We ship models but don't know how to make them safe. Now what?
You ship because the deadline is real and the danger is abstract. Everyone on the pod agrees safety matters. That is exactly why no one is responsible for it. This is not a technical gap. It is a structural one — you built a team, not an owner. You already know more than you act on. You know what red-teaming is. You know what an eval set is for. What you lack is not the knowledge. It is the nerve to let that knowledge slow the ship date, and the clarity of who, by name, is allowed to say no. What follows will not make your pod expert in alignment research by evening. It will make you stop pretending vagueness is a plan.
Safety is not a feeling you arrive at. It is a set of checks, owned by a named person, performed before release — not after. You already have most of the knowledge you need. Choose who owns safety this week. Choose what gets checked. Everything else is noise.
What changes unlock by starting
- The pod can name, without hesitating, who owns safety for the current release.
- You have a written list of specific harms, not a vague sense of risk.
- You have at least one check that has actually delayed a release — proof it has teeth.
- The team has practiced disagreeing about safety before it mattered, not after something went wrong.