Reviewing a Pair

The screen where the decision actually gets made. It shows two assets, the evidence for and against them being the same thing, and five ways to act.

Match weight

The large number at the top.

Match weight is the share of the two assets’ available weighted evidence that actually matched, from 0 to 1.

It’s not a percentage of similarity and it’s not a confidence. It’s a ratio: of all the weighted values these two assets have between them, how much found a partner on the other side?

That framing matters when you compare two pairs:

• A pair at 0.90 has almost nothing that doesn’t match.
• A pair at 0.45 matched something real, but half its evidence disagrees.
• A pair at 0.45 made of one shared IBAN is a much stronger claim than a pair at 0.45 made of six shared country codes — which is exactly what the breakdown below is for.

Next to it sits the lineage chip (shared upstream / no path / unknown) and, if the pair was already judged, the standing verdict — plus a note if it has been re-scored since, which means the judgement was made about a different number and is worth another look.

Why it scored n — the breakdown

A waterfall chart, one bar per label. Two things make it trustworthy:

The bars add up to the number above them. Nothing is hidden in a blend. If the total didn’t match the headline, that’s the one thing this screen must never do, so the arithmetic is built to make it impossible.

Evidence against is inside the sum, not omitted. Each label shows:

A label sitting on only one of the two assets produces a full positive potential bar and an equal negative one. That is the honest way to show disagreement: as weight that was available and went unclaimed, in the same units as the weight that matched.

A dashed Perfect match reference line marks the total available weight, normally 1. If it sits somewhere else, the label profiles and the scorer have drifted apart (usually mid-recompute) and the line moves visibly rather than the discrepancy being absorbed in silence.

For similar spelling pairs the bars are sums of spelling-similarity scores rather than counts of whole shared values, and the chart says so.

“This pair was scored before the breakdown was recorded.” An older pair from before this data existed. It reappears after the next scan. The score itself is still valid.

The values behind this match

The table underneath: the actual values, side by side.

This is the evidence. Everything above it is a summary of this table.

Two habits worth building:

• Read the shared values before the score. One shared account number beats five shared common words, whatever the arithmetic says.
• Use “Where else this value appears”. A value that shows up in four hundred assets isn’t identifying this pair — it’s boilerplate, and the right response is an exclusion, not a verdict.

This cluster — the neighbourhood graph

A small graph of the cluster these two sit in, with both under review highlighted. It exists to answer one question: is this pair holding together something that shouldn’t be together?

The cut point is marked: the weakest link whose removal would split the cluster in two. Its score tells you how real the split would be — a low score on the cut point means one weak match is chaining two groups, and cutting it is obvious; a high score means it’s a genuine judgement call.

If no single link holds the cluster together, there’s nothing to cut, and the Split action is disabled with a note saying so.

The five actions

Confirm

These two are the same thing. Recorded as a duplicate, reversible from the undo log, and listed in Decisions where you can push it into a case or an inquiry.

Confirm doesn’t merge or delete anything. Classifyre records judgements; it never edits your source systems.

Not a duplicate

These two are different things. The negative verdict, sitting right next to the positive one — because without it the only way to disagree was “unsure”, which means something different and produces a worse record.

Rejecting also suppresses the pair: later scans will not rejoin these two into a cluster.

Choosing it opens What made these match, which is the point. A rejection on its own changes nothing about the matcher — the next scan produces the same pair from the same evidence. So the dialog names the label carrying most of the score, shows the values that matched, and tells you how many other pairs the same combination produced. That last number is what turns one rejection into a decision worth taking: it’s the difference between dismissing one bad match and stopping four hundred.

It then offers three ways out:

The two fixes are only offered when one label genuinely dominates. On an even split there’s nothing to single out, and suggesting a weight change would be advice to break the labels that were doing their job.

Split the cluster here

These two shouldn’t be in the same cluster. Cuts the link and re-clusters the neighbourhood immediately.

Available only when there is a single link to cut — that is, when the cluster graph found a cut point. If the two are held together through several routes, cutting one edge changes nothing, and the button is disabled rather than pretending otherwise.

The verdict is what makes it stick: cluster building consults recorded splits on every later pass, so the next scan cannot quietly rejoin what you separated.

Afterwards the app tells you what actually happened. Two assets inside a larger cluster can remain joined through a third member — and saying “split” when they’re still together would be a lie you’d only discover later.

Split vs. Not a duplicate. Not a duplicate is a statement about the two assets. Split is a statement about the cluster’s shape — usually what you want on a chain whose ends drifted apart.

Unsure, next

I can’t tell. A first-class button, not a way out.

Forcing a binary decision on a genuinely ambiguous pair produces bad records, and the fix costs more than the decision saved. So “unsure” is a real verdict that routes the pair to a second look.

It has a second job: the count of unsure verdicts is a signal about your cutoffs. A pile of them means the review band is sitting where the evidence doesn’t separate, and the answer is on the tuning screen, not in the queue.

Unlike not a duplicate and split, unsure does not suppress anything — it’s explicitly not a decision about the assets.

Add to case / Create inquiry

Where a confirmed duplicate goes next.

• Add to case — both assets (optionally with their findings) become evidence in a new or existing case, with the normal case activity trail.
• Create inquiry — opens an inquiry pre-filled from this match: the labels that made it match, scoped to the sources it came from. It keeps watching for the same signature instead of starting from a blank form.

The queue advances

Every verdict moves you to the next undecided pair in the same pattern, strongest first, under the same cutoffs and lineage filter you were using. When the pattern is clear you’re returned to it with a note.

The subtitle under the score shows how many are left in this pattern, so the queue has a visible end.

Decisions are applied optimistically: you’re already reading the next pair by the time the write lands. If a write fails you’re told and the pair comes back — nothing is lost quietly.

Keyboard shortcuts

The whole point of a queue is throughput, and at a round trip per verdict you feel every one of them. The shortcuts map one-to-one onto the buttons in the action bar — nothing is hidden behind a key that isn’t also visible on screen.

They’re scoped to the pair screen only, and they’re inert while you’re typing in a field or when a modifier key is held. There are no shortcuts on the list screens, deliberately: a legend advertising keys that do nothing is worse than no legend.

A comfortable pace with these is roughly a pair every few seconds on clear-cut matches. If you’re spending a minute per pair, that’s a signal — go back to the pattern and look for the rule instead.

Next: Decisions — what happens to everything you just judged.