Ambiguity Is a Verification Problem
When legal language supports more than one reasonable interpretation, the uncertainty can change the requirement being verified and therefore the result.
In a 2019 Delaware merger dispute, the Court of Chancery encountered an earn-out provision that it could not reliably parse and even make sense of. After repeatedly reading both the contract and the parties’ proposed constructions, the court concluded:
“Because the Court cannot understand the contract, it cannot endeavor to construe it.”[1]
The case was a bit extreme, but the underlying problem is common. Verification assumes that there is something sufficiently well defined to verify. A provision identifies a requirement, the relevant facts are established from the available evidence, and those facts are evaluated against the requirement.
Natural language complicates formal verification. The same provision can support more than one reasonable interpretation, and those interpretations can correspond to different requirements. This creates a form of uncertainty that is different from missing evidence. The facts may be known but the uncertainty may instead concern which proposition the contractual language requires us to test against those facts. For verification, ambiguity can therefore be a problem of specification.
Ambiguity can change the specification of the check
Consider a simplified provision:
The Seller shall deliver the financial statements
and tax records prepared after January 1.
The phrase prepared after January 1 has at least two plausible scopes. Under one interpretation, it modifies both document categories:
financial statements prepared after January 1
AND
tax records prepared after January 1
Under another, it modifies only the second:
all financial statements
AND
tax records prepared after January 1
The evidence has not changed, nor the documents in the data room. What changes is the set of documents required by the provision. Suppose a financial statement prepared before January 1 was not delivered. Under the first interpretation, that document falls outside the requirement. Under the second, its absence is not insignificant.
Retrieving additional evidence does not resolve this difference. Before the evidence can be evaluated, the requirement itself must be specified. This is not merely a constructed linguistic example. Acquisition agreements have produced almost exactly this problem.
An actual asset purchase agreement had the same problem
In Concord Steel, Inc. v. Wilmington Steel Processing Co., the Delaware Court of Chancery considered restrictive covenants in an Asset Purchase Agreement. The dispute depended in part on the APA’s definition of Competitive Business.[2]
The definition contained multiple clauses followed by modifying language. The court observed that, because of a “dearth of punctuation,” it was unclear whether modifying phrases near the end of one subparagraph applied to both preceding clauses or only the second.
The same definition created a temporal-scope question. Language referring to activities engaged in by specified entities “prior to and after” the effective time admitted questions about which entities and periods the modifier governed.
These are not cosmetic drafting issues. Different scopes produce different contractual sets. Conceptually:
Interpretation A
The modifier applies to clauses (a) and (b).
Interpretation B
The modifier applies only to clause (b).
The underlying business conduct can remain fixed while the legal question changes.
Concord Steel also involved disagreement about what counted as a business “competitive with” the relevant type of business. A narrow interpretation could focus on competition for the same specific work, while a broader interpretation could encompass rivalry for the same customer or market. The linguistic ambiguity therefore propagated into the verification task:
Reading A
Competition for the same specific work.
Reading B
Rivalry for the same customer or market.
The evidence about what the company actually did may be identical under both branches. The status changes because the specification changes.
Interpretive uncertainty is different from evidentiary uncertainty
Separating the two forms of uncertainty makes the difference clearer. Evidentiary uncertainty occurs when the requirement is sufficiently clear but the available record does not establish whether it was satisfied. For example:
Requirement:
Document X must be delivered at closing.
Evidence:
Closing occurred.
Delivery of Document X cannot be established.
Result:
Unresolved.
The proposition being tested is stable, but the uncertainty is factual.
Interpretive uncertainty occurs when more than one reasonable proposition can be derived from the governing language.
Interpretation A:
Documents X and Y are required only if condition C applies.
Interpretation B:
Document X is always required.
Document Y is required only if condition C applies.
A complete factual record does not necessarily resolve the second problem. The same facts can produce different determinations because the interpretations specify different checks. A verification result is therefore meaningful only relative to the proposition that was actually evaluated.
One provision can produce multiple verification branches
Material ambiguity can be represented as a branch in the analysis.
Interpretation A
The modifier applies to clauses (a) and (b).
