What WikiFix catches — and what it doesn't yet

These are the problems we know about, and where each one stands.

If the thing costing you isn’t here, tell us — that’s how the next one gets picked.

Every finding shows you the exact conflicting statements, side by side, quoted verbatim and linked to where each one is said. You’re never asked to trust a score — you see what the pages actually say and decide for yourself.

Findings are consistent. The same conflicting statements produce the same finding every time — nothing about detection is left to chance. A finding you resolved won’t quietly reappear, and one you’ve already fixed stays fixed until something on the page actually changes.

Duplicate pages and departed owners work differently, and are steadier as a result. Duplicates are matched by comparing the facts already extracted from each page, so the same set of pages always groups the same way. Departed owners reads your site’s user status rather than page text, so it re-checks every page in scope on every scan — a page nobody has touched can still acquire a departed owner between scans.

• Disagreements stated directly on both pages — the most common real-world shape of a contradiction, and where WikiFix is strongest.
• Numerical, temporal, and categorical disagreements — two different dates, counts, or classifications for the same thing.
• Yes/no and status conflicts — one page says a policy applies, another says it doesn’t.

• Contradictions that require reasoning to spot — where the disagreement follows from what two pages say rather than being stated outright. This is the direct cost of never inferring: by design, not by accident.
• Near-miss numerical differences — “$50/month” vs. “$50.00 per month,” or values that differ by a rounding margin rather than a real disagreement. There’s no unit or precision model yet, so this can go either way: a miss, or an occasional false flag.
• Facts phrased differently enough that WikiFix doesn’t yet recognize them as answers to the same question. Both pages read correctly, the link between them just isn’t made. This is the single biggest thing we’re working on.

We test against public, independently-labeled contradiction benchmarks rather than grading our own homework blind — seeded into a real Confluence space and run through the same production pipeline a customer gets. Our primary published benchmark today is WikiContradict (IBM Research): 333 pages, 124 known contradiction pairs, 42 pairs that agree. We’re expanding coverage to additional public corpora as we go.

On that benchmark: WikiFix currently surfaces roughly 1 in 5 of the labeled contradictions, with zero false alarms across the pairs that agree. We hold ourselves to a strict standard when scoring this — a pair only counts as caught if WikiFix produced one finding that names both sides, not “flagged something nearby.”

Why we don’t quote a number against other systems’ benchmarks. Published research systems in this space (like Stanford’s CLAIRE, evaluated against Wikipedia at large) score a different question — “does a contradiction exist anywhere in a huge corpus” rather than “did we surface this specific pair” — on a different kind of corpus, with different metrics. Comparing headline numbers across those setups isn’t meaningful, so we don’t.

One thing that research is genuinely useful for: the best available public estimate of what real-world contradictions look like when you sample them naturally, rather than construct them for a benchmark. On that evidence, the reasoning-heavy contradictions WikiFix declines to infer are a small share of what actually occurs day to day — most real disagreements are the direct, stated kind WikiFix is built to catch.

Your content is handled the same way whether we’re running a public benchmark or your own Confluence: it never leaves the EU.

Scanning runs on AWS Bedrock in Frankfurt (eu-central-1), on OpenAI’s open-weight gpt-oss-120b by default — served in-region, not a live call out to a third party. As of August 2026, there’s no cross-region routing available for that model in any region, so there’s no mechanism by which a Frankfurt call could be served elsewhere. Our own deployment also refuses to start if it’s configured with any model identifier that would route outside the EU — checked at deploy time and again when config is parsed.

Your content is never used to train AI models. We keep a derived index and the short excerpts behind each finding, not full copies of your pages. Jurisdiction, encryption, subprocessors, and what we don’t have yet are all on the security page.

• How it works — the scan, findings, and apply loop
• Security & data — what we store and where
• FAQ — the questions we expect