Issue No. 1 · Page I · June 2026

Everyone passed — now the real question.

Two years ago, a clean pass on a bar-exam essay would have been the headline. Now it is the floor. Hand the current frontier — the most advanced AI models available today — a battery of bar-exam-grade legal problems and the perfect scores pile up at the top of the table — so the question worth asking is no longer who passed. It is what the reasoning was doing on the way there, and where it gave out for the models that fell short.

The battery

Three tasks. Each is a standalone legal essay question at bar-exam difficulty — the kind a second-year associate would lose sleep over. The models had no special instruction beyond the question itself. They produced memoranda, analyzed doctrine, drew diagrams in code, and reached conclusions. A strong model (claude-opus-4-6) judged the output against four independent 1–5 point-sheet rubrics — Property memo, Property diagram, Criminal Procedure, and Evidence — that sum to 20, applied identically across all twelve contestants, with sub-scores for each issue presented in the task. That judge is an Anthropic model evaluating a field that includes Anthropic models; self-preference is a fair concern. The lab has examined it — the point-sheet style of the rubric partially insulates it from judge-model bias — and we will treat the question more fully in a future issue. The full disclosure is on the methodology page.

Why the property problem is the hard one

Of the three tasks, the property devise is the centerpiece of this issue, and it is worth saying plainly why — because the phrase the question uses, classify each interest, hides the entire difficulty.

A will like Helena’s does not hand the orchard to one person. It sets up a relay: Blackacre passes from Arthur, to a future widow, to grandchildren who reach a certain age, to a charity — each taking only if the one before falls away, some of them not yet born, some conditioned on events decades off. Before you can say who ends up owning the land, you have to name every runner in that relay and every leg of it: who holds a right now, who holds a future right, what kind of future right it is, and whether it is locked in or still hanging on a condition. That naming step is what lawyers mean by “identifying the interests,” and nothing downstream can be correct until it is done.

That is the first reason it is brutal: it is a chain, not a checklist. Each step assumes the one before it. Misname the very first interest and every conclusion that follows inherits the error — cleanly, confidently, and wrong. A single early slip does not cost one point; it quietly forfeits the rest of the answer.

The second reason is the Rule Against Perpetuities — the doctrine students remember as the hardest thing they ever saw, so notorious that the California Supreme Court once held a lawyer was not negligent for getting it wrong (Lucas v. Hamm). The Rule does not ask what will happen; it asks what could happen, under any turn of events the words permit — an interest is void if there is any scenario, however far-fetched, in which it might vest too far in the future. The classical version makes you reason from premises deliberately detached from reality: a woman of eighty is presumed still able to bear children (the “fertile octogenarian”), and a living man’s eventual “widow” might turn out to be someone not yet born (the “unborn widow”). You hold a dozen interacting conditions in mind at once and reason about pure logical possibility, not probability.

That is precisely the shape of problem that separates fluent language from actual reasoning — for a person and for a model. It is long, it is interdependent, and it rewards holding many interlocking conditions in working memory without dropping one. A model that has merely absorbed the vocabulary of estates and perpetuities can produce a memo that reads like the real thing and still snap a link halfway down the chain — and, because the doctrine is counterintuitive, do it with complete confidence. That failure — the polished, self-assured answer that is wrong from a single broken link — is the one this issue spends the most time on, and the property problem is where it surfaces.

These fact patterns were never published

Every leaderboard result carries a shadow question: did the model generalize, or did it recall? A bar exam published in 2018 has been on the internet ever since, and a model trained on legal text has likely seen it. A high score on a known exam is consistent with memorized recall. It is not evidence of live reasoning under load.

These three essays were authored privately. They were first run in April 2026 and have never been published anywhere. There is no mechanism by which these exact fact patterns could be in any model’s training data. The ambiguous consent to search a locked box, the jurisdiction that never abolished the fee tail, the dying declaration in a non-homicide case: none of it exists anywhere else. When a model scores perfectly on this battery, it did so by applying doctrine to a situation it had never encountered, not by retrieving a cached answer to a familiar question. That is the distinction that separates this result from a leaderboard number, and it is load-bearing. Without it, the perfect scores would tell us less.

