On Case Intelligence, provenance, the boundary between detection and legal meaning — and what must remain within the professional judgment of the criminal defense lawyer
A criminal case can contain large volumes of documents, witness statements, memoranda, images, expert opinions, and multiple versions of the same event.
But volume is not the only problem.
Even when all the case materials are already in front of us, a harder question remains:
How do we turn all of that information into a picture we can actually think with?
That was one of the questions that occupied me when I began building a system for working with criminal case materials that also incorporates AI capabilities.
At first, it is easy to assume that the challenge is mainly technological: reading files faster, extracting information, searching, summarizing, or locating references across a large body of material.
As the work progressed, I realized that the more interesting problem lies elsewhere.
A criminal case is not simply a collection of information.
It contains different sources, people, events, versions, evidence, claims, dates, and contexts. It matters who said something, when it was said, on what basis, and whether another document is an independent source or merely repeats information that appeared elsewhere.
The same detail may look insignificant in one document and take on a different meaning when considered alongside another source.
That changed the question guiding my work.
Instead of asking only:
What can we make AI do?
I began asking:
What work should the system perform, where can AI assist, and which decisions should remain, from the outset, with the criminal defense lawyer?
A criminal case is not a folder of PDFs
The move from paper files to digital case materials has already changed criminal defense work significantly.
When material is searchable and properly organized, it becomes much easier to locate names, expressions, dates, and details that once required manually reviewing large volumes of documents.
That is a major improvement.
Suppose, for example, that a case includes an allegation that the defendant had a knife.
In a digital case file, I can search for the word “knife” and quickly find every place where it appears.
Search helps me find the relevant material.
But that is where a different kind of work begins.
Who was the first person to say there was a knife?
Did that person see it personally?
Is another witness describing it independently, or simply repeating what they heard?
Are all of the statements referring to the same object?
At what point in the event does it appear?
And if four documents mention a knife, do those four documents represent four independent sources — or one source repeated across three additional documents?
The problem is no longer only finding the items.
It is understanding the relationships between them.
This is where I use the term Case Intelligence.
Not in the sense that the machine “understands the case.”
I mean a state in which the case materials are organized in a way that allows the lawyer to work with the sources, people, events, and relationships between them — while still being able to return to the underlying material.
A summary can shorten material.
Search can locate material.
A knowledge system should help preserve context and make it possible to work with the relationships between pieces of information.
Who is the source? What was said? When? In what context? What connects to what? And what still requires examination?
The development I am exploring is not intended to replace search capabilities that already exist.
It is an attempt to build, on top of retrieval, a working environment that helps expose and organize the relationships between the things that have been found.
Before intelligence comes fidelity to the source
Before relationships between pieces of information can be built, there is a more basic question:
Is the material the system is working with actually the material we received?
Digital case materials pass through technical layers: files, archives, formats, filenames, encoding, text extraction, and sometimes system-generated elements that have no evidentiary meaning.
One of the decisions that gradually became more than a technical principle was the separation between source material and the working layers created from it.
The source remains the source.
Structured information can be created from it. Content can be extracted, organized, and linked. A new working environment can be built around it.
But none of those outputs replace the source material.
This distinction matters because a system can fail before reasoning even begins.
A file may be ingested only partially. Text may be extracted incorrectly. A name may be distorted. A document may fail to enter the index.
If a technical failure is not kept separate from legal information, it becomes surprisingly easy to assign meaning to a processing error that never existed in the case itself.
That led me to a simple rule:
A technical failure is not a fact in the case.
This becomes especially important when the system does not find something.
There is a fundamental difference between:
“The system did not find the information.”
and:
“The information does not exist in the case materials.”
The first describes a processing or retrieval result.
The second is already a claim about the case.
Before I can ask a system what it sees in the material, I first need to know:
What material, exactly, is it looking at?
An answer is not enough — we need to know where it came from
Preserving the source solves only part of the problem.
Suppose the system presents a piece of information.
A person said something.
An event took place at a certain time.
The same detail appears in several documents.
Now I need to be able to ask:
How do we know this?
This is where provenance becomes critical.
In my view, when working with legal material, the ability to return to the source is not an optional feature.
It is part of the ability to evaluate the output itself.
There is a difference between:
“X happened.”
and:
“A particular witness said that X happened.”
The second formulation preserves information the first one erased:
the source.
And from there, further questions follow.
Did that person see it personally?
Did they hear it from someone else?
Was the statement made in real time?
Does another document independently support the claim — or merely repeat it?
If the same claim appears in four documents, we can count four appearances.
But from an evidentiary perspective, there may still be only one source.
Quantitatively: four appearances.
In evidentiary terms: perhaps one source.
So before a system helps me decide what deserves examination, it must allow me to know exactly what I am looking at.
