Every Number Explained

A reference for every number and phrase you’ll meet in duplicate review. If you landed here from a screen, use the sidebar or search for the exact words you saw.

The core objects

Matching pair

Two assets the engine linked because they look like the same thing. A pair is one decision — it’s the unit everything else is counted in.

Pairs have no direction: “A and B” and “B and A” are the same pair, and a verdict recorded from either side applies to both.

A set of assets joined transitively by strong matches, treated as one thing. If A matches B and B matches C strongly enough, all three are one cluster — even if A and C never matched each other directly.

Every asset belongs to at most one cluster. Clusters are what splitting acts on, and the reason cluster shape is worth checking: one weak link can chain two unrelated groups.

Pattern

A group of pairs that matched for the same reason — the same set of value labels, from the same engine. email + person is a pattern. Patterns are what make a corpus navigable: 18,000 pairs is usually five or six of them.

Verdict

Your recorded judgement about one pair: Duplicate, Not a duplicate, Unsure, or Split. Verdicts are keyed by asset pair, so they survive re-scoring, renaming and re-clustering.

Numbers on the queue screen

Pairs remaining

Undecided matched pairs inside your review band. The headline number. Moves when you move a cutoff. Decisions taken by an agent are reported separately so this stays a count of human work.

Duplicate rate

Assets appearing in at least one matched pair, over all assets in the workspace.

A cluster count means nothing without a denominator — “4,100 clusters” is meaningless until you know whether it’s out of nine thousand assets or two million. This is the number to quote to someone who didn’t run the scan.

Above your merge cutoff · Needs your review · Below your review cutoff

The three bands, split by your two cutoffs. See Where the matches stand.

review cutoff

The lower line on the histogram. Below it, matches are treated as rejected — not worth looking at.

Saved, it becomes the related threshold: the minimum weighted match at which a link is recorded at all.

merge cutoff

The upper line. At or above it, matches are strong enough not to need a person.

Saved, it becomes the duplicate threshold: the minimum match at which two assets are treated as the same thing. These are the links that drive clustering, so moving it changes cluster membership, not just what you review.

Unsaved — this view only

You’ve dragged a cutoff but not saved. Every number on the page reflects the new position; nothing has been written and nothing has been re-scored. Saving re-scores the whole corpus.

The proportion of all matched pairs running between a given pair of systems. A high share on one pairing means the duplicates have one findable cause. An even spread across everything more often points at the matcher than at the data.

Numbers on a pattern

left

Undecided pairs in this pattern, inside the current band. Not the pattern’s total size.

clusters

How many clusters this pattern’s pairs touch. An upper bound: a cluster can have pairs in several score bins, and the bins can’t be added without counting it twice. The app labels it as an estimate rather than presenting it as exact.

avg

The average match weight across the pattern. A high average with a cutoff candidate label is the signature of a pattern you should settle with a threshold rather than by hand.

pairs / clusters / assets (the header figures)

What a bulk action would touch right now, under your current cutoffs and filters. Live — they follow the cutoffs.

After / Before

Total pairs remaining across the whole workspace, before and after this bulk action. The honest measure of leverage: a pattern that takes 8,400 off a backlog of 12,000 is worth doing first.

showing n of m pairs

For repeated text patterns only. Near-duplicate groups are found at the finding level and projected onto asset pairs, which is quadratic — a 50-asset group is 1,225 pairs — so the projection is capped. The row shows the capped count next to the true size, rather than silently printing a truncated number.

`+N` in a pattern name

A pattern key names at most three labels and 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, 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.

Kinds of pattern

Shared values

Matched on the concrete values inside findings — emails, account numbers, IBANs, names. The main engine.

Similar spelling

Matched phonetically: name-like values that sound the same, spelled differently — Jon Smyth and John Smith. Only applied to name-like labels, and only above a spelling-similarity floor. These are the ones most worth spot-checking as a group.

Identical content

The same bytes, twice. A stronger and differently-derived claim than value overlap, so it’s kept separate.

Repeated text

Matched by meaning rather than by literal values, using embeddings. This is what catches boilerplate — the confidentiality notice on four hundred contracts. Needs a configured embedding model; without one, this family is simply absent.

What kind of decision is this?

no judgement needed

Byte-identical content. Confirming is bookkeeping. Confirm the lot.

rule candidate

Shared boilerplate is doing the matching. Don’t grind the pairs — press Stop matching on these values, which excludes the values inside the template so it stops driving matches everywhere else. See Patterns.

cutoff candidate

These labels match closely and consistently — average weight at or above 0.85. So where you draw the line is the only real question. Go to the histogram, set the merge cutoff above this pattern’s mass, and save. Every pair above it stops needing a person, permanently.

needs judgement

The labels overlap but the rest of the evidence doesn’t agree. No threshold separates these, so they’re genuinely per-pair. Sample a dozen first: if they’re all the same mistake, the real fix is a weight change or an exclusion.