Interpretation B
The modifier applies only to clause (b).
Reporting only Supported would hide the dependence on Interpretation A. Reporting only Unresolved would discard the fact that the evidence is sufficient under another reasonable reading. The more informative result is:
Verification result:
Interpretation-dependent.
This leads to a practical definition:
For verification, an ambiguity is material when two reasonable interpretations change either the check being performed or the result of that check.
The criterion is branch divergence. Two linguistic interpretations may differ without creating a verification problem if both yield the same relevant requirement and the same result. Conversely, an apparently small difference in scope, reference, time, negation, or qualification can be material when the resulting branches diverge.
Not every possible parse creates a reasonable branch
Natural language is pervasively ambiguous. A system that generated every grammatically imaginable interpretation would create enormous amounts of noise. Contract law supplies an important constraint: competing interpretations generally need to be reasonable.
ITG Brands, LLC v. Reynolds American, Inc. illustrates the point.[3] The case arose from an approximately $7.1 billion Asset Purchase Agreement involving cigarette brands. A provision required ITG Brands to use reasonable best efforts to reach agreements with previously settled states concerning the obligations it would assume.
The disputed provision ended with the words prior to the Closing. ITG argued that the phrase limited the duration of its reasonable-best-efforts obligation, meaning that the obligation terminated at closing. Reynolds read the phrase as modifying a much nearer portion of the provision: the description of the obligations ITG was to assume.
Both readings could be articulated grammatically. The Court of Chancery nevertheless rejected ITG’s construction and applied the nearest-reasonable-referent canon. The relevant temporal modifier naturally attached to the immediately preceding language, rather than to the reasonable-best-efforts phrase roughly fifty words earlier.
This difference is helpful in verification. It isn’t so much:
Can another syntactic parse be generated?
The question is closer to:
Is there another reasonable interpretation
that changes the verification task or result?
A theoretically possible reading is not necessarily an analytical branch worth preserving. This prevents ambiguity analysis from degenerating into unrestricted parse enumeration.
Pernicious ambiguity: when both sides think the text is clear
Some of the most consequential ambiguity may be difficult to detect precisely because an interpretation feels obvious. Lawrence Solan calls one form of this problem pernicious ambiguity: different interpreters regard a text as clear while assigning different meanings to it.[4]
That phenomenon is especially relevant to verification. If both reviewers recognize that a provision is ambiguous, the uncertainty is visible. They can preserve the alternatives or ask which interpretation should control.
The more dangerous case is:
Reviewer A:
"This clearly means P."
Reviewer B:
"This clearly means Q."
Both may reason coherently. Neither experiences their own interpretation as an assumption. The hidden branch occurs before the factual analysis begins. This produces a peculiar failure mode:
If only one branch becomes visible, a downstream result can look objective even though it depends on an unreported interpretive choice.
Coherent interpretation can conceal the branch
Large language models make this problem especially interesting because they are very good at generating coherent continuations from an adopted interpretation. A model can encounter ambiguous language, select a reading, and reason consistently from it. The resulting answer may be fluent, precise, well cited, and internally coherent. Coherence does not establish uniqueness.
Research on language models has begun to examine this issue directly. Kim et al. found that language models can struggle to handle inputs admitting multiple interpretations and proposed methods for encouraging explicit recognition of perceived ambiguity rather than simply answering through it.[5]
Zhang et al. found something even more intriguing: signals associated with question ambiguity can be represented internally even when outward model behavior remains confident. They identified a small set of neurons carrying strong ambiguity signals and showed that manipulating those signals could shift behavior from direct answering toward abstention.[6]
The practical implication for verification is narrower than the general question of whether a model “understands” ambiguity.
The concern is that:
coherent answer
does not imply:
unique reasonable specification
The model may have followed one branch correctly while failing to expose another branch that would change the check.
Contract ambiguity is already being treated as a requirements problem
There is also a useful connection between contract analysis and software requirements engineering. Singhal et al. treat commercial contracts as sources of project-specific requirements and study the generation of clarification questions for ambiguous contractual language.[7] Their ConRAP system was evaluated on contracts drawn from the CUAD dataset. The authors report an F2 score of 0.87 for ambiguity detection and that human evaluators regarded 70% of generated clarification questions as useful.
