I Started With a Legal Research Desk. It Turned Into a Chamber.

I started building CiteBench because legal research felt unnecessarily fragmented.

You have a legal question, you have the facts, and you need to find the law. That sounds simple until you actually do it. You move between case-law databases, judgments, PDFs, notes and browser tabs. You find a promising authority, then another. You try to work out whether the case actually says what you think it says. You copy citations somewhere. Then, when the research is finally done, you leave the research tool and go back to the actual matter — where the hearing, the previous order, the documents, the deadlines and the next piece of work are all sitting somewhere else.

The first version of CiteBench was built to make that research process better.

Then I started thinking about what happens before and after the research.

That changed the product completely.

Imagine sitting down at your chamber in the morning.

You open CiteBench and the first thing you see isn't a conversation with an AI.

It's your day.

There is a hearing today. Another matter has a deadline coming up. There is an order that came in yesterday but hasn't been confirmed yet. Somewhere lower down is a matter that hasn't had any activity for a while.

Nothing here is particularly magical. That's the point.

A lawyer's day begins with work that already exists.

So CiteBench starts there — with the diary, the matters, the orders, the tasks and the deadlines — and puts the AI inside that workflow rather than making the lawyer leave it every time they need help.

You open the matter for the hearing.

And this is where the idea of a “research tool” starts to fall apart.

The matter is not just a folder containing a few PDFs.

It is the working history of the case.

The parties are there. The court and case number are there. The CNR is there. You can see the stage the proceeding is currently at, previous hearings and what happened at them, documents and papers, orders, tasks, deadlines and the timeline of the matter.

If you've already researched something for that case, the research is there too. If you've generated a draft from that research, it can become part of the papers. If there is a task that needs attention, it belongs to the same file.

The more you use it, the less the case exists as a collection of disconnected pieces.

And because proceedings don't all work the same way, CiteBench also understands the broad shape of different kinds of matters — civil, commercial, criminal, writ, appellate, family, consumer, arbitration, execution and others. It can tell you what stage the file is at, what that stage generally involves and what kinds of work normally surround it.

But it doesn't move the case for you.

That remains the lawyer's decision.

That principle becomes much more important when the system starts reading orders.

Suppose yesterday's hearing produced an order.

You paste it into the Inbox.

Maybe you upload the PDF instead.

CiteBench reads it and gives you a structured interpretation. It identifies what it believes the court actually directed, extracts a next hearing date when there is one, and may suggest things that need to be done as a consequence.

But there is a line between those two things.

The court may have directed something.

CiteBench may suggest something.

Those are not interchangeable.

So the order doesn't immediately rewrite the chamber. It waits for you.

You read the extraction against the actual order. You confirm it. Only then can the confirmed directions create court-origin tasks or deadlines and update the relevant parts of the matter.

The suggestions remain suggestions.

This is one of those details that seems small until you think about what the software is actually being trusted with. If an AI system says, “You should prepare a reply,” that is useful. If it silently turns that into “The court directed you to file a reply,” it has crossed an entirely different line.

CiteBench is designed to keep that line visible.

Now imagine that the order raises a legal question.

Perhaps the next hearing is approaching and you need to know whether a particular argument is sustainable. Perhaps the opposing side has taken a position you haven't dealt with before. Perhaps you simply need to understand how the courts have treated a particular set of facts.

You open Research.

This is still the heart of where CiteBench began.

You give it the facts, the legal question, the forum, the practice area and the side you're working for. You can attach the relevant papers. If you're already inside a matter, much of that context can come from the file itself.

Then the research begins.

What comes back isn't intended to be another chat response that disappears when you close the tab. It becomes a structured research memo: the issues, statutes and doctrines, relevant precedents, arguments on both sides, counters, strategy and risks, along with the underlying sources.

And this is where one of the most important design decisions in CiteBench shows up.

The system doesn't simply believe a model when it says, “This case is verified.”

Legal research has a particularly nasty failure mode: an AI can produce a completely plausible case name, citation and proposition that never existed.

So CiteBench's research process uses a defined set of legal sources and actually retrieves the authorities. The application then stamps the precedents based on what was retrieved rather than blindly accepting the model's own verification flag.

It doesn't mean a lawyer should stop checking the original judgment.

Quite the opposite.

It means the system tries to make a useful distinction between an authority the model happened to mention and an authority the research process actually retrieved.