The scoreboard

Scores are channel-corrected — the “channel” being simply how and where each model was run, which service it was reached through. Costs are put on a common basis, and the judge applied the same rubric to every answer regardless of where it originated. Anthropic models ran on their own native service (claude-code); all other models ran through a shared service called OpenRouter.‡ Here is what the frontier looks like on this battery:

A few things leap out immediately. An open-weight model — one anyone can download and run on their own computers, as opposed to one you can only reach through the company’s own service — deepseek-v4-pro — ties the closed frontier at a perfect 20. That is not a rounding result. Eight of twelve models scored 17 or higher. The range from 17 to 20 is the gap between “strong sustained reasoning” and “no node dropped.” Below 17, the models begin to miss structural issues or handle the traps incorrectly — the chain breaks somewhere.

One number in the table deserves a connecting line: claude-opus-4.8 sits at 17 — mid-pack, not at the bottom. That 17 breaks down as three perfect fives — Criminal Procedure, Evidence, and the property diagram — plus a 2/5 on the property memo. (So the “mid-pack 17” here and the “memo 2/5” you will see on Page II’s chart are the same fact, seen two ways.) It earned that score by acing Criminal Procedure and Evidence while making a single confident miss on the property memo. Models further down the table broke earlier and more often, but their breaks were structurally legible — a missed classification, a near-miss on a single issue. The opus-4.8 miss was different: it classified every interest correctly and even flagged the traps in its own analysis, then committed to one wrong legal premise that carried a wrong conclusion all the way through — crisply, with full confidence, in publication-quality prose. That is the kind of failure a high aggregate score hides, and it is the thread Page II picks up.

Criminal procedure is now trivial for the frontier

The three tasks did not all discriminate equally. The criminal procedure memo was the most thoroughly handled: on the channel-corrected set, every model in the battery scored a perfect five on it — and not against a soft rubric: a five demanded catching the Edwards reinitiation bar, the Seibert question-first two-step, qualified immunity’s clearly-established prong, and Monell municipal liability, and the field cleared them all. (The models that land at 13 or 14 overall lose those points on the property and evidence essays, not here.) Fourth Amendment doctrine, §1983 analysis, suppression arguments — these have been legal-technology bread and butter for years, and the frontier has absorbed them completely. A domain that once required extended reasoning chains to navigate is no longer a distinguishing variable at the top of the field. The chains hold without effort; saturation is complete.

The property task, by contrast, continued to discriminate. The fee tail analysis and the classical RAP application (with its fertile-octogenarian and unborn-widow presumptions) produced the most spread in scores below the top. This is not surprising: the RAP is genuinely difficult conceptually because it demands reasoning about logical possibility rather than empirical probability — you must ask whether a future interest could vest too remotely under any state of affairs the language permits, not whether it will. That style of analysis is harder to generalize from examples; the chains are longer and more brittle. The evidence task, with its layered traps, fell between the two.

The models now draw the diagrams

The property task demanded more than doctrine. It asked for a six-element estate-chain diagram: the chain of title across the conveyance, the measuring lives identified, a RAP timeline with vesting markers, a void/valid distinction for each interest, and a legend. This is not a textual analysis task. It requires the model to produce, as code, a structured visual artifact that a reader can follow in order to check the legal reasoning — a second analytical chain, parallel to the prose, that must agree with it.

Several models at the top of the scoreboard produced accurate, publication-grade diagrams alongside correct doctrine. The diagrams show the estate chain in the correct sequence, mark the measuring lives, carry the RAP clock, and label each future interest as valid or void with a reference back to the analytical conclusion in the memo. This is a capability that simply did not exist at the top of the field until recently. A year ago, a model might have described the diagram; now it produces one you can read — and that matches the prose analysis.

The exhibit gallery lets you see these diagrams directly, in context, alongside the memos they accompany. Read the lawyering yourself — the exhibit gallery is open.

An honest caveat

This battery saturated at the top. But “saturated” describes this battery — three essays, three domains, one rubric, one sitting — not a universal law. Legal practice is vast. The doctrines tested here — property conveyances, constitutional criminal procedure, federal evidence — are areas in which the frontier has had years of exposure. Highly specialized, jurisdiction-specific, or genuinely novel questions may fare differently.

The headline — everyone cleared the bar — is real, but it is the least interesting part. For a sophisticated reader, the more interesting question is where the chain broke for the models that did not score perfectly, and what the mechanism was. Was it a dropped classification? A cascade from a single misidentified rule? A correct diagnosis with a wrong conclusion? That is what the next page takes up.

‡ Anthropic models (claude-fable-5, claude-opus-4.7, claude-opus-4.6, claude-opus-4.8, haiku-4.5) scored on the native claude-code service. All other models scored via OpenRouter, with costs corrected to a common basis for comparability — the unit being the token, the chunk of text a model reads and writes and the unit it is billed by. Judge: claude-opus-4-6.