What it shows
Reading the evidence section
Each claim is shown with the source that supports it and the excerpt that support was matched against, so the attribution can be checked rather than trusted. Confidence is banded rather than scored to three decimals, because the underlying signal is term overlap between the claim and its source excerpt. Anunsupported band means the source did not carry the claim, which is the finding a reviewer needs to see first.
An empty evidence section is not automatically a defect. Claim binding runs over inline citation markers, so a template that does not cite produces no claim records — expected for social posts, scripts, and landing pages. On a research format, an empty section means the citations were never bound to their sources, which is worth investigating.
Three states, not two
Assurance components distinguish ran, did not run, and not recorded. A stage that was skipped is different from a stage whose outcome was never written down, and both are different from a stage that passed. This mirrors how executable verification reports its own findings:VERIFIED, VIOLATION, and UNVERIFIABLE are three separate verdicts. A viewer that rendered an unrun check as a blank or a tick would turn a coverage gap into false assurance, which is the failure the governance layer exists to prevent.
Reproducibility and replay
Check now compares the asset’s recorded configuration against what the deployment can currently supply, and reports drift per field.
Show replay request returns the generation payload that would reproduce the asset. Replay does not run the generation for you: the payload is submitted through the normal generation path, so a reproduction is metered, appears in history, and debits credits like any other run.
Exporting
Passport JSON downloads the complete record for a compliance file or an external audit. Its shape is versioned — a consumer reading a v1 passport can keep reading it after v2 exists. Print produces a paper or PDF copy of the dossier as displayed.Assets without a passport
Imported documents, hand-written drafts, and anything generated before passports existed have no production record. The dossier says so and stops. A partial record is not synthesized for them. An official-looking document that no pipeline stage ever vouched for is worse than an honest absence, particularly in front of a reviewer who is deciding whether to publish.Limits
- A passing check establishes that a rule was executed and held. It does not establish external factual truth.
- Confidence bands measure overlap between a claim and its source excerpt, not semantic entailment. Semantic claim verification is a separate, opt-in check.
- The dossier reads recorded state. Opening it does not re-scan sources or re-check links; run a validity scan for that.
- Reviewer identity is recorded for approvals routed through the Reviews queue. The Premium editor-review interrupt records a decision without one, and the dossier reports that as “not recorded” rather than implying nobody reviewed it.
Related
- Content Governance: how the passport is assembled and what each section means
- Review Queue: routing and recording qualified sign-off
- Content Management: the library the dossier is opened from