From primary record to public signal

1. Ingest and validate

Each connector follows the official source contract, uses server-only credentials when required, limits request time, handles pagination and provider limits, and validates every response before persistence. Invalid records are rejected rather than coerced into plausible-looking data.

2. Preserve provenance

Normalized records retain the source agency, dataset, URL, record identifier, publication and update times, retrieval time, ingestion run, source-record hash, and normalization version. Raw payloads are retained only when appropriate and permitted.

3. Detect change

Snapshots are compared field by field. New records, revisions, withdrawals, threshold crossings, large changes, and ranking moves become structured change events. Dataset-specific scoring considers magnitude, rarity, reach, economic value, history, and source importance.

4. Explain with limits

AI may summarize verified records and draft plain-English context. It may not invent a value, cause, quote, API result, relationship, or degree of certainty. Numeric claims must resolve to structured source records or deterministic calculations.

5. Review and correct

Health and safety claims, sensitive national-security subjects, low-confidence entity matches, major accusations, ambiguous data, and failed validation stay in human review. Substantive published changes preserve an update or correction history instead of silently replacing the old claim.

Interpretation cautions

• An award amount is not the same as money paid.
• Facility proximity does not prove exposure or health risk.
• Preliminary, revised, estimated, observed, forecast, and declared data are distinct states.
• Geographic detail is shown only at granularity supplied by the source.