Methodology.

If a figure we serve cannot be traced through the method described here, the figure is wrong and we want to know: corrections@atcstw.com.

1 · How we collect

ATCSTW ingests public records from their issuing sources — federal and state registries, disclosure systems, and filing repositories. Source data is preserved as retrieved. Cleaning happens in versioned processing stages, never by editing the retrieved record, so any served figure can be traced back to the bytes it came from.

2 · How we resolve entities

Records name the same entity differently across sources. ATCSTW resolves records to entities using identifier matching where authoritative identifiers exist and documented matching rules where they do not. Resolution decisions are themselves databased: an entity in the graph carries the records merged into it, and a merge can be inspected and reversed.

3 · How we dedupe and version

Each dataset passes ordered processing stages — normalize, dedupe on natural keys, validate and reconcile (counts and sums are checked against source manifests; a stage that does not reconcile does not advance), then linkage and aggregation. Every stage is versioned, and the version that produced any served figure is part of its provenance.

4 · Provenance: the source stamp

Every figure ships with its source: the issuing agency or registry, the source endpoint or file, the period covered, and the retrieval date. The stamp format is fixed — SRC: AGENCY · DATASET · PERIOD · AS OF DATE — so a citation can be checked mechanically.

5 · Measured vs. modeled

Figures taken directly from source records are labeled measured. Figures computed by ATCSTW — estimates, rollups, modeled values — are labeled modeled, with the method stated. The two are never blended into a single figure, anywhere, at any tier.

6 · Connections and confidence (Lux)

Lux, the connection engine, databases relationships the records support: ownership, affiliation, money flow, shared officers, contract history. A connection is shown with the records that support it. Connections the records do not prove are not shown. Correlation is labeled correlation wherever it appears.

7 · Corrections

Errors are corrected at the stage where they entered, re-served, and recorded in the public Corrections log — dated, specific, permanent. Our commitment to ourselves: a confirmed error is logged the same business day it is confirmed.

8 · What we do not do

We do not serve individual FEC contributor names or addresses (aggregate-only, §30111(a)(4)). We do not build individual wealth or vulnerability profiles. We do not sell user-activity data from downstream properties — it is aggregated, coarse-location, and never sold. We do not publish a figure we cannot stamp.