Understanding Verification Results
A guided tour of Fast Verify vs Deep Verify, every Assessment outcome (Authentic, Authentic with Notice, Unsure, Inauthentic with/without replacements, Invalid, Error, Exceeded), confidence scoring, and the stats bar at the top of the results list.
Fast Verify
The default first pass — runs every citation through a quick metadata + baseline-database check. Cheap, fast, but limited to citations whose details closely match a public record.
How it works
- Parses each citation into title / authors / year / DOI / URL, etc.
- Queries baseline academic databases and matches by exact metadata.
- Costs 1 credit per citation. If the answer is uncertain, click Deep Verify to escalate to a stronger search.
Deep Verify
A heavy-duty second pass that broadens database coverage and brings in the most capable AI models. Use it when the fast check comes back uncertain, or when the citation is too important to leave to a quick check.
How it works
- Queries a wider set of academic databases, including less common sources most tools skip.
- Runs the strongest frontier AI models for matching and disambiguation — higher recall and stronger reasoning than fast verify.
- The AI agent derives multiple distinct keyword combinations to query academic databases, rather than relying only on the citation text itself.
- Runs on-demand per citation and costs an extra x5 credits; recommended after fast verify when the result is unclear.
Confidence Scores
Citations that resolve to a verdict (Authentic or Unsure) include a confidence score showing how closely the matched paper aligns with the citation.
Confirmed match — citation found and fully verified.
Found a candidate but couldn't fully confirm — manual review recommended.
No matching publication found — likely fabricated or mistyped.
Summary Statistics
At the top of the results list, a stats bar shows the total citation count plus colored dots for each status that needs attention. Authentic refs aren't dotted — their verdict is already clear from the card icon.
Authentic (Found)
The citation was found and matched in academic databases with high confidence.
[1] Vaswani, A., Shazeer, N., Parmar (2017). Attention is all you need.
A Vaswani, N Shazeer, N Parmar
What this means:
- The source exists in academic databases
- Author names, title, and year match the records
- You can click to view the original source
Authentic with Notice (Found with Discrepancies)
Found the paper, but a field (year/authors/URL) doesn't match — citation may be slightly wrong.
[2] Piketty, T. (2016). Capital in the twenty-first century. Harvard University Press.
Capital in the twenty-first century
T Piketty
- Year mismatch. Please double-check.
Unsure (Needs Review)
The citation was found but some details don't match exactly. Manual verification recommended.
[3] Smith, J. (2019). The role of attention in transformer networks. Journal of ML Research.
Attention mechanisms in transformer architectures: a review
J Smith, L Wong
Found a paper with a closely matching title, but the author list has an extra name (L Wong) — possibly a name collision or a different review on the same topic. Manual confirmation against the original is recommended.
AI agents aren't always 100% correct. If a result looks off, give it another try.
Common reasons:
- Minor differences in author name formatting
- Year or volume number mismatch
- Title has slight variations from the original
Inauthentic (Fast Verify - Not Found)
After searching academic databases, no matching publication was found. Often a fabricated, hallucinated, or mistyped citation. Click Deep Verify to recheck more thoroughly before concluding.
[4] Akqdp, Z. (2020). Trnsfrmrs of Plzvktic Manifolds in Bzfrxk Lattices. Nature Press
Possible causes:
- The citation may be fabricated or AI-generated
- Spelling errors in author names or title
- The source is not indexed in academic databases
- Very recent publication not yet indexed
Inauthentic (Deep Verify - Not Found)
Even Deep Verify couldn't match the citation. The citation is very likely fabricated, hallucinated, or fundamentally wrong. From here you can search for plausible replacements — candidates stay hidden behind a CTA so credits are spent only when you ask for them.
[5] Akqdp, Z. (2020). Trnsfrmrs of Plzvktic Manifolds in Bzfrxk Lattices. Nature Press
Searched multiple academic databases with various keyword + author + publisher combinations — no match. Both author name and title look highly irregular, suggesting an AI-hallucinated or mistyped fabricated citation.
AI agents aren't always 100% correct. If a result looks off, give it another try.
Inauthentic (Deep Verify - Not Found, with Suggested Replacements)
When a citation can't be found in the real world, CiteTrue searches for papers that plausibly match your intended meaning — so the result isn't just a dead end.
[5] Akqdp, Z. (2020). Trnsfrmrs of Plzvktic Manifolds in Bzfrxk Lattices. Nature Press
Original citation matched no real paper (likely fabricated). Based on the citation context (keywords like Plzvktic Manifolds / Bzfrxk Lattices), the following related papers may be what you intended to cite.
AI agents aren't always 100% correct. If a result looks off, give it another try.
Invalid (Not a Citation)
The text isn't a citation — the verifier won't search. Use Find instead.
See chapter 3 for more details on the methodology.
Error (Verification Failed)
The verifier or a data source failed transiently — retry typically resolves.
[6] Garcia, M. (2021). Neural network optimization techniques. ACM Computing Surveys.
Server busy, please try again later.
Exceeded (Credits Ran Out)
Your credit balance ran out before this ref ran — top up to verify.
[7] Brown, K., Davis, R. (2022). Large-scale language model evaluation. NeurIPS.