Most AI-ethics pages are principles. This one is mechanisms: for every principle that matters, the platform either enforces it with a gate, measures it with a hard floor, or records it in an instrument an auditor can read — and where it cannot enforce, it abstains and says so.
No client-facing advice or document exists before a lawyer explicitly approves it — enforced by the server, not by workflow guidance. Monitoring signals after release go to the lawyer only; a client never receives an automated legal conclusion. The EU AI Act classification of the platform rests on this design fact, and weakening it is a recorded reclassification trigger.
Every legal citation must verify against its official public source, or it cannot be displayed — there is no flag to ignore.
When the verified basis is thin, the platform abstains instead of guessing. Abstention is a designed, measured outcome, not a failure mode.
Deadlines and statutory dates come from counsel-signed deterministic rules, never from text generation.
Facts taken from client documents are labelled as the client's assertions and can never be presented as law.
An automated bias-evaluation suite probes whether outcomes shift with client characteristics that should not matter. The floor is zero strict violations, and it runs on every code change: a change that breaks it cannot be merged.
Every AI-assisted output carries its assistance label; clients see the counsel-approved AI notice before any interaction.
Performance claims enter our material only as measured numbers.
Counsel maintains a list of prohibited promise wording (e.g. “your advice stays current”); an automated check fails any release in which such wording appears in a client-facing string.
Every consequential action lands in an append-only, tamper-evident audit chain.
Any firm can download its own conformity dossier: a live-verified evidence file of its oversight, releases and certificates, generated deterministically — no model involved.
The platform's EU AI Act position is a counsel-signed determination on file (limited-risk, Article 50), with its transparency surfaces counsel-approved — and its controls are voluntarily documented to the standard the higher risk tier would demand.
The underlying language model inherits the biases of its training; our verification and abstention layers mitigate the consequences, they do not eliminate the cause.
The fairness floor covers the probes we have written. It cannot certify the absence of all conceivable bias.
Family law is deliberately out of scope: the platform abstains entirely rather than serve a domain where errors are hardest to bound.
Ethical review is carried by counsel of record, the data-protection officer and recorded decision triggers — not by an independent ethics board.