The framing is interesting because it treats ambiguity as a problem that occurs before downstream requirements can safely be operationalized. That is closely related to verification. If contractual language can generate more than one reasonable requirement, the uncertainty cannot be repaired merely by improving evidence retrieval after the fact. The ambiguity exists upstream of the evidence evaluation.
The chain is:
- SourceContract language
- InterpretRequirement specification
- TestEvidence evaluation
- OutputDetermination
Ambiguity near the top can propagate through everything below it.
Ambiguity can span multiple deal documents
The interpretive branch is not always contained in a single sentence or even a single document. Energy Transfer, LP v. The Williams Companies, Inc. involved a failed multibillion-dollar merger and an ambiguity created by the interaction between a Merger Agreement and a Parent Disclosure Letter.[8] The Merger Agreement imposed interim operating restrictions. The Parent Disclosure Letter contained exceptions to those restrictions, including an exception permitting up to $1 billion in certain equity issuances.
The disputed exception appeared beneath a heading corresponding to a particular subsection of the Merger Agreement. That structure supported a narrow reading: the exception applied only to the corresponding subsection. But the documents contained additional language pointing in the other direction. The disclosure letter stated that its headings were for reference and should not affect interpretation. The Merger Agreement also provided that disclosures could qualify other sections where their relevance was reasonably apparent.
This produced two reasonable branches:
Interpretation A:
$1 billion exception
-> applies only to corresponding subsection
Interpretation B:
$1 billion exception
-> can qualify Section 4.01(b) more broadly
The Delaware Supreme Court agreed that the relevant provisions were ambiguous. The consequences were not academic. Under the broader interpretation, Energy Transfer’s challenged conduct would have been excused and Williams would not have been entitled to a $410 million reimbursement. Under the narrower interpretation, which was ultimately accepted after considering extrinsic evidence, the exception did not excuse the conduct.
Again:
same transaction
same challenged conduct
same documents
different reasonable specification
different result
This case also exposes a limitation in document-by-document analysis. The ambiguity did not arise solely from either the words of the Merger Agreement or the disclosure letter. It emerged from the rules governing how the two documents interacted. Interpretive uncertainty can therefore be cross-document uncertainty. For verification against a deal record, the operative proposition may depend on relationships among the agreement, schedules, disclosure letters, amendments, definitions, side letters, and other controlling material.
Ambiguity can be intentional
Ambiguity in legal language should not automatically be treated as a drafting defect. Contracts are negotiated instruments. The parties may have incomplete information, competing interests, limited bargaining time, or reasons to preserve flexibility. Trying to resolve every possible future interpretation can itself impose negotiation costs.
Sometimes ambiguity helps make agreement possible. The Delaware Court of Chancery recognized this explicitly in United Rentals, Inc. v. RAM Holdings, Inc., a dispute arising from a roughly $7 billion merger agreement.[9] The court confronted conflicting remedial provisions and found that both sides had advanced reasonable interpretations.
In discussing the drafting, the court observed that parties sometimes “riddle their agreements with a certain amount of ambiguity in order to reach a compromise.” That possibility is critical for verification. The objective is not:
eliminate ambiguity
or:
force software to select one interpretation
The verification objective is:
identify when the verification result depends
on an interpretive choice
An intentionally unresolved provision can still generate divergent verification outcomes. Preserving that dependency is compatible with leaving the legal ambiguity unresolved.
Facts and interpretations should not be conflated
Hidden interpretive choices can create a misleading appearance of factual disagreement. Consider two reviewers evaluating the same provision against the same deal documents. Both agree on every relevant fact. Both reason correctly from the interpretation they adopted. One reports:
Supported
The other reports:
Contradicted
If the interpretive branch is invisible, the disagreement appears to concern the evidence. It does not. The reviewers are evaluating different propositions. For auditability, a verification record should make it possible to delineate at least four causes of a changed result:
- the underlying evidence changed;
- a factual extraction changed;
- the interpretation of the governing language changed; or
- a different portion of the requirement was evaluated.
