A VETTING step has an independent reviewer check each reply against your rules. Use it when replies must follow rules you can state.
VETTING (Verification and Evaluation Tool for Targeting Invalid Narrative Generation) comes from classroom research (a middle school study). Use it for a tutor that never gives the answer away, or an assistant that says only what its sources say.
How it works
The reply model writes a reply. The reviewer model checks it in a separate call and gives a verdict and a reason. A failed reply goes back with that reason, until one passes or the Attempts run out.
The reviewer's rules live in the step, apart from the reply model's instructions, so a learner can't talk the reviewer out of them. It reads the reply and nothing the learner wrote. With an answer key, a reply that confirms an answer as the learner's own passes, one that supplies it fails, and one that calls a wrong answer right fails.
A learner who has the answer is told so. Code compares what they wrote with the key (a number equal to it, or its own few words), or reads a Questions panel's marking, and only that yes or no goes on, so no model reads their words for it. The reply then says they are right and may talk about the answer, and a reply in that turn that leaves them unsure fails. Where code can check and they haven't written it, the reviewer hears that too, so a learner who only claims to have the answer doesn't get it confirmed; an answer in words code can't read counts as not written. Where code can't check at all, as with a key that is a long sentence, a rewrite is told to keep saying they are right if they are. Each reply and each review is a model call. A review is a judgment, not a guarantee.
Fast or careful
A learner waits for every round, one reply and one review, with nothing on the page until one passes. With a strict reviewer that can be half a minute. How much care is worth how much waiting depends on your learners, so it is yours to set, and the defaults are only a place to start.
| To change | Do this | And |
|---|---|---|
| A shorter wait | Fewer Attempts, or a faster Reviewer model | More replies end unverified, or pass with a lighter review |
| More care | More Attempts, or a stricter Reviewer model | A longer wait, and more cost |
| No wait at all | Take the VETTING step out | Nothing checks replies before learners see them, and they arrive word by word |
Try it and Run history show how long each VETTING step took and how many rounds it ran, so try it on your hard cases before you choose.
Set it up
When you add a VETTING step, choose how to start: Check replies against a guideline, Don’t give a question’s answer away, Check an earlier step’s answer, or Blank. Then:
- Choose what it checks against. Under Setup → Checks against, pick A guideline or A question and its answer key. The Answer key is seen only by the reviewer.
- Set the reply model. Choose a Reply model. Add option → Reply instructions replaces its default prompt. Safety rules are always added.
- Choose the first answer. Under Answer to review, pick The reply model writes it or An earlier step’s answer.
- Set the reviewer. Add option → Reviewer model chooses it. Left unset, it is
gpt-6-luna. - Set the limits. When something goes wrong → Attempts is 3 by default, 1 to 10. End every reply with a question is on unless you add it from Add option and turn it off.
- Connect the outcomes. Under Outcomes, choose the step each one goes to.
A guideline can add a Reference for the reviewer from Add option, such as {{search.context}}, to check facts against. To try the VETTING tutor template, open Try it → Reference data → JSON and enter:
{"question": "Solve 3x = 12", "answer_key": "x = 4"}
Outcomes
- Passed: the reviewer approved the reply.
- Unverified: no reply passed, or a review couldn't run. Later steps get the last reply, not approved.
- Needs attention: the reply model found the conversation unsafe, such as self-harm. It wrote a caring reply instead, which isn't reviewed.
What learners see
Learners see only what the Reply step on each path says. They never see drafts, the reviewer's reasons or the answer key.
Warning
Decide what the Unverified path says. The VETTING templates give a fixed message instead of the unapproved reply.
Insights lists replies that didn't pass.
When a teacher sets How the AI helps
In a course activity, a teacher uses How the AI helps to say how much the AI does. Say a Grade 6 teacher chooses Gives hints. The reply model gets it as instructions, headed "From the teacher of this activity". The reviewer checks each reply against it too.
What the reviewer reads is only the reply unless you choose otherwise: in a step's settings here, or for one activity in How the AI helps. With the learner's message too, the reviewer can tell their own idea from one handed to them, and it fails a reply that doesn't answer them; their message reaches it as data, but a learner who knows it is there could write to it. A step's own choice wins over the activity's. What a reviewer reads.
The agent profile's What it reads decides which steps get it. With Only as template variables, nothing is added; use {{context.workspace.teaching}} yourself.
- How the AI helps: what a teacher tells the AI in an activity.
- Testing and run history: read each VETTING round in a run's trace.
- Knowledge bases: give the reviewer sources to check facts against.
- VIABLE Open Kit: check replies in your own app with the same reviewer.

