Skip to content

Reviewing Contradictions

When your organization's documents disagree with each other — for example, two sources describe the same procedure differently or define a term in conflicting ways — Mention flags these as contradictions. While a contradiction waits for review, Mention already prefers the side stated more recently and hides the older one from the AI's context; an admin's decision then settles the matter for good, so every article Mention writes and every answer it gives becomes consistent with your call.

A contradiction belongs to your sources, not to any one Audience: the same knowledge base serves all of them, so resolving one improves everything at once. Conflicts between sources are settled here, by an admin. The exception is a misunderstanding — a member said something on a call or in chat that conflicts with what your documents say. There is a person on one side of that, so it is routed to the Audience whose owners can tell whether the member got it wrong or the document went stale, and they handle it on that Audience's Questions tab. See Reviewing Misunderstandings.

Where to find it

  • Web — Open Sources from the admin sidebar, then the Contradictions tile under the Review section at the bottom of the dashboard, which shows how many are waiting. By default it lists only the pending conflicts routed to no Audience, worst first — the ones waiting on an admin. Switch to All conflicts for oversight of everything, including what Audience owners have pending; a side stated by a member shows who said it, and you can still decide a routed conflict yourself.
  • Mobile — Admins get a Conflicts tab listing every pending conflict, with a badge showing the number of pending items.

Both surfaces work the same queue on the same backend, so a decision made on one shows up immediately on the other.

What is a Contradiction?

A Contradiction is a pair of facts, extracted from your source documents, that Mention's AI has judged to be logically in conflict. Each contradiction includes:

  • The two conflicting facts — shown side by side with references to the source documents they came from
  • A severity score (1–10) — how strongly the two facts conflict; it orders the queue rather than appearing on the page
  • What surfaced it — the query Mention was answering, or the meeting claim it was checking, when it noticed the disagreement; recorded but not shown, and deciding which fact is right does not depend on it

Contradictions surface when two facts are considered together. Two unrelated documents making different claims about different topics are never brought into the same set, so they are not flagged.

How detection works

Detection runs in two places.

While Mention works. Every time Mention writes an article or answers a question, it gathers the facts that answer that particular query — and checks that set for logical conflict before writing.

When a meeting transcript arrives. Once Mention has extracted the facts from a connected Google Meet or Teams transcript, it looks up what your documents already say about each claim and checks the two against each other. This does not wait for anyone to ask a question, so something said in a meeting that disagrees with a written source is flagged within the normal ingestion window rather than whenever the topic next comes up.

Both paths go through the same second verification step: a stronger model re-checks each candidate pair and keeps only genuine contradictions, filtering out facts that merely describe different aspects of the same thing, restate each other in different words, or where one is a more specific case of the other. Only pairs that pass this check are stored for you to review.

A pair of facts that is never considered together is not competing to say anything, so it is not surfaced. Each contradiction records what surfaced it, and if the same pair turns up again, Mention records each occurrence — a pair with a long list attached is disrupting more of your content than a pair with one.

Mention remembers which pairs it has already examined, so rebuilding the same content repeatedly does not re-check pairs it has already cleared.

Detection consumes credits. See Understanding Credits for details.

Resolving a conflict

Each conflict is one row with the two facts side by side — the competing claim on the left, the competing claim on the right, each with the sources behind it and when it was stated. The side Mention is already preferring arrives pre-selected, so submitting without changes simply confirms it. Between two sources that is the one stated more recently, and when both were stated at the same moment nothing is pre-selected — there is no ground to prefer either. Where a member is on one side, the document is pre-selected however old it is. Work straight down the list:

  • Pick a side — click the fact you believe is correct, or keep the pre-selected pick. Clicking a selected fact clears it.
  • Dismiss — the × at the end of the row marks the conflict a false positive; both facts stay in use.

Decisions stage rather than commit. Work through as many as you like, then submit them together with the N decisions ready button at the bottom of the page. Nothing changes until you submit, so you can revise a pick or back out entirely.

Auto-resolve

Auto-resolve stages a decision for every unrouted conflict it can confidently rank, leaving the rest for you. It never commits on its own — you review the staged picks and submit them. Choose the rule that matches how you trust your sources:

  • Newest statement wins — keep the fact stated most recently, on the theory that newer information is more accurate. This is recency and nothing else, so where a member is on one side it can stage the opposite of the pre-selected pick.
  • Prefer docs — keep the fact backed by written documentation.
  • Prefer chat — keep the fact from conversations, transcripts, and calls.

Under the docs and chat rules, a side backed by a Mention-authored source — a feature brief — wins outright, and conflicts a rule can't confidently rank (an exact tie, or both sides matching) are left untouched for you to decide. Picks you have already made by hand are never overwritten.

What resolution does

When you resolve a contradiction by picking a winning fact, the losing fact is invalidated — retired from the knowledge base entirely, not just hidden. It stops appearing in every future article and answer. This means:

  • Articles become more accurate because conflicting information is removed
  • Members see consistent answers grounded in the version you chose as correct

A pending contradiction does not leave both facts in play: Mention already prefers one side and hides the other from the AI's context — the more recently stated side between two sources, and the document where a member is on one side. Your decision overrules that default: picking the hidden side restores it and retires the other. Only a tie between two sources (both stated at the same moment) or a dismissal keeps both facts in use.

The other half: gaps

Contradictions are what your sources get wrong. Unanswered questions are what they are missing — questions members asked that Mention could not answer.

Those are not org-wide. A gap belongs to the Audience it was asked in, and is closed by the people who own that area, on that Audience's Questions tab. See Answering Unanswered Questions.

Routing: misunderstandings go to an Audience owner

When one side is something a named member said — on a call, or in Slack or Teams chat — and the other is a document, Mention calls it a misunderstanding and routes it to the single Audience whose owners are best placed to settle it, normally one that member belongs to. They review it on that Audience's Questions tab. An owner resolving a routed conflict is deciding org-wide truth: the losing fact is retired for everyone, exactly as if an admin had decided it.

Everything else stays here. Two documents disagreeing — or two members disagreeing with each other — is a question about the sources rather than about one person, so there is no owner closer to the answer than an admin. The same is true of a misunderstanding whose speaker belongs to no Audience Mention can identify. That is what the default view shows; routed conflicts remain visible to admins under the All conflicts filter. See Reviewing Misunderstandings for the owner's side.

Contradictions vs. a table of contents

Resolving contradictions is about what Mention knows: it makes the shared knowledge more accurate, across every Audience.

An Audience's table of contents is about what one group reads: which topics get an article, from prompts its owner wrote. Resolving contradictions makes the knowledge better; the table of contents decides what is made from it. See Curating Contents.