The queue that decides nothing

Every marketplace needs moderation. The naive implementation is an approval queue: a listing is submitted, it waits for a human, the human approves or rejects it.

And then, on a real campus marketplace, two things happen. The queue is slow, so sellers learn that listing early does not help, and they start listing something else - a workaround, a rival channel, a DM. And the moderator, facing a backlog, starts approving things in batches without looking. The queue is now a bottleneck for the honest sellers and a rubber stamp for the dishonest ones. It is the worst of both.

The question is not whether to moderate. It is where the human goes.

The mistake in the design

The mistake is putting the human in the path of every listing. It treats moderation as a gate on all traffic, when what it should be is a risk-based filter on the traffic that matters.

The difference shows up in the numbers you can actually observe. If every listing waits for approval, the median time to live is bounded by moderator attention - a few hours at best, a day or more on a busy week. If only the listings that look risky wait, most listings are live in seconds and the queue contains only the hard ones, which is a queue a human can actually look at carefully.

The three-tier design

The structure that works is three tiers, and the tiers are decided by how much of the listing is machine-checkable.

Tier one - rules. A listing whose title, price, and category are all within the obvious bounds goes live immediately. The checks are deterministic and fast: the price is within the allowed range for the category, the title is not a contact detail, the images are the right kind of file, the seller is not suspended. This tier carries the overwhelming majority of listings, and it is the tier that has to be fast, because it is the tier everybody uses.

Tier two - signals. A listing that trips a signal - a price far outside the range, a description that reads like a duplicated one, a burst of listings from one seller, a category that does not match the title - goes to a review queue, and it is visible to the moderator with the reason attached. Not "please review". "Price is 3x the category median. Four near-identical listings from this seller in the last hour." A moderator can clear that in seconds because they do not have to work out why it was flagged.

Tier three - a human, on purpose. Fraud, harassment, and disputes do not belong in a listing pipeline at all. They arrive through reports, through payment events, and through the transaction lifecycle, and they get a case - a thread with the parties, a history of what happened, and a decision that is written down. This tier is slow on purpose, because it is the tier where being slow is the point.

The part that is easy to get wrong

A queue of unclear items is a queue nobody clears. The single biggest improvement is not a better model or a faster classifier - it is attaching the reason to the item. A moderator who knows why something was flagged is doing judgement; a moderator looking at a generic listing is doing nothing but reading.

Do not make the fast path conditional on the slow path. If tier two is a queue that is not being cleared, listings that could have been live in seconds are stuck behind it. Each tier has to be able to stand on its own.

Report the thing, not the person. A moderation system that only shows moderators who did something wrong teaches the moderator to act on people. Showing the transaction, the history and the pattern is what makes the decision fair.

The appeal path is not optional. A rule that can be wrong - and tier one rules will be wrong sometimes - needs a way to be corrected. If a seller can appeal a rejection and the appeal reaches a human, the rules improve. If the appeal is a form nobody reads, the rules do not, and the system slowly becomes unfair.

The part that is a product decision

What does the seller experience while a listing is in tier two?

This is the question that decides whether the design works. If tier two is invisible - the listing just appears later, with no explanation - then the seller experiences a delay and a feeling of being suspected, and the fastest thing a seller can do is create another account. If tier two is visible and specific - this listing is under review because the price is outside the usual range for this category; it usually takes a few minutes - then the seller knows what happened, can fix it, and waits.

And if the review is negative, the seller deserves to know what to change. A rejection that says "this violates our rules" teaches nothing and is functionally the same as a random failure. A rejection that says "this price is above the range for this category" teaches the seller how to sell on the platform.

The moderation system in FUWMarketplace is built on this shape, and it is worth being clear about what it is not: it is not a classifier that decides what is bad. It is a routing system that decides who should look at this, and what they need to see when they do.