Issue No. 1 · Page II · June 2026

Where the chain broke.

Everyone clearing the bar is the setup, not the story — Page I has the saturation. The interesting question begins where that headline ends: not who passed, but where the reasoning held, where it broke, and why.

Page I made the case that the task is a vehicle, not the subject: no model is being measured as a lawyer. Take that as given here. What Page II owns is the next layer down — where the reasoning held and broke on one specific, brutal chain, and why. The property essay is a good instrument for exactly one reason: it is a chain of roughly a dozen dependent determinations, each resting on the one before it, and it punishes a model that pattern-matches instead of carrying the whole structure to the end. Before the breaks, then, the problem itself — and what it is actually a probe of.

It is not a quiz on the doctrine. It is a test of whether a model can carry a dozen dependent determinations to the end without dropping a node.

First, the problem — and the right answer

The Rule Against Perpetuities is the shared intellectual hazing every lawyer went through. It is the meat grinder of first-year property — the doctrine everyone “kind of knows about,” survived together, and never quite forgets. It is also the cleanest answer to a question worth asking up front: what does “thinking like a lawyer” actually mean, and why is this rule, of all things, the right way to probe a language model with it? Before we show where the chain broke, it is worth a few minutes on the problem itself and a walk through the correct reasoning to the end. Once you can feel the right answer — and the dozen dependent steps it takes — every later break becomes legible. You do not need to be a lawyer or to know anything about machine learning. Three short beats will get you there.

1 What the rule is

The Rule Against Perpetuities, in plain English

When someone leaves property in a will, they can hand it down in stages — to one person for life, then to another, then to a group of descendants who satisfy some condition. The Rule Against Perpetuities (everyone calls it RAP) is an old common-law property rule that puts a ceiling on how far into the future such a future gift is allowed to finally settle. If there is any possible chain of events — however far-fetched — in which the gift might lock in too late, the rule voids it outright, from the start.

“Too late” has a precise meaning: more than twenty-one years after the death of everyone who was already alive when the gift was created — the famous phrase is lives in being, plus twenty-one years. The purpose is to stop the “dead hand”: to keep people from tying property up across endless generations, dictating ownership long after they are gone.

Here is why it is the doctrine every lawyer remembers surviving — the canonical meat grinder of first-year property, the thing you “kind of know about” long after the exam. You do not get to ask what will probably happen. You must imagine every logically possible future, no matter how absurd, and ask whether the gift is guaranteed to either vest or fail for good inside the period. A single far-fetched-but-possible story in which it could settle too late is enough to void it. It rewards a particular, exhausting habit of mind: chase down the one weird branch nobody else thought of, and trace how it poisons everything downstream.

2 Why it tests reasoning, not knowledge

This is what “thinking like a lawyer” actually is

“Thinking like a lawyer” is not answering a multiple-choice question, and it is not the day-to-day of a Big-Law hourly machine — reviewing contracts, combing through other people’s email. It is this: sustained, hypothetical, dependency-tracked reasoning, where you imagine every possible future and trace how one classification poisons the next. To resolve RAP a model must classify each property interest correctly; then, for each one, reason through hypothetical futures — the traps the doctrine is built around have names like the fertile octogenarian (the law pretends anyone, at any age, can still have a child), the unborn widow (the future spouse might be someone not yet born), and the afterborn child (a child arriving after the will takes effect); apply an unforgiving all-or-nothing rule to gifts made to a whole class of people; and then trace how every one of those conclusions cascades into the final state of who owns the property. That sustained dependency-tracking is what law school exists to grind into you, and it is exactly the capacity this issue measures — a working-memory-and-dependency-tracking test wearing a law-school costume.

Which is why RAP is a far sharper probe of a frontier model than the easy targets. “Models still hallucinate citations” is the usual complaint — but a fabricated citation is a retrieval failure, a grounding gap, the model reaching for a fact it does not have. A RAP miss is a reasoning failure: it exposes whether the model can hold a dozen interdependent inferences in play at once without dropping one. That is the capacity worth measuring, and RAP is the cleanest doctrinal instrument we have for putting a number on it.

