Adversarial Legal Review
Better your own assistant finds the weakness in your thesis than your opponent finds it in the hearing. A single 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 workflow builds that move into the method: one role 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 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 (or the opposing-counsel-attack workflow in this pack) instead of the full debate.
Workflow (4 roles + verification)
Each role is a separate pass with a clean mandate. Pseudonymise the input if the document contains privileged or personal data.
Reviewer access tiers (anti-bias)
Reviewers split into two tiers with DIFFERENT access - deliberately:
- ISOLATED (internal consistency): the logic-and-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. If this reviewer saw the sources, it would confuse "sounds-like-the-source" with "is-internally-consistent".
- AUGMENTED (grounding): the attacker and the citation check see the sources and the intake, and pull contrary authority.
When one model simulates all the roles, isolation is prompt discipline, not a hard sandbox - explicitly instruct the ISOLATED role not to reach for sources.
0. High-stakes gate
Assess whether the matter qualifies (table above). If not, stop and offer an ordinary review.
1. Builder - the strongest version of the thesis
Build the strongest possible argument for the deliverable's thesis. Gather the best rules, authority, and doctrine. Goal: give the attacker a hard target, not a straw man. Output: thesis + 3-7 pillars, each with its legal or evidentiary basis.
2. Attacker - devil's advocate
Attack each pillar like opposing counsel or a sceptical court: contrary authority (real sources for the relevant jurisdiction), a doctrinal counter-argument, a factual or evidentiary gap, over-reading of a rule, an omitted exception, an out-of-date line of authority, procedural risk (deadlines, standing, jurisdiction). Output per pillar: the charge, its strength (high/medium/low), and the source of the charge.
3. Synthesizer - the balance
For each pillar decide: survived / weakened / defeated. State what is left of the thesis after the attack, where the deliverable needs reframing, where a caveat is required ("contested risk, authority is split"). Output: a table of pillar, verdict, recommended change.
Priority hierarchy when charges collide: logic (consistency) ~ citations (grounding) ~ counter-arguments, all above style, above clarity. The substantive layer never trades against itself - on conflict, accumulate BOTH views and add a caveat; do not pick arbitrarily. Style and clarity never override substance.
4. Verifier - final check (10 points)
- Every citation grounded - the quote exists in the source word for word (BLOCK on a failure)
- Every pillar has a basis
- No defeated pillar left in the final thesis without a caveat
- The attacker's contrary authority addressed, not swept under the rug
- No categorical claims where the line is contested
- Currency of the rules (not repealed or amended)
- Internal consistency (thesis does not contradict the reasoning)
- Scope matches the client's question - no more, no less
- Procedural risks listed
- Confidence level stated explicitly (no false certainty)
Revision loop (bounded)
The debate is not single-pass, but it runs within hard limits, with a guarantee that the shipped version is never worse than the best one seen.
- Versioning: draft_v1 (after the first synthesis), draft_v2... Each iteration = targeted edits from the synthesizer's recommendation, then a re-attack and verification.
- aggregate_score (0-100) = the mean of: verifier (X/10, scaled) + consistency (ISOLATED, 0-100) + grounding (AUGMENTED, 0-100). Hard component: the citation grounding check - any failure blocks regardless of score.
- v_best: after each iteration, remember the draft with the highest aggregate_score.
- Iteration cap: high-stakes = max 3 rounds; quick mode = max 1.
The synthesizer owns the exit decision, evaluated in order:
- Clean approval - all reviewers accepted, no blockers: exit approved_on_vN.
- Regression-revert (from round 2) - the new score is lower than v_best: reject the round, restore v_best, exit with a banner "round N made the deliverable worse - reverted to the best version".
- Plateau early-exit (from round 2) - score at or above threshold (default 85) but improvement under 1 point: exit early rather than burn effort on marginal gains.
- Continue - blockers remain, cap not reached, no regression or plateau: another round of targeted edits.
- Forced exit at max - cap reached with blockers: deliver the best version with a banner listing the unresolved blockers.
Terminal states and always-deliver
Every ending MUST produce an artefact for the human - never a silent or empty exit. Terminal states: approved_on_vN, accepted_early_on_vN, forced_exit_on_v_best (regression), forced_exit_at_max, failed_with_fallback (e.g. no source material - deliver what exists plus the reason). The closing banner always shows the terminal state, which draft version was delivered, and what blockers remain.
Output
A review block containing: the stakes assessment, iteration history with scores, a pillar table (pillar, attack and strength, verdict, action), the ten-point verifier result, the confidence level after the debate, and a plain recommendation ("do NOT send before fixing the failed citation and adding a caveat to the defeated pillar"). Keep the full debate transcript (builder, attacker, synthesizer) as an attachment - it is the evidence of adversarial verification. Where AI-governance rules apply to the firm's tooling (e.g. EU AI Act Articles 12 and 14), the transcript doubles as the record of human oversight.
Governance boundary
The workflow produces a recommendation and draft versions; it does NOT send the document. The decision to send despite remaining blockers stays with the human. For privileged material, pseudonymise the input before running; pulling contrary authority uses public sources.
Position in this pack
The judicial-first-impression workflow reports neutrally how a text lands; the opposing-counsel-attack workflow is the cheap single-pass attack; this workflow is the top rung - the full debate for high-stakes deliverables. After revision, the output-scoring workflow turns the result into a send / revise / verify decision.
Attribution
The debate-plus-verification pattern is inspired by AnttiHero/lavern (Apache-2.0); roles, prompts, and the ten-point check are written from scratch. Bounded self-revision (the exit tree with regression-revert), reviewer access tiers (isolated/augmented), the synthesizer's conflict-priority rule, and the always-deliver invariant are adapted clean-room from gregmos/memoforge (MIT) - concepts of loop control and roles, not prompts or code.