What judges actually do on Rule 12 motions in S.D.N.Y.

Ask any commercial litigator in New York how Judge X handles 12(b)(6) motions, and you will get an answer. The answer will be confident, specific, and delivered with the tone of a person who has read every order on the docket. It will also, in our experience, have been formed from a sample size that — if you actually pulled the underlying motions — would not survive a graduate-level statistics class.

This is not a criticism of litigators. It is a structural feature of how the bench is read. Each partner sees a few dozen orders from any given judge over the course of a decade, weighted by recency, weighted by the kinds of cases they happen to bring, weighted by what their associates write up in the memo at the end of the matter. The result is a working model of every judge in the district — a model that is largely formed by intuition, plausible to the people who hold it, and frequently wrong.

We ran the actual numbers. Twelve thousand four hundred motions to dismiss filed in the Southern District of New York between January 2018 and February 2026, scoped to commercial matters, scored by outcome, and bucketed by the active judge at the time of decision. What we found is roughly what you would expect a real distribution to look like — wider than anyone’s intuition, with outliers in both directions, and with judges whose reputations do not match their numbers.

The bench is a distribution, not a vibe

If you stopped reading here, you would already have a more useful number than most litigators carry around. Thirty-one percent is the median, but the spread is what matters. The fifth-percentile judge — the most plaintiff-friendly on commercial MTDs in our dataset — grants in roughly 14% of cases. The ninety-fifth-percentile judge grants in roughly 58%. That is a four-fold difference in dismissal probability, controlling, to the extent possible, for motion type and matter complexity.

What is surprising is not that the spread exists. Of course it exists. What is surprising is how poorly the spread maps onto the reputations the bench carries inside the district. Two of the judges with the highest grant rates in our dataset are not, as far as we can tell, in any working litigator’s mental list of “grantors.” Two of the judges with the lowest grant rates are widely reputed to be defense-friendly. The reputational ranking and the empirical ranking, when plotted against each other, are weakly correlated — closer to noise than to signal.

Tenure does not buy predictability

There is a folk theorem in litigation circles that older judges are easier to predict. They have settled into a pattern. They know what they like and what they do not. The data does not support this. When we plot judge tenure against the variance of their grant rate across motion subtypes — Rule 12(b)(1) vs. 12(b)(6) vs. 12(b)(2), and within 12(b)(6) by subject matter — the variance does not narrow with time. In several cases, it widens. Senior judges in our dataset are not more consistent. They are, in some cases, more variable, because the volume gives them room to be.

We are not making a normative claim about whether this is good or bad. We are making a factual claim that you cannot proxy for predictability with seniority. If you are a litigator and you need to know what a specific judge does on a specific kind of motion, you have to actually pull the orders.

Why the gut is wrong

There is a pattern to where the gut breaks down. It is not random. It is structural, and it is the same set of biases that show up in every domain where human intuition is asked to do work that should be done by a model.

• Recency. The last two orders you read from this judge dominate the prior.
• Selection. You only see the motions in matters you or your firm worked on, which are not representative of the docket.
• Salience. The unusually large or unusually quotable orders are over-weighted. The boring 12(b)(6) grants in middle-of-the-night minute orders are under-weighted.
• Aggregation. “This judge denies MTDs” is true on average for some subject matters and false for others, and the partner-level model usually does not split the difference.

Each of these is well-known in any field that takes prediction seriously. None of them are corrected by simply reading more orders. They are corrected by building a model and looking at the distribution.

What to do with this

If you are a litigator working in S.D.N.Y., the practical implication is straightforward. The mental model you carry of the bench is probably accurate at the level of broad reputation and inaccurate at the level that actually matters for venue selection, motion strategy, and settlement timing. The fix is not to read more orders by hand. The fix is to start asking, on every matter, what the actual distribution of the bench looks like on the specific motion you are about to file.

If you are a GC, the practical implication is different. You are not choosing the judge. But you are choosing the outside counsel that will argue the matter in front of that judge, and you should be asking your firms to show you the empirical profile, not the partner’s opinion. If they cannot produce one, that is information.

If you are a funder or insurer, the implication is structural. You are pricing a contingent payoff that is materially conditioned on which judge draws the matter. Carrying around a reputational prior of the bench is not a model. Building a real one is now cheap. The shops that do are going to outperform the ones that do not.

You would not price an option without a vol surface. Don’t pick a venue without one either.

Notes on the dataset

The S.D.N.Y. dataset covers all 12(b)(1), 12(b)(2), 12(b)(3), 12(b)(6), and 12(c) motions filed between January 1, 2018 and February 28, 2026 in commercial matters, defined as cases assigned a commercial-track Nature of Suit code. Outcomes are coded as grant, partial grant, denial, or moot/withdrawn, with partial grants weighted at 0.5 in the headline figures. Where a motion was decided by a magistrate judge and reviewed de novo, the reviewing district judge’s outcome is the one we score. We will publish the methodology paper later this year. For per-judge profiles on an active matter, email us at info@valarhq.com.