Mechanisms,
not adjectives.

Every firm gets a single-tenant deployment: its own private network, its own compute, its own encrypted storage, its own encryption keys, in its own isolated environment on AWS. There is no shared application, no shared database, and no pool your files could be pooled into. Your files are never commingled with another firm’s and are never used to train any AI model.

The firm’s system runs in AWS ca-central-1 (Montreal), with encrypted backups to a second Canadian region, ca-west-1 (Calgary). Backups run daily and are encrypted with the firm’s own key. The record of work is tamper-evident: edits after the fact are detected and flagged for the lawyer.

Your firm’s system has no open inbound ports. The working agents are walled off from anything beyond their own matter files, with write control that denies by default.

CaseClarity reads the matter folders your firm already keeps (OneDrive and SharePoint today, via Microsoft’s own API) on a read-only basis. It cannot alter, move, or delete anything in your files. Scanned documents are read by a document-reading service hosted in Canada, in the same environment as the rest of the firm’s system, with no document retention.

Your firm’s system, files, and backups are hosted in Canadian regions. AI model processing is the disclosed exception: it runs on Anthropic’s API, hosted on US infrastructure, under contract terms with a no-training posture, data-protection agreements, and deletion of API inputs on a short retention cycle. We would rather tell you this plainly than claim an end-to-end residency the architecture does not have. The written client-consent workflow included with every deployment covers exactly this processing.

When a citation is checked against CanLII, the only thing transmitted is the citation string itself: never a client name, never a word of the file.

The system structurally cannot send anything outside the firm: no part of it has a path to email a client, a court, or opposing counsel. It cannot file and cannot sign. Sign-off and anything leaving the firm are reserved to the lead lawyer and cannot be delegated away, even by configuration.

Approvals run on single-use cryptographic tokens held so that even the firm’s own server cannot forge the lawyer’s sign-off. Before the AI touches a file, the matter must clear four professional-responsibility checks on the record: the conflicts confirmation, a limitation review, the client’s AI disclosure, and a jurisdiction match. The record proves they happened first.

Compute is metered at cost and itemized to its matter, with per-file spend caps and a firm-level backstop; the lawyer’s own instructions always run, even at a cap. Every step of the work is written into a permanent matter record as it happens. Costs, decisions, and deliverables are all traceable after the fact.