The proofreading pipeline's wire contract

The proofreader console — the F4 content cartridge in app-console-content — submits text for editorial review and gets back a rewrite. What follows describes only what's confirmed by the client's own request and response types: the language-protocol vocabulary, the shape of a proofread response, and the verdict step that closes the loop.

Protocol selection is a fixed, real list

Every submission names a protocol from a fixed set of nine, defaulting to prose-topic:

The request also carries a tenant, scoping the submission to the caller's own organisation.

The response reports what actually ran

A successful response carries improved_text (the rewrite) alongside fields that describe how it was produced. tier_used names which compute tier handled the request. degraded lists anything that didn't run at full capability — real evidence that the pipeline has more than one internal component capable of failing independently, though this contract doesn't name those components or their order. audit_ledger_id references the audit trail, and request_id is used to record the verdict afterward. A caller can tell from the response alone whether the full pipeline ran or something degraded along the way, without needing separate status calls.

The verdict is binary, and it trains the model

After reviewing the rewrite, the operator posts a verdict tied to the original request_id. The verdict is binary — accept or reject — and it's SSH-signed the same way the platform's other human-review checkpoints are: a reviewer's signature is checked against the workspace's allowed_signers file before the verdict is recorded. An accepted verdict marks the rewrite as a positive training example; a rejected one keeps the original. This is the same apprenticeship-style verdict mechanism used elsewhere in the platform for supervised model improvement — a human decision converted directly into training signal, never an automated accept.

See also

• Proofreader console — the terminal console that submits text through this pipeline
• Language-protocol substrate — the genre family definitions the protocol list is drawn from
• Customer-tier catalog pattern — the deployment model for a proofreading pipeline instance