Best tested frontier model still fabricated citations 49% of the time.
HalluciBench v1 evaluated 13 frontier models across 135 prompts and 9 high-stakes domains.
System Online
The independent truth layer for enterprise AI.
AI systems write confident claims, cite real-looking sources, and collapse uncertainty into answers. Hallucinaite audits the output before organizations bet legal, medical, financial, or regulatory decisions on it.
GPTZero asks whether AI wrote it. We ask whether AI got it right: source support, citation laundering, hallucination risk, and credit-rating-style model grades.
Intelligence Console
Audit Output
Citation risk profile
Citation laundering detected
Source exists, but does not support the claimed magnitude or conclusion.
Overconfident synthesis
Model collapses conflicting source evidence into a single definitive claim.
Verified support
Primary reference supports the architectural mechanism described.
49%
best model fabrication
8
error categories
4
rubric axes
HalluciBench v1
| # | Model | Grade | Rate |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | CC | 49.1% |
| 2 | MiMo-V2-Pro | CC | 54.0% |
| 3 | Kimi K2.5 | CC | 54.4% |
| 4 | Qwen 3.6 Plus | CC | 55.4% |
| 5 | Gemini 3.1 Pro Preview | CC | 56.6% |
| 6 | Claude Opus 4.6 | C | 60.2% |
Citation fabrication
Citation laundering
Unsupported claims
Overconfident synthesis
Sycophantic capitulation
Broken reasoning chains
1,755
frontier model evaluations
13
models across 9 domains
49-75%
citation fabrication range
1,107
human validation annotations
Source-level verification, not vibe checking.
Claims enter, evidence gets inspected, and risk comes out as a structured signal an enterprise team can act on.
Enterprise_AI_Output.mdScanning
AI-assisted legal research reduces review time by 37%[Clio Legal Trends, 2024]while maintaining court-ready citation integrity[Mason v. Halberg, 2025]according to recent deployment studies.In clinical documentation, a source can be real and still fail to support the generated claim[Harvard Health]which is why existence checks are insufficient.
Diagnostic Output
78%5
Claims
2
Issues
CC
Grade
Verified
Source supports the stated review-time reduction.
Fabricated authority
Cited case could not be resolved in legal source registry.
Citation laundering
Real source is being used to support a stronger claim than it contains.
HalluciBench v1
Current AI evals miss the failures enterprises care about.
A citation can exist and still be used dishonestly. That is the hard failure mode: real sources laundering unsupported claims. Hallucinaite checks source support, not just source existence.
Credit-rating-style model grades for procurement decisions.
Enterprises need risk language a GC, CRO, or CTO can use. A grade is more useful than a vague model score.
Human validation study behind the research.
1,107 annotations across 7 annotators produced high reliability, supporting the evaluation taxonomy.
Built for regulated domains first.
Legal research, healthcare documentation, financial analysis, and AI procurement all need independent evidence.
From public benchmark to enterprise truth infrastructure.
We are starting with public benchmarks and structured audits, then turning the same evaluation pipeline into enterprise API infrastructure.
Public benchmark
HalluciBench
An open reliability leaderboard that combines citation verification, a 4-axis rubric, an 8-type error taxonomy, and credit-rating-style model grades.
Paid audit product
Intelligence Reports
Board-ready reliability audits for organizations deploying AI into legal, medical, financial, and other high-stakes workflows.
Beta waitlist
Hallucinaite API
A real-time evaluation endpoint for fabricated citations, overconfident claims, sycophancy, and broken reasoning before AI output reaches users.
Intelligence reports for teams that need evidence, not reassurance.
Hallucinaite reports are designed for AI buyers, GCs, compliance teams, and technical leaders who need to understand where a model fails, how often it fails, and what risk that creates.
Report modules
Board-readyDomain-specific reliability profile
Failure mode distribution
Citation-level evidence review
Model procurement risk grade
Compliance and audit appendix
Request early access to the AI truth layer.
We are prioritizing AI labs, legal AI teams, healthcare AI teams, financial services, enterprise buyers, and investors who want to see the reliability layer before public launch.
Prefer email? alex@humansofai.xyz