Those are materially different events. A factual change affects the premises whereas an interpretive change can alter the proposition against which those premises are evaluated. Collapsing the two makes it harder to explain why a determination changed.
Verification should preserve consequential branch points
None of this requires software to determine the legally correct interpretation of every ambiguous provision: applicable law or negotiating history or commercial practice or extrinsic evidence may be the most important given the circumstances. Even where some questions require legal judgment, verification can still establish more tightly scoped computational facts:
The result changes depending on which
reasonable interpretation is applied.
That information is valuable before the interpretive question is resolved. A verification record might therefore preserve:
Interpretation A
Requirement: ...
Evidence: ...
Result: Supported
Interpretation B
Requirement: ...
Evidence: ...
Result: Unresolved
The reviewer can then determine whether one interpretation should control, whether additional context collapses the branches, whether extrinsic evidence should be considered, or whether the ambiguity itself requires legal analysis.
This is the key point:
Resolving the ambiguity
!=
Preserving its effect on the analysis
A verification system can do the second without having authority to do the first.
Ambiguity changes what a verification result means
A status such as Supported looks categorical. But every verification result is relative to some specification of the proposition being tested. When that specification is uncontroversial, the dependency may require little attention. When the governing language supports materially different readings, the interpretation becomes part of the result’s provenance. A reviewer may therefore need to know not only:
What evidence supported this determination?
but also:
Which interpretation made that evidence sufficient?
The same record can support different conclusions when the underlying proposition changes. Verification should preserve uncertainty at the level where it occurs:
Missing evidence
-> evidentiary uncertainty
Requirement never checked
-> coverage uncertainty
Multiple reasonable specifications
-> interpretive uncertainty
The categories should not be collapsed into a single generic uncertain state, and interpretive uncertainty should not silently disappear into a confident answer.
The central question isn’t so much whether legal language contains ambiguity; it inevitably does. The question for verification is whether another reasonable interpretation causes the analytical branches to diverge. When it does, the interpretation is part of the verification result, and ambiguity is no longer merely a defect to eliminate; it becomes a structured verification problem to preserve and resolve.
References
[1] Western Standard, LLC v. SourceHOV Holdings, Inc. & Pangea Acquisitions, Inc., C.A. No. 2018-0280-JRS (Del. Ch. July 24, 2019). ↩
[2] Concord Steel, Inc. v. Wilmington Steel Processing Co., Inc., C.A. No. 3369-VCP (Del. Ch. Apr. 3, 2008). ↩
[3] ITG Brands, LLC v. Reynolds American, Inc., C.A. No. 2017-0129-AGB (Del. Ch. Nov. 30, 2017). ↩
[4] Lawrence M. Solan, Pernicious Ambiguity in Contracts and Statutes, 79 Chicago-Kent Law Review 859 (2004). ↩
[5] Hyuhng Joon Kim et al., Aligning Language Models to Explicitly Handle Ambiguity, Proceedings of EMNLP 2024, 1989–2007, DOI: 10.18653/v1/2024.emnlp-main.119. ↩
[6] Zhuoxuan Zhang, Jinhao Duan, Edward Kim & Kaidi Xu, Sparse Neurons Carry Strong Signals of Question Ambiguity in LLMs, Proceedings of EMNLP 2025, 16081–16099, DOI: 10.18653/v1/2025.emnlp-main.813. ↩
[7] Anmol Singhal, Chirag Jain, Preethu Rose Anish, Arkajyoti Chakraborty & Smita Ghaisas, Generating Clarification Questions for Disambiguating Contracts, Proceedings of LREC-COLING 2024, 7611–7622. ↩
[8] Energy Transfer, LP v. The Williams Companies, Inc., No. 391, 2022 (Del. Oct. 10, 2023). ↩
[9] United Rentals, Inc. v. RAM Holdings, Inc., C.A. No. 3360-CC (Del. Ch. Dec. 21, 2007). ↩
Disclaimer: This article is for informational purposes only and is not legal advice. For advice about a specific matter, consult qualified legal counsel.