Large language models often answer questions about a document from memory instead of from the document. In legal work this produces the most dangerous kind of error: the model cites a real clause or case and attaches a holding or number it never contained. Current benchmarks measure this with a single score on fixed documents, or with counterfactual passages that another language model rewrote. Both approaches leave open the question a practitioner actually needs answered. If I change one fact in this contract, how far can I change it before the model stops tracking the change and falls back on what contracts like this usually say?
We propose to answer that by bringing back an idea from 1980s legal AI. Rissland and Ashley's HYPO system generated "hypotheticals" by taking a real case and sliding one legally meaningful fact along an axis, such as the number of people who know a trade secret, whose direction is known to strengthen or weaken a claim. We apply the same controlled edits to real contracts from CUAD, an expert-annotated contract dataset. Each seed contract becomes a trajectory of documents that differ in one known value. Because we made the edit ourselves, we know what the correct restated value is, what follows from it arithmetically, and which way any judgment should move. Most of the hallucination score can then be computed without a second model acting as judge. The result is a curve per model and per fact type, showing where its faithfulness to the text breaks.
Core Claim
A deformation along a single dimension with a known direction gives a hallucination test that is checkable by construction. On fixed documents, a correct answer and a memorized answer look the same. Along a trajectory they separate: a model that reads keeps tracking the edited value, while a model that recalls reverts to the seed value or to a typical value. We claim three things. First, this tracking measure can be scored mostly by exact match and arithmetic, so it agrees with expert labels at least as well as LLM-judge pipelines do, without inheriting the judge's biases. Second, the point along the trajectory where tracking fails (the "breakpoint") is a stable property of a model and a fact type, and it predicts failures on held-out contracts better than an aggregate hallucination rate does. Third, breakpoint curves can show directly whether larger models are more truthful or more stubborn, which is the tension between TruthfulQA and FaithEval.
Where This Sits
The method comes from the case-based reasoning literature. Hypotheticals as Heuristic Device defined dimensions with "focal slots," meaning editable values that have a direction. It also gave five generation heuristics, including incremental strengthening or weakening (H2), conflicting dimensions (H3), and extreme values (H4). Explaining and Arguing With Examples formalized case deformation as movement along these axes. Constrained Example Generation supplied the architecture of retriever, modifier, and judge. We reuse the judge as a constraint checker that keeps edited contracts internally consistent. None of this work was evaluated quantitatively at scale or applied to language models. We supply both.
Two benchmarks motivate the gap and are our closest competitors. HypoTermQA has GPT-3.5 invent nonexistent terms and finds error rates above 94% for GPT-3.5 and Llama2-70B. Its probes are isolated words with no structure and no graded distance. FaithEval has GPT-4o rewrite contexts into counterfactual, unanswerable, or inconsistent versions. GPT-4o's accuracy falls from about 96% to 47.5% under counterfactual contexts. Its perturbations are binary, though, and the rewriting is itself model-generated. Our edits are deterministic, graded, and attributable to a single slot.
Two legal studies show that aggregate scores hide the failures that matter. LegalHalluLens audits CUAD extractions by claim type. It finds about 30% error on temporal claims against about 70% on numeric and obligation claims, and two models with the same 52% aggregate rate that lean in opposite directions (inventing versus omitting). Hallucination-Free? shows that commercial legal tools built on retrieval-augmented generation (RAG, meaning the model answers from documents fetched for the query) still hallucinate 17% to 33% of the time, often by citing real sources with fabricated content. Both studies use static inputs. We keep their typed view and their concern with real-source fabrication, and we add the deformation axis.
One mechanistic work supplies a secondary probe. In-Context Sharpness as Alerts shows that hallucinated tokens come with blurry, high-entropy internal activations over the prompt. It has only been tested on static question answering. We ask whether that entropy rises with deformation distance.
The broader motivation is the pattern described in The Empirical Validation Gap in Systems Research, where systems are shown to work through demonstrations rather than reproducible numbers. Classic hypothetical generation is one example of that pattern, and current LLM-judged hallucination benchmarks are another.
Research Program
Dimensions. We specify about fifteen contract dimensions tied to CUAD clause categories, concentrated on the numeric and obligation types where LegalHalluLens found the worst errors. Examples include the liability cap amount, notice period, non-compete duration and radius, renewal term, the number of parties permitted to receive confidential information, and the minimum purchase commitment. For each dimension we record three things: the CUAD span types where its value appears, a validity constraint (for instance, the notice period must be shorter than the term), and a direction, meaning which way the value moves a party's position. A small panel of lawyers confirms the directions. Dimensions where they disagree are excluded from the judgment-level tests.
