The Review Queue

The first screen of Duplicate review. It answers one question — how much work is there, and which part of it is worth doing first? — and hands you off to a pattern.

The layout is constant across every section: the work on the left, the instruments on the right. Moving between sections never rearranges the page under you.

Pairs remaining

The headline number, and the one everything else is measured against.

Pairs remaining = matched pairs inside your review band that nobody has decided yet.

Three things are worth knowing about it:

• It counts pairs, not clusters or assets. A pair is one decision.
• It only counts what’s inside the band — see cutoffs below. Move a cutoff and this number moves with it.
• Decisions taken by an agent are counted separately, so an agent clearing the safe band can’t make the queue look like a person worked through it. See AI Agents & Duplicates.

Underneath it sits the duplicate rate.

Duplicate rate = assets appearing in at least one matched pair ÷ all assets in the workspace.

A count of clusters means nothing without a denominator. “4,100 duplicate clusters” sounds alarming until you learn it’s out of two million assets — or terrifying when it’s out of nine thousand. The rate is the honest version, and it’s the number to quote to anyone who didn’t run the scan.

Where the matches stand

A three-part strip that splits every matched pair into the bands your cutoffs define:

The shape of this strip is diagnostic on its own:

• Almost everything above the merge cutoff — the matcher is being generous, or you have a genuine mass-duplication problem. Check a few pairs before believing it.
• Almost everything below the review cutoff — the matcher is being strict, or your detectors aren’t producing the kind of concrete values that make good evidence.
• A fat middle — normal, and what the queue is for.

Score distribution and the two cutoffs

The histogram shows how many pairs fall at each match weight from 0 to 1, in twenty bars. Two draggable cutoffs sit on it.

Reading the shape

A healthy distribution usually has two humps and a valley — a mass of weak coincidental matches at the low end, a mass of real duplicates near the top, and a thin middle. Put your cutoffs in the valley. If there’s no valley, the two populations aren’t separable by score alone, and the work is genuinely per-pair.

Dragging is free; saving is not

Dragging a cutoff recomputes every number on the page instantly and costs no round trip — the histogram data needed for the arithmetic is already in your browser. While you’re dragging, the page reads Unsaved — this view only.

Saving the cutoffs is a different act. It writes them into the matcher’s configuration and re-scores the whole corpus, which changes what gets stored as a duplicate everywhere in the product. Explore freely; save deliberately.

Work by pattern

The pattern list is the actual work queue. Every match is filed under the reason it matched, which is what makes a corpus of 18,000 pairs navigable — it’s usually five or six patterns.

Each row shows:

How the list is ordered

Not by size, and not by score. Both of those sort badly:

• By volume, unfixable noise floats to the top.
• By score, the easy work ends up last.

So patterns are ranked by how much work one decision settles — the number of undecided pairs multiplied by how far a single decision reaches (pairs per cluster), lifted for patterns holding matches lineage can’t explain. Work top-down and you clear the most backlog per decision.

Why the cluster count is an estimate. A cluster can have pairs in several score bins, so the bins can’t simply be added without counting it twice. The number is an upper bound and the app labels it as one rather than quietly presenting it as exact.

misc and +N

Two bits of housekeeping you’ll see in pattern names:

• +N — a pattern key names at most three labels, then collapses the rest. email+person+phone+2 matched on five labels. The full set is on the pattern page.
• misc / Everything else — patterns with only a handful of pairs are folded into a remainder bucket per family. A pattern is something you write one rule for; a “pattern” of two pairs is just two pairs.

Where duplicates come from

A small graph in the instrument rail whose nodes are systems, not assets. This distinction is the whole point of the tile:

• Concentrated between two systems → a real, findable cause. Usually an integration that re-imports instead of updating, or two pipelines writing the same thing.
• Concentrated inside one system → that system’s own ingestion or deduplication, not an integration.
• Spread evenly across everything → more often points at the matcher than at the data. Go to Tuning rather than the queue.

Click a system to narrow the entire page — headline, patterns, histogram — to pairs touching it. Click several to ask “these two systems”. Touching means either side counts: restricting to pairs wholly inside a selection would hide exactly the cross-system pairs the tile is pointing at.

Reset clears the filter.

Empty queue?

Next: Patterns — what a pattern is and how to settle one.