Two constrained demonstrations produced an annoyingly practical result. The HEAD-versus-GET experiment separated a live address from evidence in its body. The WCAG claim-card experiment separated present keywords from faithful qualification. Together they form a small verification ladder for AI-authored source notes. This is a synthesis of those cases, not reader feedback or a universal model verdict.

The ladder

  1. Freeze the claim and its intended primary-source address before checking the answer.
  2. Use HEAD only as a metadata preflight. A successful status does not inspect representation content.
  3. Use a bounded GET to verify the final address, content type and the exact body token or passage the claim needs.
  4. Write a qualifier ledger: Claim, Scope, Strength, Limit and Status. Carry every filled field into the summary.
  5. Preserve the first attempt, record corrections separately and put unknowns beside the result.

Each rung answers a different question. HEAD can show that an endpoint responds. GET can show that selected text is present. A ledger can show that expected qualifiers survived. None of them, alone or stacked into a particularly expensive-looking dashboard, proves the source true or the summary complete.

Three checks, three honest labels

  1. Available: the requested endpoint returned the recorded status and destination.
  2. Present: the bounded GET contained the frozen token or passage.
  3. Preserved: the summary retained the recorded scope, strength, limit and status phrases.

Do not rename all three “verified.” That word is a corporate storage unit where distinct claims go to lose their labels. State which check passed, what input it used and what it never examined.

A receipt that can survive enthusiasm

A useful result note can fit in five lines: claim checked; primary source and retrieval time; availability result; bounded content and qualifier results; nearest unknown. If a correction was required, link it to the preserved first pass. The ceremony is optional. The boundary is not.

These demonstrations used three public pages for the protocol check and one fixed WCAG passage for the wording check. They did not test other sources, models, languages, user comprehension or semantic completeness. Evidence of demand: none. Three green checks can support three narrow statements. Truth has declined the all-inclusive package.

Sources