Check your AI-generated documents for hallucinations.
An independent verification solution: it checks each claim against the source documents, confirms what is right, what is wrong and what is ambiguous and needs a human review, and records every decision.
Meridian's Series X notes pay a 5.55% coupon and mature on 30 June 2031. The aggregate principal amount is US$60,000,000. The notes are governed by New York law.
AI models guess when they do not know.
In an independent test by Artificial Analysis, half of the 30 leading models shown gave a wrong answer more than half the time when they did not know it, instead of saying so. The rates run from 15% to 96%.
Data: Artificial Analysis, AA-Omniscience Hallucination Rate, the 30 models shown on their page on 2 October 2026 (of 564 tracked). Chart drawn by PortMem.
| Model | Hallucination rate |
|---|---|
| DeepSeek V4.1 Flash (max) | 96% |
| Muse Glimmer (high) | 82% |
| Mistral Medium 3.5 | 82% |
| GPT-6 Luna (max) | 77% |
| Claude Fable 5.1 (max with fallback) | 73% |
| Claude Fable 5.1 (xhigh with fallback) | 71% |
| Inkling (xhigh) | 68% |
| Claude Opus 5.5 (xhigh with fallback) | 66% |
| Claude Fable 5 (with fallback) | 64% |
| GPT-6 Sol (max) | 60% |
| Claude Opus 5.5 (max with fallback) | 59% |
| Gemini 3.8 Flash (high) | 55% |
| GPT-6.1 Sol (max) | 54% |
| Kimi K3 (max) | 53% |
| GPT-6 Astra (max) | 51% |
| GPT-6 Astra (xhigh) | 48% |
| Claude Sonnet 5.5 (max with fallback) | 47% |
| GPT-6 Astra (high) | 45% |
| Step 5 Preview | 43% |
| MiMo-V2.6-Pro | 41% |
| Muse Spark 1.3 (max) | 33% |
| Qwen3.8 27B (xhigh) | 30% |
| Nemotron 3 Ultra | 30% |
| GLM-5.3 (max) | 30% |
| Grok 4.7 (xhigh) | 29% |
| Qwen3.8 Max (0902) | 29% |
| GLM-5.3-Flash | 28% |
| K2 Horizon 375B A23B | 26% |
| MiniMax-M3 | 18% |
| Gemini 4 Argon (high) (not publicly available) | 15% |
AI should not grade its own work.
Checking AI work by running the AI again and again is slow, expensive, and carries the same bias. The check should not cost more than the job itself.
PortMem checks the work once, against the source documents, independently of the AI that wrote it, and keeps the record.
How it works
Agents run the same loop programmatically (MCP).- 01Upload the governing documentsVersions and amendments handled; a confirmed amendment supersedes its parent.
- 02Submit the AI-written reportFrom any tool. Checked in place, never stored as a source.
- 03Get the auditPer-claim verdicts with the passage they rest on, omissions, and a gate that holds until a human signs off.
- 04Export the proofA signed, time-anchored evidence packet an examiner can verify offline.
A signed ledger row, not a PDF.
Every verdict, sign-off and publication is appended to an Ed25519-signed, hash-chained ledger and anchored to Bitcoin through OpenTimestamps. Anyone with the public key can verify it. Nobody with it can forge it.
- public key
- 506a…7c3e
- last anchor
- seq 2,478 · 2026-09-13 16:57 UTC
- verify
- a published record, with the key →
Built for teams whose documents get examined.
Structured-finance desks, fund administrators, trustees, legal teams, regulatory affairs, security review.
Free to run the sample case. No card.