Bernio OS

Bernio OS is an agentic legal operating system in which each legal matter runs in one governed environment, from intake through strategy, drafting, review and approval.

Key features and functions include:

Confidentiality by architecture
Client documents, matter data and the audit record are stored in the user’s own environment and are not shared with Bernima or with model providers. There is no vendor-held copy of matter data.

Context obfuscation
Before any request leaves the user’s environment, its context is obfuscated on the user’s own hardware. Privileged information, party names and client identifiers are removed, and the matter context itself is obscured so an external model receives the legal task without the client’s confidential context. Each routing decision is recorded, and calls to external models can be disabled entirely.

Local models and token economics
Each task is routed to the model best suited to it. Bernima’s own small and medium-sized models, built for legal tasks, handle sensitive work on the user’s hardware, and purpose-built models handle classification, extraction, ranking, deduplication, entity resolution and date calculation. Large language models are called only where a task requires them and firm policy allows, limiting external token use.

Supervised AI agents
Each agent works with declared skills, a runtime-enforced autonomy ceiling and a named supervising lawyer, inside the matter’s own documents, chronology and strategy. Agents cannot be assigned until a practitioner is bound to them, and every output passes practitioner approval before it reaches the client.

Matter workspace
Every matter runs from intake to close in one controlled workspace. New matters are held at a gate until a named practitioner releases them, with potential conflicts surfaced for a partner to clear. AI-extracted chronology events are held until a lawyer confirms, edits or rejects them, and every entry is pinned to its source document. The case theory is versioned and cited to the instrument, with each issue setting out the rule, application, conclusion and a stated confidence level, and authorities are retrieved from legal sources rather than generated. Case law, legislation, regulator guidance, dockets and other defined sources are monitored against the matter, and agent work is billed on its own line, at its own rate.

Isolation and record
The matter boundary is applied in the runtime before any retrieval or model call, and agent memory is sealed when the matter closes. Every AI action enters a signed, append-only record showing who ran what, when, on which matter, with which model version and who approved it. The record stays in the user’s environment, can be exported in full at any time, and handles erasure requests by cryptographic severing rather than rewriting entries.