A provenance checklist for public-signal monitoring

Store observations separately from conclusions

A monitor should preserve what it observed before deciding what the observation means. An unavailable page is a retrieval result. It is not, by itself, proof that an organization withdrew a statement. A changed page may reflect a new claim, a template update or an unrelated edit.

Use separate records for the captured source, extracted signal and analytical conclusion. This makes a later correction possible without silently rewriting the original observation.

The capture record

For each observation, retain enough context to retrieve or reconstruct the evidence within the rights and retention rules that apply to it.

• Requested and final source location, with redirect information where available.
• Capture time and outcome, including failures.
• Original content or a permitted snapshot, content type and fingerprint.
• Extraction method and version, with the exact supporting passage.
• Source-published dates separately from capture dates.
• Known omissions, permissions and retention constraints.

Compare the right representation

Comparing raw bytes catches template changes as well as substantive edits. Comparing extracted text can reduce that noise, but an extractor may omit tables, images or meaningful formatting. Retain the raw capture and the derived representation so the comparison can be inspected.

Treat an unreachable source as unknown rather than unchanged. Retry according to the source's access rules, and record persistent failure. Do not bypass authentication or access controls to complete a monitoring record.

Review before action

A review record should identify the claim under consideration, the evidence used, the decision and remaining uncertainty. Use stronger review for actions with higher consequences. This checklist is an AEY design proposal informed by the W3C primer; it is not a field study or a guarantee that monitoring is lawful in every context.

AI assisted preparation. The company owner approved editorial/disclosure publication on 3 October 2026. No live-source collection or client information underlies the examples.

Sources

• W3C PROV Model Primer