Edits and trajectories. The modifier rewrites every occurrence of a slot with templates, not with a language model. The judge rejects any edit that violates a constraint. From each seed contract we build three kinds of trajectories. Incremental trajectories (H2) step a value through K levels on a log scale. Extreme trajectories (H4) push it to degenerate values such as zero or ten million. Conflict trajectories (H3) move two dimensions in opposing directions. Near-miss edits (H1) insert a clause that was absent, for example adding a non-compete to a contract that had none. This tests whether the model notices the new obligation rather than defaulting to "none."
Three question tiers. Restatement questions ask for the slot value and are scored by normalized exact match. Consequence questions ask for something computable from the slot, such as whether a stated damages figure exceeds the cap or which date a notice must be given by. These are scored against arithmetic ground truth. Judgment questions ask for a strength rating of a party's position and are scored by a metamorphic check: across a trajectory, ratings must move in the dimension's direction, and they should level off at extreme values. Only the judgment tier is ever scored with human or model judgment, and even there we score only the sign of change.
Metrics. For a trajectory with slot values $v_0$ (the original) through $v_K$, let $F(k)$ be the fraction of restatement and consequence answers consistent with $v_k$. Let $R(k)$ be the fraction consistent with $v_0$ or with the most frequent value of that dimension in CUAD. $R$ measures reversion to memory or to typical values, as distinct from random error. The breakpoint is
$$k^{*} = \min\{k : F(k) < \tau\},$$
where $\tau$ is a fixed fidelity threshold, set to 0.8 in advance. For judgment questions we report the monotonicity violation rate. This is the fraction of adjacent pairs $(k, k+1)$ whose ratings move against the known direction $d \in \{+1, -1\}$:
$$M = \frac{1}{K}\sum_{k=0}^{K-1} \mathbf{1}\left[ d\,(s_{k+1} - s_k) < 0 \right],$$
where $s_k$ is the model's strength rating at step $k$.
Every quantity is reported per dimension and per claim type, following the typed reporting argued for by LegalHalluLens. We also report the share of errors that are reversions ($R$) versus other errors. This share separates "recalled the prior" from "made something up," the distinction Hallucination-Free? draws between real-source fabrication and outright invention.
Models. The panel spans scale and tuning. It includes Phi-3-medium and Mistral-7B, the small models FaithEval found more faithful to context, along with Llama-2-7B and Llama-2-70B, Gemma-4-26B-A4B, GPT-4o, and Claude 3.5 Sonnet. This lets breakpoint curves be compared across both size and instruction tuning.
Baselines
The naive reference point is static evaluation on undeformed CUAD contracts. We ask the same three tiers of questions about original seeds and report an aggregate hallucination rate per model, which is what current leaderboards would report. It runs on the same seeds, models, and decoding settings. No reported result was found on Lacuna for question answering on undeformed CUAD with these models. The nearest number is the LegalHalluLens extraction profile of about 30% temporal error and about 70% numeric and obligation error. Deformation beats this baseline if it changes the model ranking, or splits models with indistinguishable static scores into significantly different breakpoints, and if those breakpoints predict held-out failures better (see decisive checks).
The strongest existing evaluation method is FaithEval-style unconstrained counterfactual rewriting, applied to the same seeds. GPT-4o is prompted to produce counterfactual versions of each contract, matched in number to our deformed items, and scored with an LLM judge as FaithEval does. FaithEval's reported anchor is GPT-4o at 47.5% accuracy on counterfactual contexts, against roughly 96% on originals (FaithEval). This is the right competitor because it is the cheapest way to produce counterfactual legal documents at scale. We give it the same generation budget, the same question tiers where it can support them, and the same human validation subset. Our method beats it if three things hold: a higher fraction of its items are judged valid and unambiguous by lawyers, its automatic scores agree more closely with expert labels, and it finds more distinct confirmed failures per thousand queries.
For typed auditing, the strongest existing system is the LegalHalluLens category breakdown and Risk Direction Index, run on static CUAD. We compute both on our deformed items. This tests whether deformation adds information beyond typing, or only reproduces the typed hierarchy.
On the mitigation side, we measure how far each of the following shifts the breakpoint $k^*$ when applied as-is: chain-of-thought prompting, calibrated multi-agent debate as in LegalHalluLens, and Activation Decoding from In-Context Sharpness as Alerts. LegalHalluLens reports that debate cut fabricated detections by 45% on CUAD extraction. Activation Decoding reports an 8.6-point gain in Truth*Info on TruthfulQA for Llama-2-70B and an average 16.1% gain in HotpotQA F1. No reported result was found on Lacuna for any of these mitigations on deformed or graded counterfactual inputs.
