Bring an existing review without inventing new authority

You can begin with a versioned paper and an ordinary review letter. The prototype asks what that review actually supports, then keeps unresolved questions visible. An agent can prepare the structured wrapper immediately; a human must still make or authorize any real responsibility assertion attributed to them.

A six-question card, with the matching agent action

Human question	Agent action in this example	If the answer is missing
1. Which exact claim and version? Point to the precise result, immutable release and evidence.	Preserve claim ID, declared git object and evidence digest; keep the ordinary URL as context.	Stop the executable conversion; do not treat a moving page as an exact release.
2. What was checked? Name the method, scope, result and limits in the reviewer’s terms.	Keep declared check scope/result/method separate from the recommendation to accept.	Preserve unknown; never fill in a positive result from tone or recommendation.
3. Who can take responsibility, and for what use? A reviewer, author and editor have different roles.	Leave decision authority unknown until a separately explicit assertion exists.	No scoped reliance or publication permission is inferred.
4. What connections could limit independence? Include shared control and inaccessible prerequisites.	Preserve disclosures; never infer independent groups from different names or many team members.	Unknown remains unknown. Nonempty unsupported disclosures must be reconciled, not overwritten.
5. What effort and conditions can be supported? State purpose, capacity and expiry.	Require explicit fixture capacity, exact scope, purpose and expiry before the toy run.	Do not promise completed coverage or an appeal service.
6. What would make us reconsider? Identify dependency changes and a correction path.	Record dependencies separately and exercise a material amendment.	Missing dependencies remain a known blind spot; the adapter cannot discover them.

The machine-readable path is runnable from the extracted packet root:

python3 -m adapters.demo

First, the original legacy metadata is refused for scoped reliance. Then a separate invented fixture supplies the missing assertions solely to demonstrate the actual toy protocol. Its old reliance becomes pending after amendment. The original source, every absent/null/empty distinction, and the semantic loss report stay attached to the result.

The example does not parse arbitrary prose, contact a journal, authenticate a reviewer, establish consent, or claim conformance with an existing interoperability standard. It shows a practical entry point that agents can execute today while keeping the human questions explicit. A real adapter should earn trust through independent round trips and permission-aware integration tests before live use.