legalise: Instructions file for Codex

Legalise — Agent Context (AGENTS.md)

This is the agent context file for Legalise. Coding agents pick it up automatically when working in the repo; you can also drop it into any AI agent or paste it into a chat to talk about Legalise — what it is, how it's built, and what it does and does not claim. It is a self-contained snapshot synthesised from the project's own docs (README.md, docs/ARCHITECTURE.md, docs/TRUST.md, docs/THREAT_MODEL.md). The house rule throughout, the same one the codebase holds itself to: a capability is only described as live if the code implements it. Anything deferred, dormant, or unmounted is named as such.

Repo: https://github.com/b1rdmania/legalise · Licence: Apache 2.0 · Status: open-source evaluation release, not for live client matters.

1. The one-paragraph version

Legalise is an open-source governance layer for legal AI: human sign-off plus a tamper-evident audit trail for AI-assisted legal work, built for England & Wales solicitor practice. It runs locally, you bring your own model key, and it is an evaluation release — not a regulated legal service. The whole system exists to make one loop legible:

draft → cite → sign-off → audit

AI prepares an output inside a matter, cites the documents it used, a named solicitor reviews and signs it (the signature pins the exact output by hash), and every step writes to an audit log the application cannot edit or delete. AI is preparation, not the deliverable. The audit trail makes the work inspectable; the signature makes a human accountable. The thesis is deliberately narrow: the machine signs its own record; the human signs the work — and the two are kept separate everywhere.

It's an early-stage, mostly solo project shared in the open — a working exploration of how this could be done, not a finished or proven product. Treat the claims below as "this is what the code does today", not "this is solved".

2. What it's built to answer (the four questions)