Two ablations isolate the key ingredient. The first, random perturbation, edits the same number of tokens by the same relative magnitude, but in non-dimensional numbers such as section numbers or exhibit dates that carry no legal direction. This tests whether failures come from any change to the document or specifically from changes to legally loaded facts. The second, endpoint-only, keeps only the seed and the extreme value (H4) and drops the intermediate steps. This tests whether the graded trajectory adds anything beyond a binary counterfactual.
Why not just…?
Why not just prompt GPT-4 to generate hard counterfactual contracts? This is the real threat, and we treat it as the main baseline rather than dismiss it. There are three reasons to expect it to fall short. The generator brings its own priors into the test, a bias HypoTermQA acknowledges in its own GPT-3.5-generated data. Its edits do not come with a known correct answer, so scoring falls back on a judge. FaithEval also reports that open-ended conflict prompting makes models sycophantic about conflicts that do not exist. The head-to-head comparison on item validity, agreement with experts, and failure yield settles the question. If unconstrained generation matches us on all three, the dimensional machinery is unnecessary, and we would say so.
Why not just use chain-of-thought prompting so the model reasons through the edited facts? FaithEval reports that chain-of-thought gives marginal gains on counterfactual contexts but leaves models defaulting to their priors. HaluEval finds it weak where facts must be combined across a document. We include it as a mitigation arm. If it pushes $k^*$ past the extreme end of every trajectory, the benchmark is too easy and needs harder conflict trajectories. That result is itself informative.
Why not just ground the model with retrieval? Here the deformed contract is the retrieved document, so our setting already is retrieval-grounded. The question is whether the model uses what it was given. Hallucination-Free? shows that grounded commercial tools still hallucinate 17% to 33% of the time, often by misreporting real sources. This is exactly the reversion that $R(k)$ measures.
Why not just report typed aggregate rates as LegalHalluLens does? Typing shows which fact types fail, not how much change a model tolerates before failing. Two models with identical type profiles could have very different breakpoints. The typed-auditing baseline tests this directly. If breakpoints add nothing to typed rates in predicting held-out failures, the contribution shrinks to cheaper scoring alone.
Proposed Experiments and Decisive Checks
The data are roughly 200 CUAD seed contracts (CUAD), each carrying five or more of our dimensions. With eight steps per incremental trajectory plus extreme, conflict, and near-miss variants, and three question tiers, this gives on the order of tens of thousands of scored queries per model. Lawyers will label a stratified subset of 600 items for validity and correct answers, drawn from both our generator and the unconstrained baseline.
The first decisive check is whether automatic scoring is trustworthy. On the human subset, restatement and consequence scoring should agree with expert labels at a rate at least equal to the LLM-judge pipeline. For reference, TruthfulQA reports 90–96% agreement between its judge and humans, and HypoTermQA's best evaluator error was 6.66%. If our exact-match scoring falls well below these, the claim that the method frees evaluation from judges fails.
The second check is predictive validity. Fit breakpoints on half of the seeds and predict per-item failures on the held-out half and on two held-out dimensions. This should beat static aggregate rates and typed rates under the same predictive model, measured by area under the ROC curve.
The third check is the endpoint-only and random-perturbation ablations. Graded legal-dimension trajectories should find failures that the endpoint-only set misses, especially in the middle of the range. Error rates should be significantly higher, and reversion more common, under directional edits than under random edits of the same size.
The fourth check is the scale question. Plot $R(k)$ against model size and tuning. FaithEval's "strong priors" account predicts that larger models reach their breakpoint earlier and that their errors are mostly reversion. An imitation account in the spirit of TruthfulQA predicts reversion to the CUAD-typical value rather than to the seed value. Our metrics distinguish these two predictions.
The final, exploratory check is whether contextual entropy from In-Context Sharpness as Alerts, measured on the slot-value tokens of open-weight models, rises monotonically with deformation distance. If it does, a model's breakpoint can be estimated from its internal activations without asking any questions.
References
- TruthfulQA: Measuring How Models Mimic Human Falsehoods
- FaithEval: Can Your Language Model Stay Faithful to Context, Even If "The Moon is Made of Marshmallows"
- Hypotheticals as Heuristic Device
- Explaining and Arguing With Examples
- Constrained Example Generation: A Testbed for Studying Issues in Learning
- HypoTermQA: Hypothetical Terms Dataset for Benchmarking Hallucination Tendency of LLMs
- LegalHalluLens: Typed Hallucination Auditing and Calibrated Multi-Agent Debate for Trustworthy Legal AI
- Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools
- In-Context Sharpness as Alerts: An Inner Representation Perspective for Hallucination Mitigation
- HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models
- CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review