| name | adversarial-legal-review-en |
| description | A red team for a high-stakes deliverable - takes a finished, high-stakes document (a legal opinion, a DD/M&A memo, a court/board submission, a recommendation) and runs an adversarial debate: a builder constructs the strongest version of the thesis, an attacker tears it apart with counter-arguments and contrary authority, a synthesizer reconciles, a verifier does a final check. Goal: find the weakness BEFORE your opponent, the court, or the client does. Cost-gated - for high-stakes work only, not every query (token-expensive). Use when: "attack this opinion", "red team", "devil's advocate for this document", "find the weaknesses", "stress-test the argument", "what will the other side say", "pre-mortem", "defend this thesis", "adversarial review", verifying a high-stakes deliverable before it goes out. English counterpart of adversarial-legal-review-pl (this version drops the Polish case-law layer and is jurisdiction-agnostic).
|
| license | Apache-2.0 |
| allowed-tools | ["Read","Write"] |
| data-residency | local |
| requires-human-approval | false |
| pii-egress | none |
| attribution | [{"source":"AnttiHero/lavern","license":"Apache-2.0","relationship":"pattern-only","note":"Debate plus 3-layer verification pattern. Roles and prompts written from scratch.\n"},{"source":"gregmos/memoforge","license":"MIT","relationship":"clean-room","note":"Bounded self-revision with regression-revert, reviewer access tiers and the always-deliver invariant. Concepts adapted clean-room, not prompts or code.\n"}] |
| metadata | {"author":"Wiesław Mazur / MateMatic","version":"1.0.0"} |
Adversarial Legal Review (English) - a red team for a high-stakes deliverable
Philosophy
Better your own agent finds the weakness in your thesis than your opponent finds it in the hearing.
A single LLM pass produces an argument that sounds confident - because the model is trained to sound confident. That is an illusion of competence. The real value of an expert is to foresee the counter-argument, the contrary authority, and the gap in reasoning before the other side does. This skill institutionalises the adversarial move: one agent defends, one attacks, one reconciles, one verifies.
This is NOT writing a deliverable from scratch. It is a stress-test of a finished deliverable.
Cost gate (read BEFORE running)
The full cycle is token-expensive. Run it only for high-stakes work:
| Run (high-stakes) | Do NOT run (routine) |
|---|
| A legal opinion going to a client | An internal note |
| Due diligence / an M&A memo | A summary of a judgment |
| A court submission before a deadline | A working draft |
| A recommendation to a board | FAQ, marketing content |
| A position carrying material financial or reputational risk | A routine query |
If the matter is not high-stakes, say so plainly and offer an ordinary single-pass review instead of the full debate. Do not burn tokens on routine work.
Workflow (4 roles + verification)
Each role is a separate pass with a clean mandate. Pseudonymise the input (e.g. with a humaniser/anonymiser step) if the document contains privileged or personal data.
Reviewer access tiers (anti-bias)
Reviewers split into two tiers with DIFFERENT access - this is deliberate, not an accident:
- ISOLATED (internal consistency): the logic/structure reviewer sees only the current draft - no sources, no authority. Mandate: does the thesis contradict the reasoning, do the conclusions follow from the premises, are there logical gaps. It does NOT judge whether a citation is real. Warning: if this reviewer saw the sources, it would confuse "sounds-like-the-source" with "is-internally-consistent".
- AUGMENTED (grounding): the attacker (role 2) and the citation check (verifier point 1) see the sources + intake, and pull contrary authority. Mandate: grounding of claims and contrary authority.
In single-agent mode (one LLM simulating the roles), isolation is prompt discipline, not a hard sandbox - explicitly instruct the ISOLATED role not to reach for sources. Hard isolation via separate passes is the premium option.
