Perfect 20/20 4
Price spread 54×
Cheapest 20/20 $0.18
GunnerBench · Issue No. 1 About · The Ronin Advisory Group LLC

Issue No. 1 · June 2026

Everyone Passed. Now What?

In about two years, expert legal reasoning went from something only the most advanced AI could do to something nearly all of it can. That part is settled. The more interesting question is what happens inside the reasoning chain itself — where it holds, where it snaps, and why.

GunnerBench is a lab that measures how AI models actually behave on hard, real tasks — a workbench for deciding which model to trust with work that carries an obligation. This is its first dispatch. Each issue runs a different experiment; this one puts the June 2026 frontier — the most advanced AI systems on the market, the leading edge of what the technology can currently do — on a single hard legal-reasoning task. One kind of test among many the bench will run.

Four of twelve frontier models scored a flawless twenty out of twenty on bar-exam-grade essays they had never seen — expert legal reasoning is now table stakes. So we stopped asking who passed. We followed each reasoning chain link by link to find where it snapped — most memorably in the flagship that held every structural step, then committed to a single wrong premise and drew a flawless, confident diagram of a conclusion that was exactly backwards. A leaderboard cannot show you that. This issue does.

The failure modes are the story. Cost and trust are one downstream axis; where the chain breaks, and why, is the payoff.

Where the reasoning chain broke

The hardest of the three tasks is a property problem, and it is the centerpiece of this issue. A will hands a family orchard down a chain of conditions — to a son and the heirs of his body, then to a future widow for life, then to the grandchildren who reach twenty-five, then to a charity — each taking only if the one before it fails, some of them not yet born. Answering it is not one judgment but roughly a dozen interdependent ones in sequence: first name every present and future interest in that chain (what lawyers call “identifying the interests”), then test each contingent one against the Rule Against Perpetuities — the doctrine students remember as the hardest they ever met, which asks not what will happen but what could, under any turn of events the words permit. Miss one link early and every conclusion downstream inherits the error.

Every model in this field can now name the legal interests correctly — that part is fully saturated. Where they differ is the hard multi-step reasoning that follows: whether a model can hold a long chain of dependent conclusions together all the way to the title, or whether it drops a node somewhere in the middle. Even the best sometimes produces a confident, polished wrong answer — not because it lacks knowledge, but because a single wrong commitment early in the chain carries forward into everything that follows. That is the pattern this chart shows.

Two numbers in this issue measure different things, so hold them apart. Holding the full reasoning chain is not the same as a perfect score: a model can hold every property-link in the chain and still drop a point — on the required diagram, or on one of the other two essays. So seven models held the full chain on the property problem; of those seven, four also scored a perfect 20/20 across the whole battery.

Where the thirteen-link rubric chain snapped across twelve models. A shared dependency spine of thirteen rubric links labelled a through m runs left to right; links a through f are the classification half, links g through l are the back-half Rule-Against-Perpetuities cascade, and link m is an optional unscored footnote drawn muted and dashed. A dashed vertical divider sits between f and g. At the top, a teal band spans a through l: the seven models that held the entire chain. Below it descend five rupture lanes in capability-ladder order. grok-4.3 ruptures early at b, a hollow ring with a snapped spine, the lone front-half structural break and the ladder's bottom rung. deepseek-v3.2 ruptures at e. gpt-5.4 ruptures at g, and a thin arc carries the failure into the title link l. opus-4.8 holds all structure then ruptures at h and i, with two arcs converging on l. haiku-4.5 ruptures at i with a partial at h, and an arc into l, plus partial half-rings at j and k. Held nodes are solid teal discs; broken nodes are hollow rings stroked red with a gap in the spine; partial nodes are half-filled rings. a b c d e f g h i j k l m classification (a–f) RAP resolution (g–l) Held the full chain a–l (seven models, memo 5): claude-fable-5, claude-opus-4.7, gpt-5.5, deepseek-v4-pro, claude-opus-4.6, gemini-3.5-flash, gemini-3.1-pro. 7 models held a–l memo 5 / 5 classification saturated across the frontier grok-4.3, link b BROKE: calls the post-fee-tail interests executory interests, not remainders. An early, structural classification mis-step — the ladder's bottom rung. grok-4.3, link f PARTIAL: measuring-lives identification muddled. grok-4.3, link k PARTIAL: Octavia-as-measuring-life muddled. grok-4.3 memo 3 / 5 early & structural — foundational miscategory, but the RAP conclusions held bottom rung deepseek-v3.2, link e BROKE: classifies the Society gift as an executory interest. A near-miss; the core RAP result essentially held. deepseek-v3.2, link h PARTIAL: no explicit all-or-nothing (Leake) citation. deepseek-v3.2, link k PARTIAL: Octavia muddled. deepseek-v3.2 memo 4 / 5 near-miss — one misclassification, core RAP result essentially held gpt-5.4, link g BROKE: concludes the widow's life estate VOID when it is VALID (it vests at Arthur's death). The single wrong decision — the originating break that cascades into a second broken link at l. gpt-5.4, link l BROKE: wrong resulting title — a consequence of the wrong determination at g. gpt-5.4 memo 3 / 5 one wrong decision (g) cascades into a second broken link (l) claude-opus-4.8, link h BROKE: flips grandchildren VOID to VALID under one wrong premise (post-fee-tail remainders RAP-exempt). Knowledge intact; a single wrong commitment overrode it. claude-opus-4.8, link i BROKE: flips the Society gift VOID to VALID under the same wrong premise. claude-opus-4.8, link l BROKE: wrong resulting title — carried from the flipped h and i holdings. claude-opus-4.8 memo 2 / 5 held all structure (a–g), then one wrong premise detonates h, i and the title l haiku-4.5, link h PARTIAL: never identifies the afterborn mechanism; all-or-nothing not invoked. haiku-4.5, link i BROKE: calls the Society's gift VALID. Working-memory overload across the long downstream cascade. haiku-4.5, link j PARTIAL: mishandles the fertile-octogenarian trap. haiku-4.5, link k PARTIAL: Octavia-as-measuring-life mishandled. haiku-4.5, link l BROKE: wrong resulting title — cannot carry the full dependent resolution at once. haiku-4.5 memo 2 / 5 holds the taxonomy (a–g) but cannot carry the long dependent resolution (h–l) capability ladder ↓ early-structural to late-conclusional
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).

