How Justine uses AI — and your data.
The governance behind the reasoning: what the models see, where they run, how citations are grounded, and how the attorney stays in control. Verified against the running product.
1. No training on customer data
Customer matter data is never used to train or fine-tune any model. The production system uses models only for inference — there is no fine-tuning, file-upload, or batch-training path against customer data. Eve-Genesis™, the synthetic substrate used to fine-tune the legal reasoner, is synthetic by construction and held in a storage account, subscription, and identity boundary separate from customer data.
2. Where inference runs
All model inference routes to Azure OpenAI / Azure AI Foundry endpoints inside MindHYVE.ai™’s Azure tenant, authenticated by Managed Identity. No customer content is sent to a third-party model-vendor SDK on the production path. Frontier-model providers are under contractual terms prohibiting retention of, or training on, inference content.
3. Model identity on customer surfaces
The customer-facing model label is Eve-Legal F5/reasoner. The underlying foundation model is not exposed on any customer surface; user-facing error handling is sanitised so that no vendor or model identifier can surface.
4. Grounding & citation integrity
AI-suggested case-law citations are verified against the CourtListener (Free Law Project) public corpus. A citation is presented as verified only when it resolves to a real opinion; unverified suggestions are explicitly labelled research leads for the attorney to confirm. Only the public citation string is transmitted for verification — no matter content.
5. Human-in-the-loop & disclosure
Client-facing intake discloses, unconditionally, that the user is interacting with an automated AI system — not an attorney — that responses are AI-generated and may be inaccurate, are not legal advice, and form no attorney-client relationship. All work product is produced under the supervision of the attorney of record, who reviews and attests. The AI reasons; the attorney decides.
6. Auditability & prompt handling
AI-initiated actions are captured in the structured audit log (action, actor, timestamp, matter reference). Matter-derived reasoning is surfaced to the attorney in-product; it is not persisted into the audit store, which is deliberately kept free of matter content. Prompts and responses are never written to plaintext operational logs, and matter data is processed only to generate the requested artifacts.