "COMMITTEE 300" row) — treat amounts above ~$5M as suspect anywhere.
+- A few thousand rows carry wild dates (years 0677-9206); clamp to the
+  cycle window before bucketing.
+- Recent-cycle `committee_contributions.CAND_ID` values fail to resolve to
+  same-cycle candidates at 1-5% (FEC master-file churn).