The natural key of each source — md_record

What identifies one record in each source, so an observation's id (dev/IDENTITY.md §3) can be computed from the raw file and stays the same across re-releases. Phase 5 step 2.

Usage

md_record_keys()

Value

A data frame, one row per source in md_sources().

Details

Columns:

source, kind: As in md_sources().
filing_key: The column(s) naming one filing or one record: OBJECTID for e-file tables, FORM_ID_NUMBER for Form 8871, the EIN for one-row-per-EIN files.
grain: record: filing_key alone is unique. row: a filing has many rows (one per grant, related organisation, officer), so a row is identified by filing_key plus a hash of its fields plus an occurrence counter among identical rows. Identical rows inside one return are real - the 2022 Schedule I sample has 11 in 4,646 - and the counter keeps them apart.
trace_key: A publisher's row identifier stored with each observation for tracing by hand, but not hashed into its id: e-file TABLE_ID is the row's position in the return, and a parser change could renumber it, while the row's content cannot change.
period: The column giving the observation's year, if any.
asserted_by: For a third-party source, the column naming who wrote the row (the filer).
latest_by: For sources with amended returns: count only the filing with the greatest value of this column per (filer, period) - the latest return (dev/IDENTITY.md §9, decided 2026-10-24).
evidence: How the key was confirmed.

Confirmed on the first 2 MB of each 2022 e-file table and of the 2026-07 e-Postcard and revocation files, downloaded 2026-10-24. A sample can show a key is not unique but cannot prove it is, so md_check_record_key() runs the same test on a whole file at ingest.