The teal band marks the seven models that held the full property chain (a–l); of those seven, four also scored a perfect 20/20 across the whole battery. Holding every link is not the same as a perfect grade — a chain-holder could still drop a point on the diagram or another essay.
held = solid teal node · broke = hollow ring + snapped spine · partial = half-filled ring · m = optional/unscored (muted dashed)
cascade arc = an upstream break (g/h/i) propagating into the title link l · rows ordered by break location = the capability ladder

View the full chain-break diagram →

Read the lawyering yourself

A score is a claim. The work is the evidence. Read the lawyering yourself — here is what a 20/20 memorandum on a Rule-Against-Perpetuities problem actually looks like, in the model's own words.

claude-fable-5 · 20/20 · verbatim excerpt # MEMORANDUM — RE: In re Estate of Helena Cartwright, Construction of Article IV (Blackacre). Preliminary Note on Governing Law: Franklin is a common-law jurisdiction that (1) has never abolished the fee tail, (2) retains the common-law Rule Against Perpetuities in its classical, what-might-happen form — no wait-and-see, no USRAP, no cy pres reformation, (3) applies the conclusive fertility presumptions (fertile octogenarian, unborn widow), and (4) construes class gifts under the rule of convenience.

The beautiful wrong answer

Not every confident memo is a correct one. The clearest exhibit in this issue is a model that wrote a polished, well-drafted analysis — and got the law exactly backwards, while scoring a perfect five for drawing the wrong conclusion crisply.

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

“a remainder limited after a fee tail is not subject to the Rule Against Perpetuities. That retention is dispositive.”

The judge's verdict: it invents a fee-tail RAP exemption, holding the grandchildren's and the Society's gifts VALID where the rubric requires them VOID — and the diagram scored a perfect 5 for rendering that wrong conclusion crisply.

A polished, confident wrong answer is the one most likely to be waved through. Where that failure comes from — and how it climbs the capability ladder across the field — is Page II.

In this issue

I Everyone passed — now the real question How expert legal reasoning became a commodity in two years Narrative
II Where the chain broke The thirteen-link rubric spine (twelve scored, a–l; m an optional footnote), the five rupture lanes, and what failure modes reveal Narrative
III Cost & trust When capability is flat, price and provenance are the whole decision Narrative
IV The exhibits (index) An index, not a narrative page: the actual memos and diagrams for the Property problem; all twelve models’ answers, read in full Index
V Methodology & colophon The battery, the rubric, the judge, and channel correction Method

‡ The battery: three bar-exam-grade essays — Property (fee tail + Rule Against Perpetuities, with a required hand-generated estate-chain diagram), Criminal Procedure + §1983, and Evidence (hearsay + Confrontation Clause). Same tasks, same grading checklist, scored by a separate strong AI model (claude-opus-4-6) that reads each answer against that checklist. Anthropic models were run on their own native service (claude-code); all other models (marked ‡ in the full scoreboard) were run through a shared service called OpenRouter, with costs corrected so every model is compared on the same basis.

Capability stopped being the variable.

© 2026 The Ronin Advisory Group LLC