ADR 302 · Accepted
The final answer of an editorial run is checked before it is stored
Context
An A2A client asks for a text it uses outside TYPO3: a social post, a reply to a comment. ADR-300 added a service that checks such a text against the rules of a configuration at every length. Nothing called it. The pipeline checks only what a writing tool would store in a run that offers tools, so the answer of a run that only read is not checked at all, and the judge skips a text under 30 words.
Two more needs follow from external editorial work. An editor changes a draft outside nr-llm and wants the changed text checked, not a new one written. And the check must not run under the run-row lock that completes an A2A task (ADR-282): a judge call there would hold the lock for the length of a model call.
Decision
An end-of-run check. The run executor asks the services tagged nr_llm.run_output_check whether a run's answer is checked. The A2A implementation answers yes when the task's profile names a handler that declares itself end-checked. After the loop, before the transaction that completes the run, the executor hands the final answer to that check. It runs outside every transaction and renews the run's lease first. On the three paths a run takes (a start, a queued run, a resume) the decision is made once, in the same place.
What is checked. The text the content artifact will carry: the answer as the current answer level keeps it, shaped by the handler. A level that keeps no answer delivers nothing, and nothing is judged. The check is the service of ADR-300 with the run's configuration, the run's user and the run's uuid as correlation id.
Its result. A run event output_check keeps the report without any passage of the text: status, reasons, digests, and the rules and terms found broken. A term found and the digest of a short text still say what the text held, so below redacted (the run's event level, when the event is written and when the level tightens later) the event keeps status and reasons only. A run event rule_state keeps the rule-state digest at the start of the run. A blocking report (a broken term, a rule broken under a judge in block mode, a check that could not run) stops the run through the guardrail arm, as a policy stop with a reason that names lists and rules, never a term or a passage. The task ends FAILED.
One judge call per text. While the loop of such a run runs, the pipeline judge leaves the run's answer to the end check. Only the answer: the arguments of a writing tool are judged in the call that proposes them, as before. The flag lives in a container-shared service the executor sets; no request field and nothing a client sends switches it.
Two handlers.
nr_llm.editorial_text writes the text from the client's message, with the read-only tools the profile allows, and is end-checked. It offers no writing tool and no tool that asks the client a question (ask_choice and the like): the question of a pause reaches the client before the final answer exists, so it would carry model text past the check, and a write would stay behind a text the check stops.
nr_llm.editorial_check checks the text the client sends. The run takes the submitted text as its answer: no generator call, no tool call, the text byte for byte. A comment or a draft that asks for a tool gets none.
nr_llm.text is unchanged.
Consequences
A short answer, and the answer of a run that only read, are checked when the profile chooses an editorial handler. Such a run cannot ask the client a question in this version; a missing fact is a refusal or a gap the text names.
A run checked at its start stays checked: a segment whose profile no longer names an end-checked handler fails. The decision is made when a worker first takes the run, not when the task is accepted: an administrator who changes a profile's handler while a task waits in the queue changes how that task runs. A result whose check could not be stored is not delivered either.
A runtime built by hand without the tagged checks (AgentRuntime falls back to an executor without them, as it does for the start and resume guards) does not check; the container always wires them. A profile with nr_llm.text behaves as before.
A changed draft can be checked without being rewritten; the judge is the only model call.
The judge call happens before the completing transaction. The completion itself makes no provider call.
The client of a blocked task sees The task failed.; the reason reaches it with the report artifact of ADR-301.
A judge in block mode that cannot answer stops every text of that profile. That is the point of block mode.
Under the answer level redacted the check-only handler judges the redacted text, because that is what is delivered, not the text the client sent.
Known limits, accepted after review (each needs an administrator's edit or a lost lease inside one worker's window between the check and the completion):
An answer level loosened between the check and the completion delivers a text the check read at a stricter level, or not at all at metadata. Its report says not_checked (the digests differ), so it is never passed, but it is delivered.
A run whose settle was lost after a blocking check and that a second attempt stops by another guardrail before its own check reports the stop with the record of the first attempt. Its state is needs_human_review.
Revisit when
A handler needs structured input (channel, site, language per profile). That needs a profile field and changes ext_tables.sql.
A paused run must keep the rule texts it started with.
Alternatives considered
A flag on ``AgentRunRequest``. Rejected: the request is @api with a frozen constructor, and a field the request carries could be set by whoever builds the request. The server derives the decision from the task's profile.
Check inside the A2A completion. Rejected: the completion holds the run-row lock, and a judge call there holds it for a model call.
Run the check in the TCK-style request path. Rejected: a judge call would hold the HTTP request.