Crash-free sessions, a quiet method
Stability numbers live on engineering walls. Product people glance, nod, and return to funnels. That division is a convenience, and it is costing rooms their honesty. A crash during the first collect, the first lesson, or the first transfer is not a quality incident. It is an activation incident with a stack trace.
Crash-free sessions, as commonly reported, is a blended rate. New builds, old builds, and a long tail of devices that should have been retired all sit in the same percentage. A 99.6% that does not name the build is a lullaby. The quiet method is to treat crash-free as a cohort metric: among sessions that began after the first successful identity, on the build you actually shipped this week, what share ended without a fatal. Then split once, by the screen that matters for activation, and stop.
Put it in the Signal Note when it changes a decision
Not every crash belongs in a product review. A crash in a settings page used by 2% of people may be a ticket. A crash on the collect confirmation used by everyone on Saturday morning is a product event. If the rate on that screen moved, the Signal Note should say whether you will hold the release. If you will not hold the release, the Refusal field should say you are accepting the rate. Hiding the number in an engineering channel is a third option, and it is the one that produces surprise later.
We do not teach crash symbolication in the studio. That is a limitation, and it is written into the flagship FAQ for a reason. We teach how to ask for the number in a population that matches the rest of your App Analytics. Engineers who have sat the table tell us the request is a relief: they are tired of being asked for a single percentage that means nothing.
Sessions are not users
A person who crashes three times and still completes the job will inflate session counts and still look like a success in a funnel that only watches completion. Join the two. If you cannot join them, write that you cannot. Incomplete joins are more respectable than a dashboard that pretends the join happened.
The method is quiet because it adds one field, not a new religion. When stability is fine, the field is empty. Empty fields are a luxury. Use them.