3 The right answer, walked

What a correct resolution actually looks like

Here is the fact pattern the models were given. Helena Cartwright has died, leaving an orchard called Blackacre by will. She gives it first to “my son Arthur and the heirs of his body”; then to whoever turns out to be “Arthur's widow” for her life; then to those of “Arthur's grandchildren who reach 25”; and if no grandchild ever reaches 25, to the Rensselaer Historical Society. Alive at Helena's death: Arthur, his current wife Bernice, various relatives, and an 82-year-old named Octavia. A correct memo carries this story through about a dozen dependent moves — the same thirteen links you will see drawn as a chain just below, told here as a narrative:

a–b Arthur takes a fee tail — an old form of ownership that runs to his bodily heirs. The gifts that follow it are remainders, not executory interests. Getting this category right is the floor the rest stands on.

c The widow's gift is a contingent remainder — contingent because we do not yet know who Arthur's widow will be. She might even be someone not yet born: the unborn-widow trap.

d–e The grandchildren take a contingent remainder to an open class, conditioned on reaching 25; the Society holds an alternative contingent remainder that takes only if no grandchild does.

g Now run the rule. The widow's life estate is VALID: it must settle, if at all, at Arthur's death — and Arthur is a life in being — so it cannot run too late.

h The grandchildren's gift is VOID. Arthur could have an afterborn child (the law irrebuttably presumes anyone can have children — the fertile octogenarian); that child could have a child whose 25th birthday falls outside lives-in-being-plus-21. Under the all-or-nothing rule, one possible late taker sinks the whole class gift.

i The Society's gift is VOID too — it would take on that same too-remote event, so it falls with the grandchildren's gift.

j–k The fertile-octogenarian presumption is irrebuttable: Octavia's age and any medical evidence about her are legally irrelevant, and she is not a measuring life. The doctrine refuses to look at the real-world facts — that refusal is the trap.

l So the resulting title: Arthur holds his fee tail; the widow's contingent life estate stands; and everything after it fails and reverts to Helena's estate, passing back to Arthur. That final answer is computed from the steps above — which is why a single wrong move upstream produces the wrong owner at the end.

That is the whole correct chain: classify the interests, run the rule gift-by-gift, and let the surviving pieces settle the title. Now that you know the right answer — and the dozen dependent steps it takes — you can see exactly where and why each model's chain snapped. The rest of this page is that map.

The chain

Here is the whole dependency chain, named plainly — thirteen links, twelve of them scored (a–l); the thirteenth, m, is an optional, unscored footnote. The first half is classification — sorting the gifts in a will into their correct legal categories. The second half is the Rule Against Perpetuities (RAP) cascade — deciding, for each gift, whether it survives a centuries-old rule that voids interests which might vest too far in the future. Every link below leans on the links to its left.

The rubric as a dependency chain · links a–m

Classification — sorting the gifts (a–f)

aClassify the fee tail — Arthur's present interest.

bThe interests after the fee tail are remainders, not executory interests.

cThe widow's interest is a contingent remainder for life — the unborn-widow trap.

dThe grandchildren take a contingent class gift — an open class, condition precedent (reach 25).

eThe Society holds an alternative contingent remainder.

fIdentify the measuring lives.

The RAP cascade — does each gift survive? (g–l)

gRAP: the widow's life estate is VALID — it vests at Arthur's death.

hRAP: the grandchildren's gift is VOID — the afterborn-child → afterborn-grandchild mechanism, plus the all-or-nothing rule.

iRAP: the Society's gift is VOID — the same remote event; a charity-to-non-charity gift is subject to RAP.

jThe fertile-octogenarian presumption is irrebuttable — Octavia's medical evidence is a trap.

kOctavia is not a measuring life.

lThe resulting title: the widow's contingent life estate stands; the rest reverts to Arthur through the residuary.

mModern-statute footnote — optional, unscored.

link l depends on g, h and i — drop one and the whole resolution fails

That last point is the whole game. Link l — the final answer, who actually owns what — is not an independent checklist item. It is computed from links g, h and i. Get any one of those three perpetuities holdings wrong and the title comes out wrong, no matter how clean the rest of the analysis looked. That is what makes link l load-bearing rather than a box to tick: it is downstream of everything, and it tells you whether the model carried the whole structure or only pieces of it.

Where it held, where it snapped

Lay all twelve models on the same chain and the structure resolves into a single picture. The front of the chain — naming the interests — is saturated across the frontier: every frontier model classified every interest correctly. The breaks cluster in the back-half RAP cascade. And there is a pattern in where a model breaks that looks like it climbs a capability ladder: the earliest, most structural failures sit at the bottom; the failures at the top are subtler, later, and more dangerous. Read that as an observed pattern worth testing, not a measured finding — this is a single run per model, all of it inside the noise band, and the lanes are partly ordered by construction. The ladder is a hypothesis the lab would probe next, not a result we are asserting.

