Running checks and governed runs
Two new pages in your portal let you try Clara before anything reaches a customer. Compliance check gives a fast pass, warn, or block verdict on a draft message, judged against a small built-in floor plus the rules you have given Clara. Governed run asks Clara a real question through the full governance pipeline and shows you exactly what she found and what to do next. The governed run writes a Record you can point to later; the compliance check gives you a reference, not a Record.
This page is for everyone. No technical background needed.
Compliance check
Open Compliance check in the portal sidebar, paste the message you are about to send, and press Run check.
- Pass means nothing the check knows about was tripped. Read the next section before you read that as "this message is safe".
- Warn means something needs a second look. Each warning names the rule and says what it saw, for example a written refund promise.
- Block means the message must not go out as written.
The check reports a verdict; it does not block the message itself. Your own code decides what to do with a warn or a block.
What the check is checking against
Two things, and only two.
The floor. A small set of built-in rules every account gets, whether you have taught Clara anything or not. It is deliberately narrow: attempts to talk Clara out of running her own checks, a short list of prohibited content, refund promises made in writing, legal threats, and runs of digits shaped like a card or a national ID number. We do not publish the exact wording it looks for, because that would be a guide to writing around it.
Your rules. The policies and compliance documents you have given Clara, turned into rules the check can apply. This is the part that knows your business. When you have a rulebook, the check can also have a model read the message against it. With no rulebook there is nothing to read the message against, so that step does not run at all.
So the honest reading of a pass is "nothing in the floor and nothing in your rulebook was tripped". If you have uploaded nothing, only the floor ran, and the floor knows nothing about your industry, your customers, your promises, or your obligations. It will not notice a decision made for a reason you are not allowed to use, or an email going to people who asked not to be emailed. Those are questions about your business, and the check can only ask them once you have told Clara what your business rules are. See Teaching Clara. Every policy you give her makes this check stronger, and that is the intended way to use it. For a verdict that weighs the whole situation rather than a rule list, use a governed run.
Why it is still worth running
- It is the same rules every time, not an opinion, so the same message gets the same verdict today and next month.
- It runs outside the model that wrote the message, so a model cannot mark its own homework.
- It is fast and cheap enough to sit in front of every outgoing message.
- It catches messages that try to switch governance off. A draft claiming the compliance checks were disabled for this batch and should be sent anyway is blocked in a fraction of a second.
The sandbox switch is on by default, so trying things costs nothing. Turn it
off when you want a billable check with a hold receipt. Every check returns a
decision_ref, prefixed cc:, for joining your logs, webhook deliveries and
support tickets to one verdict. It is a correlation handle, not a Record: it
does not resolve through GET /v1/records/{run_id}, and a compliance check
never appears there. For an audit artifact, use a governed run.
Governed run
Open Governed run, type a question the way your team would ask it, and choose a mode:
- Briefing only builds the briefing document and stops. Use this to see what Clara knows and what she would need before acting.
- Brief + governed execute also lets Clara produce the answer when the decision allows it.
The result page leads with What to do: the decision (answer, ask, escalate, or block), the rules that bound it, and the clarifying question to send back when the decision is ask. Below that you see the findings from each of Clara's advisors, and a plain-words note whenever a check was unavailable so you know what the decision did not consider.
The internal verdict shown next to the badge is Clara's own reasoning note. The decision in What to do is the one to act on.
Direct run
Direct run skips governance: your instructions go straight to a managed model. It is always metered and the result is marked not governance-attested. Use it for drafts and experiments, never for anything that needs an audit trail.