The problem isn’t that AI is wrong.
It’s that we believe it’s right.
Charlotin’s legal hallucination database now tracks 1,610 court and tribunal decisions including 1,122 U.S. matters. 9+ careers ended. Most were preventable with verification.
ZH-1 catches fabricated citations, verifies regulatory eligibility, and scans documents for hallucinated facts — so you never file, submit, or publish something that isn’t real.
Historical figures are updating.
This is what happens when you trust AI without verification.
Lawyers sanctioned. Patients endangered. Businesses fined. These are real stories — and they happened because nobody checked whether the AI’s answer was actually true.
Stories sourced from public news coverage. Cost descriptions are editorial characterizations, not direct quotes.
Source-linked AI failures. Real costs. Real consequences.
From $110K court sanctions to $250M fraud schemes — we documented every case so you don’t become the next one.
Trace how an answer was reached.
Pin the legal text, the question, and the model version. Then anyone can see what each system answered, which pages it cited, and how the answer was reviewed.
See the difference for yourself.
Ask about Public Law 119-21. ChatGPT, Claude, and Gemini answer live; ZH checks every claim against the enacted text and shows the evidence behind each finding.
One Big Beautiful Bill Act(P.L. 119-21; H.R. 1, 119th Congress)
This comparison uses the enrolled bill corpus. ZH retrieves relevant section excerpts before drafting an answer; the other systems use their own sources. Open a citation to read the passage and its page in the authenticated Public Law PDF.
Ask your own
Type a question. Three chat models answer live; ZH checks every claim against the enacted text.
Ask a question about H.R. 1. Citation chips show the cited section, its excerpt, and the official PDF page where mapped. A live answer is not an attorney-reviewed eligibility decision.
30 fixed eligibility questions. Run pending.
The live comparison above shows individual responses. It does not establish an error rate. We will publish a numerator, denominator, model IDs, run dates, source pages, and the blind attorney-reviewed sample when the complete set has been run and reviewed.
Question set
30 frozen prompts
Model versions
Recorded per answer
Citations
Both sides with source pages
Attorney review
20 blind answers planned
No pass rate is asserted yet. A provider error, missing citation, or unresolved claim is recorded separately and is never counted as a model failure without review.
AI alone scores ~2%. With ZH-M: 70%.
ZH-M is a deterministic verification layer for mathematics. It amplifies frontier AI models on the FrontierMath benchmark — problems that took professional mathematicians hours to solve — pushing convergence from ~2% to 70%+ with zero hallucinations.
Convergence by Model (37 FrontierMath problems)
Full Results
| Model | Converged | 1st Attempt | Avg Iter | Tokens | Cost | Halluc. | Runtime |
|---|---|---|---|---|---|---|---|
| CSonnet 4.6 | 26/37 (70.3%) | 26 | 1.0 | 115K | $1.73 | 0.0% | ~37 min |
| COpus 4.8 | 26/37 (70.3%) | 23 | 1.2 | 115K | $8.64 | 0.0% | ~28 min |
| XGrok 4.1 | 26/37 (70.3%) | 25 | 1.0 | 507K | $15.22 | 0.0% | ~118 min |
| GGemini 3.5 Flash | 25/37 (67.6%) | 25 | 1.0 | 601K | $2.10 | 0.0% | ~22 min |
| LlMaverick 4 | 25/37 (67.6%) | 20 | 1.3 | 98K | $0.05 | 0.0% | ~25 min |
FrontierMath contains problems that took professional mathematicians hours to solve. Frontier AI models score ~2% alone. ZH-M’s deterministic verification loop pushes convergence to 70% with zero hallucinations.
Results from testing 37 FrontierMath problems. “Converged” means the model produced a verified correct answer within the iteration limit.
Real answers for the questions that matter most.
Built for regulated industries.
Here’s exactly what we do and don’t do with your data and our architecture.
- No document content stored after processing
- No training on customer data
- Encrypted in transit and at rest
- Architecture designed toward SOC 2 Type II readiness
- SHA-256 report hashing
- Every decision cites specific rules
- Report integrity verification
- Deterministic outputs for auditors
- Full processing integrity evidence
- Federal rule updates tracked and encoded
- Patent pending (US utility application filed)
In August 2026 we moved to a stricter counting method where every figure traces to an individually verifiable record; live counters below, prior-period totals above.
Every fact traces to its source — click through to the funder's own page.
When a source doesn't support a value, we show the gap, never a guess.
Live verification activity
Source-checked field decisions since August 9, 2026, shown separately from prior-period totals.
Demos only, for now.
There are no self-serve accounts yet. We are working directly with a small number of enterprise, government, and research partners. Tell us who you are, what you are working on, and where you need ZH to bring integrity, and we will reach out.
Questions we hear most.
Stop wondering if the AI got it right.
See ZH check real documents against the actual source text, with an audit trail for every decision. We respond within one business day.