Provenance tells us where the information came from.
Only then can we move to the next question:
What about it deserves examination?
Detecting a difference is not the same as finding a contradiction
Suppose the system places two statements side by side and identifies a difference.
In the first version, a person describes the event one way.
In a later version, the description changes slightly.
There is a detectable difference.
But is it a contradiction?
Not necessarily.
Perhaps a different question was asked.
Perhaps additional details were added.
Perhaps the difference is immaterial.
And perhaps, of course, it is highly significant.
The system may help me see the difference.
The meaning of that difference is a different professional question.
The same principle applies throughout a criminal case.
A detail that appears in only one source is not necessarily suspicious.
An event that is absent from a particular document is not necessarily evidence that it never occurred.
A connection between two people is not necessarily relevant to the case.
And similarity between two descriptions does not necessarily prove a common source.
In each of these situations, there is a transition between two different operations:
Detection — what was found or what differs.
and
Legal Meaning — what the finding actually means.
That boundary should not be blurred.
The legal meaning of a finding is not a field hidden inside the document, waiting for the system to extract it.
It emerges from law, procedure, evidence, context, possible explanations, and strategy.
That is why one sentence keeps returning throughout my work:
AI may say: “This is interesting.”
It should not, on its own, say: “This proves it.”
A system that helps direct attention
This led me to another question.
Perhaps one of AI’s most valuable roles in criminal case work is not necessarily drawing conclusions.
Perhaps it can help organize the defense lawyer’s field of attention.
In a complex case, information may be scattered across many sources.
A version in one statement.
A date in another document.
A further detail somewhere else.
Each item may be completely visible on its own.
The difficulty is that we do not always see them together.
A system can help bring together information that may relate to the same event, place descriptions side by side, return the lawyer to the underlying sources, and surface possible connections that deserve examination.
Sometimes that arrangement is enough.
The system does not need to tell me what the information means.
When the material is placed correctly in front of the lawyer, the next question often becomes visible.
Return to the source.
Check whether the gap is real.
Ask why a certain detail is missing.
Or discover that something that initially looked unusual has a simple explanation once the broader context is restored.
That is where I began thinking about:
AI as an attention engine for the criminal defense lawyer.
Not a system that says:
“I found the decisive piece of evidence.”
But one that can say:
“There may be something here worth examining.”
Even that role, however, is not neutral.
What the system chooses to surface becomes more visible.
What it fails to find, connect, or rank highly may receive less attention.
So it is not enough to ask:
What did the system find?
We also need to ask:
What might it miss?
For me, that is one of the questions that requires much more work as these systems become more significant.
The criminal defense lawyer is not the machine’s quality-control layer
The term human-in-the-loop is sometimes used to describe a system that performs the work while a person remains at the end of the process to check that nothing went wrong.
In legal work, I see the relationship differently.
The criminal defense lawyer is not merely the system’s quality-control layer.
Some decisions are human from the beginning.
Not because the technology is not yet good enough.
Because they are a different type of work.
Is a difference between two versions actually significant?
What alternative explanations might there be?
What weight should be assigned to a piece of evidence?
How does it relate to the legal issue?
How does it fit with the rest of the picture?
And should it even be used as part of the defense strategy?
These are not simply information problems.
They require context, experience, law, procedural understanding, and strategy.
There is another important professional act that the system should not obscure:
the decision not to make an argument.
A point can be interesting and still be weak.
A gap can exist and still have a reasonable explanation.
A surprising connection can be identified and still be useless to the defense.
The ability to say “I found something” and the ability to say “this is something we should build an argument around” are not the same ability.
That is why I do not see human judgment as a safety layer added after the system has finished its work.
It is part of the professional architecture itself.
The question is not how much work we can hand over to the machine
This may be the central point that has become clearer to me during the development of the system.
The question is not:
How much of the defense lawyer’s work can be transferred to AI?
The more useful question is:
Where, within professional legal work, should computational and AI capabilities be placed — and where should they stop?
I currently see three distinct stages.
First, we need to know what we are looking at — which is why source integrity and provenance matter.
Then the system may help surface and organize what deserves examination.
Only after that do we reach the stage where someone must decide:
What does this mean, and what should be done with it?
That final stage is not simply another operation on information.
It is where professional judgment and legal decision-making begin.
My goal, therefore, is not to build an AI that becomes the criminal defense lawyer.
It is to build a better working environment around the defense lawyer — one that helps her see more, return to the source, examine connections, and ask more precise questions.
Without blurring the line between information and meaning.
Without turning a claim into a fact.
Without turning a difference into a contradiction simply because a system flagged it.
And without losing sight of the question that should remain at the center:
What, out of all of this, actually matters to the case?
That boundary is not, in my view, a limitation of the system.
It is part of how the system should be designed.
