How the community review pipeline works: a model first, a person when it is unsure

Show & TellHysen Yan3 replies

Hysen YanStaff#1

Every post on this forum is reviewed before it appears. This is what that actually means, mechanically, because the rule is on the board and the mechanism is not.

The three outcomes

A model reads the post and returns one of three verdicts: approved, rejected, or unsure.

Approved means the post is public immediately. Most posts are, and most are public within a minute of being written.

Rejected means the post is not published, and the author is told why in a form they can act on. The reasons are specific: promotion, a referral link, recruiting, an unedited personal attack, a post that is nothing but a link.

Unsure means the model is not confident, and the post goes to a person. This is the case that makes the system honest: the model is not asked to be right, it is asked to be willing to say it does not know.

What happens when the model is not available

Nothing is published. A post whose review could not complete stays pending and waits for a human. The rule is that the model failing must never result in a post going live unreviewed, which means the retry job exists and the attempts are counted.

The two rules that are not about safety

Links carry rel="nofollow ugc". Every link in a post, no exceptions. A community that passes link equity gets spammed for it, and removing the incentive removes most of the spam before it starts.

No images. A remote image is a tracking pixel the reader never agreed to see, and it doubles the moderation work, because every image has to be looked at separately from the text around it. Markdown is enough: code blocks, lists, quotes, links.

The parts that are policy rather than machinery

Trust levels do not skip review. A member with a long history of good posts is not exempt, and that is deliberate: an account that gets sold or stolen should not come with a licence to publish unreviewed.

The daily limits are on the mechanism, not the intent. A new member can start two topics and write fifteen replies a day; the limits grow with what they have done, and they are computed from behaviour rather than granted.

Nothing is editable, only deletable. An edit would have to go back through review and would invalidate the notification that already went out to the people who replied. So a correction is posted as a reply, which is more honest anyway.

3 replies

EdenMember#2

The "the model is asked to be willing to say it does not know" framing is the right one, and it is unusual.

Most systems have two outcomes and a threshold, and the threshold drifts towards approval because every false rejection costs a moderator time. Three outcomes with an explicit unsure is a better shape.

Wister XueMember#3

In reply to Eden (#2)

Worth saying what the model is actually checking, because it is narrower than people assume.

It is not judging whether the post is good. It is looking for the small set of things the rules forbid: a link that is clearly a referral, a recruiting sentence, an attack on a person, a post that is only a link. Whether the reply is useful, whether the question is well-posed, whether the argument holds: none of that is its job, and a forum that tried to make it the model's job would be worse.

Hysen YanAuthorStaff#4

In reply to Wister Xue (#3)

That is exactly right, and it is the line we were careful to draw. The reviewer answers one question — is this post one of the things the rules forbid — and everything else about quality is left to the readers and to the author's own judgement.

A moderation model that tried to judge quality would reject good posts constantly, and the people who would notice are the people who wrote them.