Before You File

Bottom line: the most expensive unknown in a litigation budget is whether your case turns on the law or on the judge. Ariadne's Thread, a paper Ross Brodskiy and Nathan Pokov posted to SSRN in July, proposes a number for exactly that. It is not ready to run your budget, and the authors say so plainly: the method is built, the math about the instrument is proved, and the test that would show it tracks real cases has not been run. But you should know what the number would change and what to ask before you trust it, or anything like it that a vendor puts in front of you.

From the client's chair, the question is never academic. The CFO wants a reserve. The board wants to know whether to worry. Outside counsel says "it depends on the judge," which is true and useless. This paper tries to turn "it depends on the judge" into something you can put on a page.

The number in one paragraph

The method runs a panel of simulated judges over the same motion. Each one has a different but lawful temperament: one ends weak cases early, one lets them develop; one is generous with discovery, one keeps it tight. Every judge reads from the same verified, current body of law. Each gives a score from 1 (the claim gets through this step) to 10 (it gets thrown out), with a one-line reason. Each judge is also run more than once, to measure how much the machine wobbles on its own. The score, called Z, is the disagreement between judges that is left after you subtract the wobble, on a scale from 0 to 1. It is reported step by step: motion to dismiss, discovery, class certification, experts, summary judgment and so on.

Three possible answers:

• Low Z. The law decides this step, whoever is on the bench.
• High Z. The judge's lawful temperament decides it.
• No number. The disagreement is too close to the machine's own wobble to call, so the instrument abstains.

That third answer is a feature. A tool that always gives you a number is sometimes giving you noise.

If Z is low

Budget. The case turns on the merits, so spend on the merits. You do not need to pay for anyone's theory of the judge. The path is set by the law, which makes staffing and timing easier to plan.

Settlement. A low-Z case is one where both sides can compute the answer. The paper points out that this is exactly the kind of case that tends to settle, because once the parties see that the merits decide and the judge is irrelevant, there is little left to fight about. The later versions of the paper put the practical point bluntly: a low-Z case with a poor prediction is a bad case. If your merits are weak and the openness is low, you are not unlucky. Settle early and stop spending. If your merits are strong, do not discount your position for judge risk that is not there.

Motion strategy. Brief the law and brief it hard. The motion lives or dies on it.

If Z is high

Budget. Build two budgets, one for the branch where the claim survives and one for the branch where it does not. A ruling that goes against you can sometimes be reversed on appeal, but the time and money spent before the reversal are never refunded. Reserve for the path, not just the outcome.

Settlement. Price in the risk that a lawful but differently disposed judge changes the path. The paper frames the output as a bargaining tool, not an instruction to pay a particular number: you can show that the dispute is closed at one step and open at another, and a mediator can see the live fork. It also narrows an old gap. Repeat players often know which steps are judge-sensitive in a given courthouse; one-time litigants usually do not. A shared map lets both sides argue about the same uncertainty.

Motion strategy. Find where the openness lives. The method produces what the authors call an anti-pretext ledger: each simulated judge marks the exact point where its ruling is a lawful choice, not a command of the law. That is where your argument should go. It is also what you preserve for appeal. If the trial court later dresses a discretionary call as legally compelled, the ledger has already located the seam.

The step that should drive your reserve is probably not the motion to dismiss

This is the most useful practical point in the paper, and it cuts against instinct.

The motion to dismiss feels like the big event. It is early and visible. But dismissals and summary judgment are reviewed fresh on appeal, so a trial judge's call there can be undone. Discovery rulings, class certification, expert admissibility and the scope of remedies are reviewed deferentially, so the trial judge's call there tends to stick. The paper calls the first group path-determinative but outcome-provisional, and the second outcome-durable.

So the method reports two things beyond the headline number: a durability-weighted score that focuses on the deferential steps, and the durable kill point, the deferential step where a differently disposed judge most durably ends the case. A high score at the expert gate may be harder to unwind than a high score at the pleading gate. If you are setting a reserve, that is where the money should be watching.

One caution: the authors state this pattern as a conjecture to be tested against real reversal data, not a proven fact. Treat it as a strong hypothesis grounded in how appellate review works.

What to ask before you trust any such number

Whether it comes from this method, a vendor, or your own firm's analytics team, ask these. If the answers are vague, the number is decoration.

1. What law did the machine read? Every simulated judge should reason over the same verified, current law. If each one pulled its own authorities, the disagreement measures retrieval error, not openness. The later versions of the paper add a hard rule worth borrowing: no language model writes a citation or a validity verdict.
2. Did the arguments clear a validity check first? A brief resting on an overruled, invented or misread case is not an open case. It is a defective one, and the method gives it no score at all. Aggressive but valid arguments go through, flagged.
3. What is the noise floor? Was each simulated judge run more than once? Show me the within-judge spread. The authors' own demonstration ran its panel judges once each, could not apply the correction exactly as designed, and said so in the write-up. That is the standard of candor to expect.
4. Did it abstain, and by what rule? Ask for the test and the threshold. A tool that never abstains is guessing some of the time.
5. What is the interval? A score to two decimals looks authoritative. The paper's demonstration produced 0.38 at the pleading step with an interval running from about 0.14 to 1. That locates the step as open. It does not say how open.
6. Which panel? The score depends on which temperaments are on the panel. A panel of near-identical judges reads low on everything. The panel should be versioned and published with the number. Two vendors with different panels will give you different numbers for the same motion.
7. Is my case in scope? The paper claims trial-level procedural rulings in large-dollar civil litigation, with the initial deployment aimed at the New York Commercial Division. Not criminal sentencing, not immigration, not constitutional questions, not the merits.
8. Has it been validated? Not yet. The only logged run so far is one engineered case, one step, one underlying model and a hand-built stand-in for the full legal library. The authors label it a test of the mechanics, not evidence that the score tracks real cases.
9. Can the simulated judges actually disagree? This is the biggest risk the paper names. Recent research reports that frontier models tend to behave like one strict formalist no matter what temperament you assign. If that holds, the score reads low everywhere, and a clean low number looks like certainty when it is really a panel of clones. Ask whether the panel still disagrees on cases independent experts rate as hard.

One more piece of context. The authors are affiliated with Legawrite.AI, which builds the system and discloses a commercial interest in it. Read accordingly. They have also stated every claim that matters as something that can be proved wrong, which is more than most vendors offer.

What it will not do

It will not tell you who wins. It will not predict with certainty what a particular judge will do. It will not decide anything, and it is not meant to replace your lawyer's judgment. The paper is explicit that it belongs under human control, as decision support and an audit, never as an adjudicator. If anyone sells you more than that, walk.

Here is what I want from you

To outside counsel, starting with the next case:

• Tell me, step by step, whether this case turns on the law or on the judge. You already do this by instinct. Write it down in plain English.
• Put the durable kill point in the budget memo. I want to know where this case is most likely to die for good, not just where it is most visible.
• Show me the fork. One paragraph on the exact point where a lawful judge could go either way, and what we are doing about it.
• If you bring me a number, bring its interval, noise floor, panel and abstention status. Otherwise do not bring the number.
• Do not call anything validated until the validation exists. When the calibration results come out, send them to me.
• A human signs everything. No exceptions.