Two model families from two companies, checking each other

A model has blind spots, and a second copy of the same model tends to share them. Two model families built by two different companies were trained differently, make different mistakes and notice different things. Putting them on the same file, with a rule that neither one's say-so counts as a source, turns their differences into a check. The engine does not treat agreement between the two as proof of anything on its own; proof comes from the record, the rule text and the opinions actually read.

The engine currently runs on Claude as the main working model family: it researches, drafts and verifies. Grok works as the second model family: it reads the complete file for itself and tries to find what the first pass missed. The roles are fixed by the process, not by which model sounds more confident.

On the default full lane, the second model family works at up to six points:

1. Deadlines. It computes every deadline again, from the same document and the same rule text. A mismatch is argued from the rule text, never split down the middle.
2. Research. It hunts for authority and angles the first pass missed. Everything it finds is a lead, verified by the engine's own standard before it can be used.
3. The second-model consult. Before strategy, it reads the complete case file and proposes its own plan. Two independent agents from that model do this one after the other; where they agree, the input carries more weight, and where they differ, both views are recorded.
4. The strategy debate. It debates the finished strategy under the debate protocol below.
5. The strongest attacks. On the full lane it can be asked for the three strongest attacks the other side will make, each with the counter that answers it.
6. The final text. At verification, it reads the complete final draft on its own, and both models sign their verdicts.

On the express lane, used for routine filings, the research, second-model consult, strategy and attack phases are skipped, and the second model's work in those phases is skipped with them.

See where the second model works in the process.

The second model family is there to widen the search, and a wider search brings in more candidates, some of them wrong. So the rule is strict: a case citation offered by either model is a lead only. It passes the same four-part standard as any other case citation (primary source, read in full, supports the point, still good law), or it waits in quarantine. The second model is also told to use its own tools to check every case and rule it offers, and to read each opinion in full before suggesting it; an answer that shows no such check is rejected. That is how its finds are leads re-verified before use.

When the two models reach different answers, they don't vote. They follow a written protocol:

1. Evidence only. Each side states its position with evidence: record citations, rule text, opinions actually read. "I would do X" with nothing behind it is a preference, not a position.
2. Three rounds at most. In each round a side either agrees or counters with evidence. A round that adds nothing new still counts as a round.
3. Signed agreement. Agreement is recorded with both signatures and its basis.
4. Escalation, not surrender. If they still disagree after three rounds, both positions and their evidence go up to the checkpoint that fits the question. Neither side gives way just because of who the other is.
5. Word for word. Every exchange is filed as labeled, verbatim turns, never a summary.

Engineers can read the debate protocol in detail.

Illustration, fictional matter. Scripted turns in the engine's real format. The matter, record and cases are invented.

Strategy debate · Avery v. City of Example (fictional) · rounds 1–2 of 3

Claude · round 1 · position. Lead with the claim against the City itself. The complaint pleads a written policy (Compl. ¶¶ 14–19), and the policy is in the record (R. 31). Basis: R. 31; ledger rows 12 and 14.

Grok · round 1 · counter, with evidence. The claim against the individual officer has the stronger record (R. 52–60). Leading with the City claim puts the whole case on the policy theory. Lead with the officer claim, answer the immunity defense directly, and keep the City claim in full. Basis: R. 52–60; ledger rows 7 and 9.

Claude · round 2. Agreed that the officer record is stronger. But the adverse sweep found controlling authority on immunity that cuts against the theory as pleaded (ledger row 21, marked adverse). Leading with the officer claim walks straight into it.

Grok · round 2. Row 21 held the right was not clearly established on facts from before a later controlling decision (ledger row 22). The conduct here came after that decision (R. 55). Distinguish row 21 in terms, on that timing, and lead with the officer claim.

Claude / Grok · AGREED: lead with the officer claim; distinguish ledger row 21 in terms on the timing of the conduct (R. 55; ledger row 22); keep the City claim in full. Basis: R. 31, R. 52–60, R. 55; ledger rows 7, 9, 12, 14, 21, 22.

Had they not agreed by round 3, both positions and their evidence would have gone to the strategy sign-off, with both positions and the evidence.

The same protocol applies whether the question is a date, a reading of an opinion or a choice of strategy: the answer that stands is the one the source supports, and the record shows why.

By default, the second model family is also part of how the engine attacks its own work. The plans are attacked as opposing counsel would attack them, and the draft is read as a hostile panel would read it. A second model family is well suited to that job because it did not write the thing it is attacking. More on the hostile-panel read.

If the second model can't be reached when the second-model consult is due, the event is recorded and the matter pauses before strategy. It does not quietly continue on one model's plans. The pause clears in one of two recorded ways: the second model returns and the consult completes, or a waiver to proceed without it is recorded and logged. It is never a silent skip.

A client's standing instruction to proceed through the checkpoints does not stand in for this waiver. See the recorded waiver.

• We do not claim that two models review the work in every circumstance: they review the work by default, and the waiver path exists.
• We do not claim any model is a lawyer, or that agreement between two models makes an argument right.
• We do not claim either company endorses Legal Ops Depot.

Claude is a trademark of Anthropic, PBC. Grok is a trademark of xAI Corp. Other names are trademarks of their respective owners and are used only to describe the technology in our process. Legal Ops Depot is not affiliated with, sponsored by or endorsed by Anthropic or xAI.

The model mix can change, which is why we say "currently." Read the full trademark and non-affiliation notice.

Why not just use the strongest single model? Because the failure that matters most in legal drafting is a confident mistake, and a single model cannot reliably catch its own. A second model family that was built and trained by a different company brings different blind spots, so it is more likely to notice what the first one missed. The engine then settles the difference on the source, not on which model is more persuasive.

Does the second model just repeat the first one's work? No. By default, it reads the complete record on its own, builds its own candidate strategy, and runs its own search for authority. Its job is to disagree where there is something to disagree about.

Can one model outvote or overrule the other? No. There is no vote. Agreement is recorded with both signatures and its basis. A disagreement that survives three rounds comes to you at the matching checkpoint, with both positions and the evidence. It is never averaged.

Is the debate record kept? Yes. Debates between the model families are filed word for word, so a supervising attorney can read how a point was settled and on what basis.

What does "by default" mean? The full lane runs both model families. A routine filing on the express lane skips the phases where most of the second model's work happens, and a matter can proceed without the second model only on a recorded, logged waiver.

Is Legal Ops Depot affiliated with Anthropic or xAI? No. We name the models only to describe how the engine works. Legal Ops Depot is not affiliated with, sponsored by or endorsed by either company.

Will the engine keep using Claude and Grok? The mix can change. We name the current models on this page and update it when they change.

Email us with questions about how the two model families work a matter, including how a debate is settled.

Start a request (opens the main site) See the adversarial review

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