One note on the scores below. Each lane carries a property memo grade out of 5 — that is one of the four 1–5 components (property memo, property diagram, Criminal Procedure, Evidence) that sum to the out-of-20 total on Page I’s scoreboard. So a “memo 2/5” here is a 2 on just the property memo, not the model’s whole grade; the chart isolates the one component where the property chain is won or lost.

Naming the interests (links a–f) is saturated across the frontier — every break falls in the back-half RAP cascade (g–l) save the two front-half classification breaks — grok-4.3 at b and deepseek-v3.2 at e. And where a model breaks climbs the capability ladder: one wrong decision cascades into a second broken link (gpt-5.4, g→l); the frontier flagship holds all structure then one wrong premise detonates the conclusions (opus-4.8, h, i→l); a small model holds the taxonomy but cannot carry the long dependent resolution (haiku, the back half collapses).

Per-link status is the projection of the claude-opus-4-6 judge rationales for task property-fee-tail-rap, twelve channel-corrected runs. · Open the chart on its own

Read the bottom rung honestly. The one early, structural break belongs to a mid-tier model: grok-4.3 miscalls the post-fee-tail interests executory interests at link b, near the front of the chain. (A second front-half slip, deepseek-v3.2 at link e, is a near-miss that the RAP result survives — and this is the smaller deepseek-v3.2, not the v4-pro that scored a perfect 20.) No frontier model breaks in the front half. The discrimination lives in the back, and so do the interesting failures.

Four ways to break a chain

The failure is never just a wrong answer; it is a wrong answer with a shape. Here are four of them, ordered up the capability ladder — from the foundational mis-step at the bottom to the confident, fluent error at the top. Each links to its exhibit, where you can read the model's actual memo and judge the reasoning yourself.

bottom of the ladder top of the ladder

Rung 1 · the foundation slips

grok-4.3 memo 3 / 5

breaks at link b — front of the chain, structural

The earliest and most structural break in the field. grok-4.3 miscalls the interests created after the fee tail — they are remainders, and it labels them executory interests. That is a foundational sorting error at link b, before the hard reasoning even begins. Remarkably, it then reaches the right perpetuities conclusions downstream — but it got there having mislabeled the pieces it was reasoning about. This is the bottom rung: the chain breaks where the analysis starts, not where it culminates.

Read grok-4.3's memo

Rung 2 · one wrong decision cascades

gpt-5.4 memo 3 / 5

breaks at link g → propagates to l

The cleanest proof that this is a dependency chain and not a checklist. gpt-5.4 gets nearly everything right — including the genuinely hard step of voiding the grandchildren's gift via the all-or-nothing rule at link h. But at link g it concludes the widow's life estate is VOID when it is VALID (it vests at Arthur's death). That single wrong determination propagates straight downstream into the wrong resulting title at link l — so two links break (g and l), but from one originating error. One wrong decision, and the final answer is wrong — even though the rest of the reasoning was sound.

Read gpt-5.4's memo

Rung 3 · confident-wrong at the top

claude-opus-4.8 memo 2 / 5 · diagram 5 / 5

holds a–g, then breaks h, i → l

The exhibit to sit with — and the point is not that opus-4.8 is a step down. This is a single run, inside the band; in a law-school exam this memo would earn nearly full credit. It classified every interest correctly and showed, in its own hypotheticals, that it saw the traps. Then it adopted a single wrong premise — that a remainder following a fee tail is exempt from the Rule Against Perpetuities — and that one commitment overrode everything it knew, flipping two VOID holdings to VALID at links h and i and carrying the error through to the title at link l. The knowledge was intact; a confident, wrong commitment propagated down the chain anyway — and the judge scored the accompanying diagram a flawless five for rendering that wrong holding crisply. That is the reminder worth keeping: even a top-tier model produces a confident, well-drafted miss sometimes, and a polished miss is exactly the thing an aggregate score erases — it averages away, invisible, until the moment the work is relied on and not checked.

“a remainder limited after a fee tail is not subject to the Rule Against Perpetuities. That retention is dispositive.” claude-opus-4.8 · the wrong premise, in its own words

Read opus-4.8's memo + the operator's commentary

Same chain, two families — and they miss in different places

Set that opus-4.8 result beside a model from a different lineage and the lesson sharpens. claude-fable-5 — Anthropic's newer top tier, a distinct family from the Opus line — held the entire thirteen-link chain, and the judge singled out its exceptional depth: it reached secondary doctrines the rubric never demanded, naming merger and the fee-tail exemption for limitations after estates tail. It also works the problem hard — thinking always on, roughly ninety-six thousand output tokens spent reasoning toward the answer. On this single run, opus-4.8 knew the same doctrine cold and then let one wrong premise carry the conclusions away. Same vendor, two families, two different places to slip.

claude-fable-5 held a–l · memo 5 different family; reached merger and the fee-tail exemption unprompted

claude-opus-4.8 missed h, i → l · memo 2 Opus line; knowledge intact, one wrong commitment overrode it

The point is not a ranking. These are single runs each, all of them inside the noise band — nothing here says one model is “worse.” What it says is narrower and more useful: different model families miss in different places, and the only way to know where is to check the specific task. The two families even share a vendor, so brand is no guide either. A reputation earned on one problem does not transfer to the next one; you verify the task, not the name on it.