Cluster shapes

Shape	Meaning	Read it as
pair	Two assets	The simple case
clique	Every member matches every other	Genuinely one thing
chain	Matches form a line; the ends don’t match	Suspicious — transitivity dragged in unrelated members
partial	Some members match, others don’t	Mixed; worth opening
mixed	(pattern level) Its clusters aren’t all one shape	No signal either way

Lineage phrases

Shared upstream

One of these assets derives from the other, or both come from the same place. A derived copy. Expected, not a problem — and the reason most metadata-only duplicate tools get ignored is that they report this as if it were one.

No path — both sides have lineage

“Two teams appear to have built the same thing independently. This is the case worth chasing.”

The most valuable cell in the product. Both assets have lineage — so we’re not guessing from missing data — and nothing connects them, yet they look nearly identical. Somebody rebuilt something that already existed, and nobody knows.

It’s expensive (two pipelines, two sets of maintenance, two chances to diverge) and invisible (neither team has a reason to look). Patterns with a fifth or more of these get the alarm colour, and an agent is never allowed to clear them.

Lineage unknown

We have no lineage for at least one of these assets. A coverage gap, not evidence either way. Judge on the values.

If everything reads unknown, check whether your sources produce lineage at all — see Lineage & Relationships.

Lineage is too interconnected to tell assets apart

A safety valve. The similarity/lineage test approximates “is there a path between these two” with “are they in the same connected component”. When one giant component swallows most of your lineage graph, that approximation makes everything look derived and nothing would ever be escalated. Past a threshold Classifyre reports unknown instead, which is honest about not being able to tell.

On the pair screen

Match weight

The share of the two assets’ available weighted evidence that actually matched, 0 to 1. Not a percentage of similarity and not a confidence. See Reviewing a Pair.

Why it scored n

The waterfall breakdown, one bar per label. The bars add up to the number above them — nothing is hidden in a blend.

for / against

Per label: what it contributed to the score, and the gap between that and what it could have contributed. A label present on only one asset produces a full positive potential and an equal negative — that’s the evidence against, shown in the same units as the evidence for.

Perfect match = n

The reference line: the total weight that was available. Normally 1. If it sits elsewhere, the label profiles and the scorer have drifted (usually mid-recompute) and the line moves visibly rather than the discrepancy being hidden.

weight n × m values

The tooltip on a bar: this label’s weight, times how many values it matched.

The values behind this match

The table of actual shared values. This is the evidence — everything above it is a summary of it. Read it before the score.

Where else this value appears

A reverse lookup. A value present in four hundred assets isn’t identifying this pair — it’s furniture, and the right response is an exclusion, not a verdict.

cut point

In the cluster graph: the weakest link whose removal would split the cluster in two. Its score says how real the split would be. A low score means one weak match is chaining two groups; a high score means splitting is a genuine judgement call.

No single link holds this cluster together

There’s no cut point, so cutting one edge would separate nothing. Split is disabled rather than pretending.

Already decided · re-scored since

This pair has a standing verdict, and the score has moved materially since. The judgement was made about a different number — worth another look.

Actions

Confirm

These two are the same thing. Records a duplicate. Doesn’t merge or delete anything — Classifyre records judgements and never writes back to your systems.

Not a duplicate

These two are different things. Also suppresses the pair: later scans will not rejoin them into a cluster. Opens the cause dialog so you can fix what caused the match, not just dismiss it.

Split the cluster here

These two shouldn’t be in the same cluster. Cuts the link and re-clusters immediately. Available only when a single link holds the cluster together. The verdict makes it stick across future scans.

Afterwards you’re told whether they actually ended up apart — two assets inside a larger cluster can stay joined through a third member.

Unsure, next

I can’t tell. A real verdict, not a way out. Forcing a binary on an ambiguous pair produces bad records. It suppresses nothing.

A pile of these is a signal that your review band sits where the evidence doesn’t separate — the fix is on the tuning screen.

Confirm all in band

Records Confirm on every undecided pair in the pattern, inside the current cutoffs and lineage filter. Never touches an already-decided pair. Reversible from the undo log.

Stop matching on these values

On a boilerplate pattern only. Writes an exclusion rule for each value inside the repeated passage, and records Not a duplicate on the pattern’s own pairs. The values are listed with how many assets hold each one, so you approve a list rather than a promise. One undo-log entry; undoing it removes every rule.