All guides
    Mission
    4 min read

    Why Navōn fails closed when the safety check can't run

    Every filter has an outage sooner or later. What matters is what the product does when the check cannot run. Navōn stops and asks you to try again, and here is why.

    Why Navōn fails closed when the safety check can't run

    Software that checks things sometimes cannot check. A network call times out, a provider returns nothing, a key is missing after a configuration change. At that moment the product has to decide between two behaviours. It can answer anyway, on the theory that most requests are fine. Or it can stop. Navōn stops. The usual name for this is failing closed, and it is a deliberate choice.

    The two ways a filter can fail

    A filter that fails open treats "I could not check" as "nothing found". It is the more comfortable design, because outages are invisible to users: everything keeps working. It is also the worst failure for a platform whose promise is that requests are screened before they reach a model. Nothing on any screen would say that the screening had quietly stopped, and a child's question would go through unexamined.

    A filter that fails closed treats "I could not check" as "not yet cleared". The cost is plain: during an outage, people are told to wait. We decided that a short, visible pause is a better trade than an invisible gap. Our own code comments say it in nearly those words: an outage stopping generation is better than an outage quietly turning the product into unfiltered AI for children.

    What is covered

    The same rule applies at each point where Navōn checks something:

    • The question. If the safety check on a request cannot be completed, the request is denied rather than passed to a model.
    • A missing configuration. If the safety service is not set up at all, generation is paused rather than running without it.
    • The answer. The text a model writes is checked as well, and if that check cannot be completed the answer is not released as safe.
    • Pictures. An image is checked before it is stored, and if that check cannot be completed the picture is not kept or shown.

    One retry happens first. If the first attempt fails, Navōn waits a moment and tries once more, using a shorter version of the rule list in case the provider rejected the longer one. Only if the second attempt also fails does the request stop. A brief network hiccup therefore usually passes without anyone noticing.

    What you see

    This is not presented as a block. The message says the safety check could not be completed and asks you to try again in a moment. We worded it separately on purpose: telling a student that their question broke a rule, when in fact nothing was decided about the question, would be untrue and would teach them the wrong lesson. The failure is logged as a check that was unavailable, so it is not mixed up with real refusals. Our post on what happens when Navōn refuses a request explains how genuine refusals look and why they appear inside the conversation.

    What it costs

    We will not pretend this is free. If the checking service is down for an extended period, people cannot generate anything until it returns, and that includes perfectly ordinary questions. A teacher preparing a lesson may be told to try again later. That is a real inconvenience, and it is the price of a guarantee that does not depend on luck.

    There is a second, quieter cost. Checks can fail for reasons other than an outage. A provider can decline to respond to exactly the sort of message the check exists to judge, such as a child describing something frightening. Failing closed there would be the wrong result for a child asking for help, and we have had to correct that kind of case in the checker itself rather than simply accepting it. Failing closed is a safety property, not an excuse for a check that fails often.

    Why this matters to a school or a parent

    The practical question for anyone choosing an AI tool is what it does when its own safeguards break. Ask vendors directly. Because Navōn checks on the server and not on the device, the answer does not depend on which phone or laptop the child is using. We make no claim that any filter is perfect. The claim is narrower: when the check cannot run, Navōn does not guess.

    See what runs on every request in Navōn.

    Common questions

    What does fail closed mean for an AI filter?

    If the safety check cannot be completed, the request is not answered unchecked. It is stopped, and the person is asked to try again shortly.

    Why does Navōn say the safety check could not be completed?

    The check on your request or on the answer did not finish, for example because of a network problem. Nothing was decided about your question, so try again in a moment.

    Does a failed safety check mean my question was blocked?

    No. A block means a rule applied. An unavailable check means nothing was decided, and the message is worded differently for that reason.

    Will Navōn ever answer when the filter is down?

    No. Requests, answers and pictures that cannot be checked are not released as cleared, even though this means a short wait during an outage.

    Published by Navōn. How these guides are written and checked.

    Read next