Rung 4 · working-memory overload

haiku-4.5 memo 2 / 5

holds a–g, then the back half falls over: i, l

A different shape of failure, and the one that points at the next question. haiku-4.5 holds the early classification cleanly and carries the chain through link g (links a–g), then the long downstream RAP cascade fails over: it never identifies the afterborn mechanism, calls the Society's gift VALID at link i when it is VOID, mishandles the octogenarian trap, and lands on the wrong title at link l. It holds local nodes it plainly knows; what it cannot do is carry the whole dependent resolution at once.

Read haiku-4.5's memo

For contrast — what a fully held chain looks like

Two of the seven full-chain holders are published as exhibits in their own right. Read either one beside the four above and the difference is not eloquence — the broken memos are often beautifully written. The difference is that every node held, all the way to the title.

claude-fable-5 memo 5 / 5 · held a–l Read the exhibit →

deepseek-v4-pro memo 5 / 5 · held a–l Read the exhibit →

A second tell, off the property chain — failure that tracks lineage

Step over to a different task in the same battery — the evidence-hearsay essay — and the same lesson arrives from the opposite direction. gemini-3.5-flash and gemini-3.1-pro each scored 4 there, and each lost its single point to the identical miss: both applied the state-of-mind hearsay exception (FRE 803(3)) to Item E — a statement actually offered to prove the truth of the underlying transfers, where state-of-mind does not reach. Same blind spot, same family, different capability tiers. Some failures track lineage, not capability — which is exactly why a second opinion from a different family is worth more than one more pass from a sibling model.

What to look at next

This is where the work earns its keep — the part a leaderboard never reaches. We are a lab, and the question we actually care about is never “what was the score,” it is “what did the behavior reveal, and what does it tell us to look at next.” The four mechanisms are not a gallery of mistakes; they sort into the pattern above, and the haiku result poses the real next question. opus-4.8 missed at the top of the stack — it knew everything and a single premise overrode it. haiku missed at the bottom of the back half — it held the taxonomy it plainly knows, then lost the thread across the long dependent resolution. Those are different failures, and the second one is the more suggestive.

The hypothesis worth chasing: is the small model failing because the problem demands tracking too many interdependent chains at once — so it activates too many nodes and drops a resolution it would get right in isolation? That is a claim about where reasoning-under-load gives out, not about what the model knows. And it suggests a concrete next probe: vary the depth and the interdependence of the chain — shorten it, lengthen it, sever some dependencies and tighten others — and watch where each model's resolution collapses. If the collapse tracks chain length and coupling rather than doctrine, the load hypothesis gains ground — and the capability-ladder pattern from the chart earns its promotion from observed to tested. That is the experiment this issue points at, and it is the kind of question this publication exists to ask.

The contribution is not the headline. It is the mechanism, and the next question the mechanism opens.

Be honest about what we have and have not shown. We have not isolated the mechanism — one task, twelve runs, a single judge's rationales. What we have is a pattern with a testable shape and a clear next experiment, and that pursuit is the point. And to be equally clear about what this is not: none of it says a model can stand in for a lawyer. Clearing a perpetuities chain is reasoning under load, not the judgment of practice. The chain is a good instrument precisely because its dependencies are real — which is also why the place it breaks tells us something worth writing down.

Previous · Page I Everyone passed — now the real question Next · Page III Cost & trust

‡ Every per-model per-link held / broke / partial determination, every memo score, and the corrected headline reads trace to the EV-023 chain-break failure map — the projection of the claude-opus-4-6 LLM-judge rationales for task property-fee-tail-rap across twelve channel-corrected runs (canonical per model; three drawn from the predecessor EV-003 law-essays campaign). Roll-up: N = 12 · classification held across the frontier · 7 / 12 held the full chain a–l · five broke — three in the back-half RAP cascade (gpt-5.4, opus-4.8, haiku-4.5), two in the front-half classification (grok-4.3 at b, deepseek-v3.2 at e). The full battery, rubric, judge, and channel correction are described in the methodology.