Adversarial review: the engine attacks its own argument first

A brief that has only been read by the people who wrote it has been tested by the people least likely to see its gaps. Good litigators moot their arguments, invite a colleague to play the other side, and read the draft the way a skeptical bench will. The engine builds those habits into the process, as written steps, so they happen on every full-lane matter, not only when there is time.

The adverse-order audit

When a government actor has won an adverse order, the engine starts there, before any strategy talk. The order is laid side by side with the filing it answered, and the audit records:

• which arguments the order addressed, and which it passed over;
• whether each rule the order cites says what the order says it says, checked against the rule's full text;
• whether each case the order quotes is quoted fairly, checked against the full opinion.

Anything passed over or misstated is kept as ground that may matter on rehearing or appeal, and at strategy every finding is turned into an affirmative argument. The audit makes no assumption about why an order says what it says; it records what the order, the record, the rules and the cases each say.

Every known trap, answered one by one

Research runs a loss-pattern sweep: the known ways a claim of this type gets dismissed in this kind of court, drawn from the engine's record of past losses and denials. Each trap is listed with how it is avoided, or marked "none applicable" with the reason. At strategy, the list is answered trap by trap. A trap table with an unanswered row does not pass.

Before anything is drafted, planners each write a full plan independently. Each plan explains every avenue of relief, including what each one cannot do. Then the plans attack each other in structured rounds:

• As opposing counsel. Where would the other side hit this plan, and does the plan survive it?
• As a hostile panel. Read the plan as a skeptical court would read it. What would such a panel ignore, misread or reject, and what would it be forced to address?

By default the second model family files its own independent plan and critique, and the merged plan is debated against that critique. Points the two cannot resolve come to you. That is the second model as adversary.

No legal theory dropped without a reason

Every plausible theory for the relief sought is written into a theories ledger, as included or excluded with a reason. Arguments may overlap where overlap helps; none may be weak filler. A theory is never dropped silently, and a ledger with a blank decision or a blank reason does not pass. When a supervising attorney asks "did we consider this?", the answer is on the record, with the reason it is in or out.

Illustration · fictional

Theory	Decision	Reason
Unreasonable seizure (officer claim)	Included	Strongest record (R. 52–60)
Municipal policy (City claim)	Included	Written policy in the record (R. 31)
State-law battery	Excluded	Notice-of-claim period ran before filing (R. 9)

The preservation map

The plan attack also asks a long-range question: if this argument loses here, is it still alive above? A preservation map records where each claim must be raised now to keep it alive in every higher court. See preservation for every court above.

After the filing is drafted against the court's requirements, the attack phase runs again on the actual text.

1. The other side's attack. The draft is attacked as opposing counsel would attack it. The argument pivots away from anything that falls rather than leaving a weak point in place. The attack is not finished until at least three attacks have been answered or pivoted.
2. The hostile-reader simulation. The draft is read as a hostile panel would read it, one looking for a way to skip, misquote or sidestep each point. Each vulnerable point is rewritten so that skipping it would create visible, reviewable error: the record is cited, the objection is explicit, and the standard of review is named.
3. The final citator check. The citator runs again over the final list of case citations actually relied on.
4. The second model's attacks. On the full lane, the second model family can also be asked for the three strongest attacks the other side will make, each with the counter that answers it. If one of them breaks the strategy, the matter goes back to strategy.

This matters most in motion practice, where the other side's best argument is predictable and the response has to meet it directly. See attacking a motion to dismiss.

The engine reads the draft as an unreceptive panel might: passing over an argument, quoting out of context, or quoting selectively. Then it rewrites every vulnerable point so that doing any of those things becomes a visible, reviewable error:

• the facts cited to the record;
• the objection stated explicitly;
• the standard of review named;
• a denial framed so that it can be reviewed.

The aim is to keep the point alive for the court above, even if the court below never reaches it.

Illustration, fictional matter.

Point	How an unreceptive panel could avoid it	Rewritten so avoiding it is visible error
Exclusion of the (fictional) witness's testimony	Treat the issue as never preserved	"Plaintiff made an offer of proof (R. 212–14), and the court ruled on it (R. 215). The issue is preserved under Fed. R. Evid. 103(a)(2). Standard of review: abuse of discretion."
The limitations question	Rely on the order's paraphrase of the rule	Quote the governing text in full, including the clause the paraphrase leaves out, and cite where it was argued below (R. 88).
The City's immunity defense	Decide it on a ground the City never raised	"The City argued only one ground below (R. 40–46). No other ground was pressed or passed upon."
Request for findings	Rule without findings	"Plaintiff asked the court to state its findings (R. 118). The order states none."

The table is kept with the matter, so a supervising attorney can see each weak point that was found and exactly how the draft now meets it.

In the element sweep, each element counts as answered only when four things line up: a paragraph in the draft addresses it, a record file supports it, a verified authority backs it, and real facts are stated. The element sheet, written when the filing type was chosen, lists every legal element the claim must prove. A draft with an unanswered element is caught here, even if every argument in it survived attack. Attacking an argument is not the same as proving a claim. So the verification phase adds a different test: an element sweep.

The final reader answers four questions in writing: does it read well, does it follow the court's rules, is it the best that can be produced, and does it ship or go back? If it goes back, the draft returns to drafting with its defects named, and the attack and verification receipts of phases 11 and 12 are voided and re-run on the new text. An argument that survived attack on an earlier version does not keep that status once the text changes.

The loop is bounded. Two send-backs without convergence return the matter to strategy, and a second strategy escalation on the same matter puts it on hold for you to decide.

See re-verification after a kickback and what a kickback re-runs.

Sent back: the attack and verification checks of phases 11 and 12 are voided and re-run on the new text.

• It is not a prediction of how a court will rule. It is a test of how the argument holds up against the strongest attacks the engine can find.
• It does not replace a lawyer's judgment. For firms, the attack memo and survival table are work product for the supervising attorney.
• It is not a one-shot answer. It runs at the plan stage and the draft stage, and it re-runs after a kickback.

The glossary has hostile panel and theories ledger defined, along with the other terms used here.

Is this the same as asking an AI to critique its answer? No. Each attack is a required step with a written output, done, wherever the roles allow, by a different role from the one that drafted, and a failed attack changes the draft or the strategy. The survival table and the trap table must be complete before their phases can close.

Who plays opposing counsel? At strategy, the two planning agents attack each other's plans and, by default, the second model's plan. In the draft, each section is attacked by an agent that did not write it. On the full lane, the second model family can also be asked for its own attacks. Wherever the roles allow, the builder never grades its own work.

What happens to an argument that does not survive? It is dropped or rebuilt. The argument pivots away from anything that falls, and the theories ledger records what was excluded and why, so the decision can be reviewed later.

What happens if the attack breaks the strategy? The engine pivots: a new theory, a new vehicle, or honest distinguishing. If the second model's attacks break the strategy, the matter returns to strategy, and the sign-off is asked for again if the strategy changes materially.

Email us with questions about the plan attack, the survival table and kickbacks.

Start a request (opens the main site) Hard cases

Legal Ops Depot is not a law firm and does not give legal advice.