Most information systems are designed to produce a coherent answer. Contentious work often requires a different discipline. A lawyer may need to preserve two incompatible dates, three competing accounts and an unexplained gap rather than compress them into a tidy narrative.
That is why contradiction and incompleteness should be treated as first-class outputs in litigation intelligence. A system that only finds material supporting the current proposition can make a case appear cleaner than the record actually is.
1. Contentious work is not a confirmation exercise
Imagine that one email records notice on 3 March, a later minute refers to 7 March, a witness remembers “early March” and metadata suggests that an attachment was not created until 8 March. A summary engine may be tempted to state that notice was given “in early March”. That is linguistically efficient and evidentially destructive.
The professional output should preserve the conflict: who says what, which source supports each date, what the source status is and why the divergence matters.
2. Adverse material is not an edge case
Practice Direction 57AD provides a useful litigation analogue. Within its scope, an adverse document includes material that contradicts or materially damages a party’s contention or supports the opposing party’s version on an issue in dispute. The Practice Direction also imposes continuing duties concerning known adverse documents.
Legal AI retrieval is not disclosure and should not be described as though PD57AD directly regulates every retrieval workflow. But the procedural principle is instructive: evidence that damages the current theory can be as important as evidence that supports it.
3. Search against the proposition as well as for it
If the proposition is “the board had approved the transaction by 10 June”, a useful investigation should not stop after locating approving language. It should also look for later requests for approval, drafts marked unapproved, correspondence showing conditions remained outstanding, or documents suggesting a different decision-maker.
This is a form of structured challenge. The purpose is not to make the AI adversarial for its own sake. It is to help the lawyer avoid treating the first coherent evidence trail as the only one.
4. Qualifying evidence is different from contradiction
Some evidence does not directly contradict a proposition but narrows it. An email may confirm that a decision was made while leaving the date uncertain. A witness may remember attendance at a meeting but not the words used. A board minute may record approval “subject to finance”.
Those qualifications should survive the analytical process. A binary supports/does-not-support model is too crude for many disputed facts.
5. “Not found” is not “did not happen”
This is one of the most important disciplines in document-heavy litigation. If a search does not locate an approval email, the defensible statement is often:
No approval email was identified in the material searched.
It is not:
No approval email exists.
The second statement silently assumes that the corpus is complete, preservation was adequate, all relevant repositories were collected, the search was sufficient and approval would necessarily have taken the form of an email.
PD57AD itself recognises the significance of missing known adverse documents: where such a document cannot be located, that fact must be disclosed within the applicable framework. The broader lesson is that absence can become a question to investigate rather than a gap for software to fill.
6. Three different absences should stay different
- Not in the retrieved passages: the retrieval step did not surface it.
- Not identified in the matter corpus: the wider indexed material was searched but the item was not found.
- Does not exist: a factual conclusion that normally requires a stronger evidential basis.
Collapsing these into one statement overstates what the technology knows.
7. Contradiction can drive the next legal task
A useful matter system should turn a conflict into work. A disputed date may justify checking metadata. A missing approval may justify identifying a custodian or repository. An inconsistency between a witness recollection and contemporaneous messages may require further interview or advice. The output is not necessarily a conclusion; sometimes it is a better question.
A practical review pattern
For every material proposition, ask: what supports it; what qualifies it; what contradicts it; what expected material has not been found; what part of the matter was actually searched; and what professional step follows from the uncertainty?
8. Temporal contradiction deserves special attention
Some of the most important contradictions are chronological. An account may require a person to know something before the email said to inform them was sent. A decision may be described as final even though later documents continue to seek approval. A witness may place a meeting before the document that appears to have triggered it.
These are not merely date errors. They can affect knowledge, causation, reliance, limitation, credibility and the plausibility of the overall account. A structured chronology can therefore act as a contradiction detector—not because software proves which account is true, but because sequencing makes tension easier to see.
9. Gaps should create tasks, not invented bridges
When the record jumps from one event to another without the expected intermediate document, the system should resist filling the narrative. Instead, the gap can be converted into an investigation task: identify the likely custodian, check attachments, search an alternative repository, confirm whether the decision was oral, or ask whether the document was ever created.
This is a productive use of uncertainty. It changes the output from “AI could not answer” to “the matter requires this next evidential step”.
10. Contradictions should not be scored away
A single confidence score can obscure the character of a dispute. Two strong but incompatible sources do not necessarily produce a medium-confidence truth. They may produce a high-confidence conclusion that the evidence is genuinely conflicted.
For professional review, the system should preserve the competing propositions and the basis for each. Resolution may depend on credibility, context, disclosure, cross-examination or legal burden—questions that cannot responsibly be collapsed into a model score.
11. The lawyer decides when the conflict is material
Not every inconsistency matters. A one-day difference may be irrelevant in one case and decisive in another. The system can surface the conflict; the lawyer determines whether it changes the issue analysis, requires further evidence or affects the advice.
That division of labour is important. Litigation intelligence should improve visibility of the record without pretending that materiality is a purely computational property.
12. Contradiction should be connected to issue significance
A long matter may contain hundreds of inconsistencies that are immaterial. The useful question is which conflict affects a live factual or legal issue. A contradiction about a meeting time may be trivial; a contradiction about whether a decision preceded consultation may be central.
This is where matter-centred organisation matters. The contradiction should be linked to the proposition and issue it affects, so that the lawyer can decide whether it changes the theory of the case, the evidence plan or the advice. The system can surface the tension; professional judgment determines materiality.
That approach also helps avoid sensationalising minor discrepancies. Litigation intelligence should not reward the number of contradictions found. It should help the team understand which unresolved conflicts genuinely require attention.
Conclusion
Contentious analysis becomes safer and more useful when the system is allowed to say “the evidence conflicts”, “support is partial” or “this has not been found”. Conflict and incompleteness are not failures of the answer. They are often the substance of the case.
Related LegalRAG Pro Insights
Professional context. This article discusses legal-technology workflow and professional-risk questions. It is not legal advice and should not be treated as a substitute for checking the current procedural, regulatory and factual position in a particular matter.