That distinction follows the work forward.

Because when you read the research memo, you usually aren't finished.

You find another question.

Maybe you want to test the opposing argument. Maybe one fact changes the position. Maybe you want to understand what happens if a particular objection succeeds.

So you ask a follow-up.

CiteBench doesn't throw away the original research and start again. The new research stays connected to the original memo, giving the case a growing research trail rather than a collection of unrelated conversations.

Eventually, though, research has to become something more concrete.

You need to send a notice.

Prepare a reply.

Draft a petition.

Write a written statement.

Now the research can become a draft.

And again, the system deliberately doesn't start searching the internet all over again at the drafting stage. The draft uses the authorities that have already passed through the research process.

The sequence matters:

research, retrieve, verify, then draft.

The idea is not to make the AI sound authoritative.

It is to make the path from a question to usable work more controlled.

But there is another kind of work that doesn't require research at all.

Imagine that the morning dashboard is still showing a task:

Prepare written statement.

You open it.

CiteBench can prepare a draft using the matter itself — the papers, notes, parties and orders already sitting in the file.

No new web search is necessary.

The system already has the context.

And that reveals something I hadn't fully appreciated when I started building the original research desk.

There are really two kinds of assistance happening here.

Sometimes the lawyer asks:

“What does the law say about this?”

That's research.

Other times the lawyer is asking:

“Given everything that has happened in this matter, what can I do next?”

That's the chamber.

The second question is much harder to answer well if the software knows nothing about the first.

Then the hearing approaches.

Instead of opening five documents and trying to remember what happened last time, you can ask CiteBench to prepare a hearing brief from the matter.

It can bring together the last hearing, the last confirmed order, open tasks and deadlines, the current stage, the positions and issues, the relevant papers and authorities already associated with the file.

Again, it isn't pretending to be a lawyer.

It's doing something much more mundane and, hopefully, much more useful:

turning the accumulated context of the matter back into something you can work with.

And this is where the pieces begin to form a loop.

The hearing produces an order.

The order produces work.

The work may require research.

The research may produce a draft.

The draft belongs to the matter.

The matter needs to be prepared for the next hearing.

And the next morning, all of that is waiting on the same desk.

That loop is what CiteBench has become.

Not because any one of those features is revolutionary.

A diary isn't revolutionary.

A case file isn't revolutionary.

A task list isn't revolutionary.

A research assistant isn't revolutionary.

A document drafter isn't revolutionary.

The interesting part is what happens when they share the same context.

The diary knows there is a hearing.

The matter knows which case it belongs to.

The order knows what the court said.

The task knows why it exists.

The research knows which facts it was investigating.

The citations have a retrieval trail.

The draft knows which matter it belongs to.

And the lawyer doesn't have to keep rebuilding that context every time they move from one piece of work to another.

That is also why I don't think of CiteBench as an AI that should eventually “do the lawyer's job”.

That isn't the interesting future.

The interesting future is a system that becomes increasingly useful simply because it has been present throughout the life of the matter.

A system that can notice that tomorrow's hearing is approaching, understand that the last order created an unresolved task, know that the task may require a filing, know that the relevant research was already done, and bring those pieces together when the lawyer sits down to work.

Not making the decision.

Not pretending to be counsel.

Just reducing the amount of context the lawyer has to carry in their head.

The closer the software gets to understanding the whole working life of a matter, the less it needs to behave like a chatbot.

It can behave more like the desk itself.

That's ultimately what CiteBench is now.

It began as an attempt to build a better legal research experience.

Somewhere along the way, it became an attempt to build the place where the research belongs.

A chamber where the day's work, the case file, the court's orders, the legal research and the next piece of work can live together.

It is still very much a work in progress.

And that is why I'm putting it in front of people now.

There is a sample chamber you can explore without starting the trial, but a demo can only tell me so much.

What I really want is for lawyers to bring an actual problem to it.

Take an actual matter.

Try the research.

Put an order through the Inbox.

See whether the case file makes sense.

Try preparing for the next hearing.

And then tell me where the experience breaks down.

Because at this stage, I don't need more people telling me that the idea sounds good.

I need lawyers to use it.

CiteBench is live: https://citebench.ordereasy.win

Not legal advice. Dates, directions and citations should always be verified against the original court record before being relied upon.