Grounded generation
Every claim is mapped to a retrieved span before it is written. Unsupported sentences are deleted, not softened.
Mirneth is a research assistant for teams whose work cannot cite a hallucination. Every sentence it returns is traceable to a passage in your own corpus — with the paragraph, the page and the version.
1. Under Article 6(1), a system is high-risk where it is a safety component of a regulated product.[1] Our classification module is not a safety component; it informs human decisions.[2]
2. Obligations that still apply include documentation (Art. 11)[3] and logging (Art. 12).[4]
3. Where the feature is later extended to biometric identification, the analysis changes materially…
In use at institutions including
Retrieval first, generation second. The model may only write sentences it can support.
Every claim is mapped to a retrieved span before it is written. Unsupported sentences are deleted, not softened.
Pinned embedding model and index version per workspace. The same query on the same corpus returns the same sources.
Export a full reasoning trail — query, retrieved passages, rejected passages and final output — as a signed PDF.
When confidence falls below the configured threshold, Mirneth states what is missing instead of improvising an answer.
Third-party evaluation on a 12,000-question benchmark drawn from legal, clinical and financial corpora.
| System | Citation recall | Unsupported claims | Refusal accuracy |
|---|---|---|---|
| Mirneth 4 | 99.2% | 0.4% | 97.6% |
| General assistant A | 71.4% | 9.8% | 62.1% |
| General assistant B | 64.0% | 12.6% | 58.4% |
| Keyword + LLM baseline | 81.3% | 6.1% | 70.2% |
Methodology and raw data available under NDA. Benchmark maintained by the Mirneth research group; auditors are independent.
A research task, from intake to signed memo.
“The memo arrives with its evidence attached. My partners stopped asking me to re-check the citations.”
Dr. Adrian Vale · Head of Knowledge, Harrow & Voss LLP
Connect a document store, DMS or data room. OCR, layout parsing and redaction run on ingest; nothing is retrained.
Select the corpus subset and the matter or project. Permissions are inherited from the source system, not re-declared.
Ask in natural language or build a structured query. Each answer shows its sources inline with page-level anchors.
Export the memo with a provenance appendix and a hash of the corpus version used. Reviewers sign inside Mirneth.
Built for buyers who read the DPA before the demo.
EU (Frankfurt), UK (London), US (Virginia) or your own VPC.
SAML SSO, SCIM provisioning, and per-document permissions mirrored from source.
AES-256 at rest with per-tenant keys; TLS 1.3 in transit. BYOK available.
Your corpus and queries are never used to train or fine-tune any model.
Configurable from 24 hours to 7 years, with legal hold support.
SOC 2 Type II, ISO 27001, ISO 42001 (AI management). HIPAA and GDPR ready.
Annual agreements. Seat bands, not per-question meters.
Small teams, single corpus
£1,200/mo
Multiple practice groups
£4,800/mo
Regulated, multi-jurisdiction
Bespoke
A citation is verifiable when the quoted span exists in the indexed version of the source and semantically supports the claim in the sentence. Each citation carries a document hash, page and character range.
Yes. Mirneth supports model routing to your own provider endpoints, including on-premise deployments, while keeping the retrieval and verification layer intact.
Typical cloud deployments reach first value in ten working days; VPC and air-gapped deployments run six to nine weeks including security review.
The retrieval layer handles 24 languages and cross-lingual queries; generated memos are produced in the language of the requesting user.
We will run it live against a sample corpus in your industry and show you every source behind every sentence.