source-evaluation

Invoke the source-evaluation skill with a person ID or name. It reads the profile's attached sources via person_read, fetches each indexed record with record_read, and compares indexed values against recorded facts to produce a classified findings report.

Detach only when the record is genuinely about a different person — multiple fields disagree, the record names unrelated family members, or dates fall outside the person's lifetime. Indexing errors should be fixed by re-reading the original and correcting the index, never by detaching.

An indexing error is a single mis-transcribed field in a record that otherwise fits the person; a misattribution is a record about an entirely different person. Ambiguous single-field cases are classified as indexing errors with a recommendation to re-read the original first.

No. The audit is read-only — it writes nothing to research.json or tree.gedcomx.json. Its only output is the findings report, and genuine source conflicts are handed to the user to take to the conflict-resolution workflow.

Sources without a readable record-persona ARK — user-uploaded documents, external links, or memories — cannot be fetched via record_read. The audit reports each of these as unchecked rather than guessing at their contents.

No. The skill holds no image tool and none of its three tools returns an image id. It reports what the index says and names what the researcher should go back to — the original page for image collections, or the underlying certificate for index-only collections.