Decisions

A queue on its own is write-only: you judge a pair and it disappears. That makes the judgement worthless five minutes later — you can’t check it, change your mind, or act on it.

Decisions is the other half.

Column	Meaning
Pair	Both asset names. Click to reopen the pair screen.
Decision	Duplicate · Not a duplicate · Unsure · Split
Score	The match weight you saw when you decided, and the score now
Pattern	Which pattern it belonged to at the time
Used in	The case or inquiry it went to — or not used yet
Decided	When, and by you or by an agent

A verdict stores the score it was made against. If the pair has since been materially re-scored — because weights changed, an exclusion was added, or new findings appeared — the row is marked Re-scored since you decided. Worth another look.

This is the honest handling of a real problem: the standing judgement was made about a different number. It isn’t discarded (that would throw away real work), but it isn’t presented as current either.

You, or an agent

Agent decisions are stamped and counted separately, everywhere. An agent clearing the safe band must never make the queue look like a person worked through it. See AI Agents & Duplicates.

Went nowhere yet

The filter that earns this screen its place.

Went nowhere yet = pairs confirmed as duplicates that were never taken into a case or an inquiry.

This is the common outcome, and a deliberate one — during a review pass you’re judging, not investigating. But those confirmed duplicates are ready-made evidence: someone already looked at each one and said yes. Coming back to this list a day later is the cheapest way to start a real investigation.

Acting on decisions in bulk

Select any number of rows, then:

Action	What it does
Add to case	Every selected pair’s assets become evidence in a new or existing case. Optionally attaches their findings too.
Open inquiry	Creates one inquiry watching for the signature these pairs share — the labels that matched, scoped to the sources they came from.
Reopen	Removes the verdicts and returns the pairs to the queue.

Once a pair goes to a case or inquiry, its row shows the destination, so Used in doubles as a record of follow-through.

Reopen, and what it undoes

Reopen removes the verdict. For Not a duplicate and Split it also un-suppresses the pair, and the neighbourhood is re-clustered immediately — so those two assets can be joined again if the evidence still supports it.

Reopening a Confirm or an Unsure only puts the pair back in the queue; neither of those changed clustering in the first place.

The undo log

Reachable from the pattern screen. Every action — a single verdict, a bulk confirm, a split, an exclusion — is written as one entry with what it touched, so the log reads as a list of decisions rather than a list of 840 rows.

Undoing an entry removes its verdicts. For a split it also re-derives the link from the evidence as it stands now and re-clusters — which is the correct behaviour, not a restore of an old snapshot: if the evidence has changed and the pair should no longer be linked at all, replaying the old link would be wrong. For an exclusion it removes the rule that was written.

Undo is not unbounded time travel. Once the review index has been rebuilt — after any scan or recompute — the pairs a batch referred to may have been re-scored or re-clustered, and replaying the batch might not restore what you saw. Those entries are shown greyed out rather than offered and then failing quietly. Undo within the same working session.

For anything older, Reopen from this screen is the reliable route: it works off the current state rather than trying to replay a past one.

Which verdicts change what

Verdict	Records a judgement	Stops later scans re-joining the pair	Re-clusters immediately
Confirm	Yes	No — it agrees with the engine	No
Not a duplicate	Yes	Yes	On reopen
Split	Yes	Yes	Yes
Unsure	Yes	No — explicitly not a decision	No

This is why unsure is safe to use freely, and why not a duplicate and split are the two that genuinely change the shape of your data.