Evidence & relationships

Every artefact in a register carries the trail of where it came from and how it relates to other items. This page covers three things that cut across all the registers: the evidence chain, typed relationships, and duplicate review.

The evidence chain

Open any item — a risk, decision, action, and so on — and its detail page shows an Evidence panel: the communications the item was extracted from. For each source you see the subject, who it was from, when it was sent, the classifier’s confidence, and the exact quote (the claim excerpt) that supports the item:

“FRP pan failing in guestrooms”

If an item came from an attachment rather than an email body, the panel shows a Source: filename link that opens the original file. This is how you verify a fact — you can read the text it was lifted from.

Caution

A source can show a fabricated excerpt warning. That means the quote the classifier produced doesn’t appear verbatim in the original message. It’s a flag on the quote, not on the item itself — confidence in the underlying artefact is unaffected — so treat the excerpt with suspicion and check the source.

You only see sources you are allowed to open, so two people can see different evidence on the same item. An empty panel means there is no source here that you can open. The same goes the other way: a message or file lists only the items extracted from it that you can open, and a count of sources (such as N evidence links in duplicate review) counts only the ones you can open.

Relationships

Artefacts can be linked to each other with a typed relation (see the Glossary):

Relation	Meaning
supersedes	A newer item replaces an older one (directed, newer → older)
duplicate	The same thing said twice
contradicts	Opposite positions — a sign the project reversed course
relates	Connected, but none of the above

The system works these out automatically across decisions, risks, issues, actions and meetings — you don’t review them one at a time. Relations show up as badges on register items (for example ↺ 3 prior, ⧉ 2 duplicates, ⚠ 1 contradicts), and they drive each item’s effective register state: the register shows the live head of each chain, and a superseded item is dimmed.

You only see relations to items you are allowed to open. A badge counts, an item’s relationship panel lists, and chat mentions only the linked items your clearance lets you see, so two people can see different counts on the same item. An item replaced by something you cannot see still shows as superseded.

The relations review

The Relationships page lists relations the system was least confident about under Needs attention (switch to All to browse everything by type). On each one you can re-type the relation or confirm it. Direction for supersedes comes from the items’ dates, so a wrong direction usually means a wrong date.

Your changes are staged until you click Re-resolve registers, which rebuilds the current/superseded state from the reviewed relations. An override you make sticks even when the classifier re-runs.

Duplicate review (dedup)

Use the Duplicate review page to find and merge artefacts that describe the same thing. Click Run dedup and the system:

1. Finds candidate pairs by similarity. LLM-gated mode (the default) reads each pair and catches paraphrased duplicates; Cosine mode is a cheaper geometric match that can miss reworded ones.
2. Judges each pair as same, different, or unsure, with its reasoning.
3. Groups the same pairs into clusters.
4. Merges each cluster into one entry — keeping the most-substantiated row as the winner, re-pointing evidence links to it, marking the others superseded, and updating the knowledge graph.

The work runs server-side, so it keeps going even if you close the tab. A progress panel shows the live log.

Run dedup and every other button here that works on many pairs at once (judging, re-judging, resetting verdicts, auto-merging, auto-linking, re-resolving) work only on records you can see. A pair is found, judged or merged only if you can see both records. A person is linked, or a chain re-resolved, only if you can see every record involved. Anything else is left exactly as it is for someone who can see it. A workspace admin’s run covers everything. See Operations.

Pending clusters appear under Pending (clusters). Tick the clusters you want, optionally change the winner with the radio button, and Approve to merge in a batch. You can also tick Auto-merge confident clusters to skip the review for clear-cut ones. Pairs the system judged different or unsure can be accepted as kept separate so they stop reappearing on every scan. The Merged and Kept separate tabs are a read-only history of past decisions.

Note

Auto-merge only ever fires for stakeholders — the same person under two records genuinely is one identity, so the rows are combined and their contact details and roles are unioned onto the winner. Event-type items (decisions, risks, issues, actions, meetings) are never auto-merged; the system steers them toward a relation instead so the project’s history isn’t lost. You can still merge an event cluster by hand from this page if you’re sure they’re the same.