The moderation model
Moderation follows one loop: a member reports a message with a category and a note; the report goes to a private queue; an owner or admin decides; enforcement is structural — dismiss, delete, kick, timeout, ban or escalate — and the action produces an audit record and notices.
Two design commitments shape it. Reports are signals, never triggers: nothing is actioned automatically, and there is no automated moderation. And there are only two privileged roles, owner and admin, so accountability lands on named people.
Communities enforce their own rules. Buzz supplies no defaults, which means an unstated standard is no standard.
Above the community layer sits a second one: platform safety — illegal content, network-level abuse, legal reporting obligations. That class is never delegated to community admins; an owner or admin can escalate a report upward and the escalation is recorded durably, but the decision belongs elsewhere. Upstream stated in v0.5.18 where exactly: the platform-safety layer belongs to whoever operates the relay. On a hosted multi-community deployment that is the hosting platform's safety process. On a self-hosted relay it is you, "because the party hosting the content carries the legal accountability". The relay/platform layer also has its own operator-and-moderator roster, separate from community owner and admin.
Why it matters to you
In a workspace that contains clients, external specialists and colleagues, moderation is a duty of care rather than a community-management feature. If someone behaves badly toward a client's staff member in your workspace, the response is yours.
The human-gated design is right, and it means the model only functions if someone actually works the queue.
One piece of governance has moved from convention into the system. Since v0.5.19 the built-in buzz-agent harness authorises every tool call through a permission request before executing it, with the answer decided by client-side policy. It is an action-level control — a gate on what an agent may do in a single step — and it is distinct from workflow approval gates, which would suspend a run until a human decides and which still do not work. Read the two separately: the first governs tool calls on one harness; the second is the decision-level control this guide still meets with written instructions.
How to apply it
Write the community's standards down, however short and obvious they seem. They are what an enforcement decision refers back to, and without them every decision is ad hoc.
Report rather than handle informally when something crosses a line — the private queue exists so the reporter is not exposed.
Work the queue on a rhythm, since nothing will notify you. See Notifications and the Home feed.
Distinguish the action from the conversation. Deleting a message and speaking to the person are different responses, and often the second is the right one.
Decide in advance what warrants telling a client's sponsor, before you are deciding it under pressure.
Before you escalate anything, know who receives it. If you run your own relay, escalation routes to you — there is no backstop behind a self-hosted operator. That is a consideration for the hosting decision, not only for the moderation one. See Decide your data residency and hosting model.
If you ignore this
An unmoderated professional workspace fails slowly and then suddenly: small unaddressed behaviour sets a norm, and by the time it is visible enough to act on, acting looks arbitrary.
The structural enforcement actions are blunt — removing someone from a community is not a small thing when that someone works for your client. Written standards are what make a blunt action defensible.
Examples
A chair rather than a filter. Nothing is screened automatically; someone is responsible for the room.
A report arrives about a dismissive comment toward a client's specialist. The queue is private, so the colleague who reported it is not exposed, and you can respond by speaking to the person rather than by removing them.
Assuming a small professional workspace does not need standards because everyone is a professional. The standards are not for the good days.
Last checked: Buzz v0.5.20, 2026-09-02.
Verified against Buzz v0.5.20 · Updated 2 Sep 2026