Tim Schartman

Documentation

Crosscheck for Confluence

Getting started

1. Install Crosscheck from the Atlassian Marketplace. A site administrator must approve the read-only permissions.
2. Open Apps → Crosscheck from the Confluence top navigation. Crosscheck is a global page, so it is not in any individual space's sidebar.
3. Choose a space from the dropdown and select Scan this space.

A scan runs in the background across many short steps. Progress advances in visible stages rather than smoothly. A few hundred pages typically complete in a minute or two; larger spaces take longer, dominated by the contradiction-checking stage.

Reading the results

Contradictions

Pairs of pages making claims that cannot both be true about the same subject. Each finding shows both page titles, the conflicting statements quoted verbatim, a one-sentence explanation, and a confidence score.

Findings are deliberately conservative. Two pages stating different numbers about different subjects — staging versus production, or a documented exception to a rule — are not contradictions and are filtered out before analysis. The intent is that a reported finding is worth your attention.

Duplicates

Pages that overlap substantially in meaning, with a similarity percentage. This is semantic rather than title-based: two pages describing the same process in different words still register. Pages with almost no content are never compared, since anything empty resembles anything else empty.

Needs attention

A prioritised remediation queue, worst first. Only pages scoring below 80 appear — listing everything would make the tab meaningless.

The score starts at 100 and deducts for:

The inline page panel

Near the byline of any scanned page, Crosscheck shows whether that page conflicts with another, which pages duplicate it, and its health score with reasons. This is where most people encounter a finding in day-to-day use.

How detection works

Comparing every page against every other page is quadratic — a 5,000-page space is 12.5 million pairs — so Crosscheck narrows the field before spending anything on analysis.

1. Signatures. Each page is reduced to a compact numerical fingerprint. Pages that look alike are identified without comparing full text.
2. Claim extraction. Concrete values — durations, percentages, currency amounts, versions — are extracted with their surrounding context.
3. Candidate selection. Pairs are proposed two ways: pages that look alike, and pages that claim different things about the same subject. The second catches paraphrased conflicts that similarity alone misses entirely.
4. Confirmation. Only surviving candidates are analysed by an Atlassian-hosted language model, in two stages, and only high-confidence results are reported.

Steps 1 to 3 use no AI at all, which is why the free tier can offer duplicate and health analysis at no cost.

Scans and scheduling

• Manual scans run whenever you select Scan this space.
• Nightly rescans run automatically for spaces you have scanned at least once. A space is never scanned automatically until you have scanned it yourself.
• Reset clears all stored analysis for a space and allows a fresh scan.

Plans and limits

Free plans (up to 10 users) include duplicate detection, orphan detection, staleness and health scoring. No language model is called on a free plan, so no AI cost is incurred by anyone.

Paid plans add contradiction detection. Each site has a monthly AI budget that scales with seat count. When a scan reaches the budget it stops rather than continuing to bill, and the dashboard reports how many candidate pairs were checked out of how many were found. Budgets reset monthly.

What Crosscheck does not do

• Modify, create or delete any Confluence content. It requests read-only access only.
• Scan personal spaces. These are excluded entirely.
• Send content outside Atlassian. See Security.
• Detect every contradiction. Conflicts expressed without concrete values, or spread across many pages rather than stated in two, may not be found.