First I wrote about it. Then I built it.

A moment in every litigator's career. Not about a specific case. About the system itself.

Clauses drafted a thousand times before. Evidence mapped manually into Excel. Statement structures rebuilt that already exist in dozens of variations. Not legal work - assembly work. And clients who pay dearly for it, because there's no alternative.

15 years of international litigation. Arbitrations. Complex lawsuits. The accumulation of thousands of hours - and a moment every litigator knows: <strong>The best lawyers in the country wasting time on work a machine could do.</strong> I saw it from the inside.

The problem runs deeper than time. The system creates built-in inequality. Budget - perfect statement of claim from a large firm. No budget - mediocre document, or don't file at all. Access to justice depends on what you can pay for technical work.

August 2025. Before a single line of code. Sat down and wrote an article. Not a business plan. Not a pitch deck. A professional article on AI and international arbitration. The premise: international arbitration is the natural laboratory for integrating AI into law.

Described a scenario for 2032: AI conducting interviews. Scanning thousands of documents. Identifying contradictions. Preparing witnesses. Generating reports. Wrote that it was far off. Years of development. Conceptual shift needed.

Finished writing. Read what I'd written. Realized I was wrong. Not the analysis - the timeline. Technology is already here. Not 2032. Now.

Let's be honest. Salaried partner at a large firm. Not a programmer. Never wrote code. Mortgage. Family. Leave and build a tech product alone - not a light decision.

Something I learned in 15 years: if you see the case clearly, you must act. What I saw was clear. No tech company has 15 years of litigation experience to know what "correct" means in a legal document. Almost no lawyer understands enough tech to build the system.

That gap isn't a bug. It's the opportunity.

Started building. Not with a team - alone, with AI as partner. Every legal structure. Every definition. Every quality check. Collaboration between legal knowledge and AI capabilities. Building complex systems through dialogue, without a programming background.

Sounds magical. Wasn't magical.

ORCA is not a document generator. It is a legal reasoning system that renders its conclusions as prose. The five structural moves behind every document: read the full methodology.

Methodology →

Rebuilt. Not code - approach. Quality checks for legal content, not just structure. Fabrication prevention at the core. Quality control by an uncompromising quality reviewer - no mercy, no discounts.

The principle: Facts in, Law out. Facts only go in. System does everything - causes, arguments, evidence, relief. Doesn't fabricate. Not a fact. Not a document. Not a date. Not an amount. Every word traces to input. Not a technical limitation. An ethical decision from courtrooms.

Every rule in the system was born from a specific moment when a bad output almost shipped. Four examples, out of twenty-eight:

Rule 23 was born after I made a small change to a configuration file - a change that looked entirely technical - that silently moved a statutory damages doctrine under the Defamation Law. In the claim that came out of the system afterward, the plaintiff demanded a sum the correct subsection didn't entitle them to. I found it only because I ran a second legal review on the output. Since then, every change to damages calculation mechanisms passes two parallel reviews - code and legal. A change that looks "only technical" can shift the boundary of what the law permits claiming.

Rule 26 was born after I caught myself fixing the same defect over and over. Once in contracts, once in defamation, once in unjust enrichment. Each time it looked like a separate bug. Until I realized it's the same pattern - and the fix that closed it in one place wasn't propagated to the other. Now, before any fix, one question is asked: is this a specific case of a broader class? If yes - we fix the class, not the case.

Rule 27 was born from the same insight, from another angle. A solution specific to defamation that has to be rebuilt for contracts, unjust enrichment, copyright - is future debt that accrues interest. Every new legal cluster costs exactly as much as the first, instead of almost nothing. Now every solution is measured first at scaling depth: if it isn't general, it isn't written.

Rule 28 was born latest. I ran an automated check that verifies every statute citation in the system against a verified registry of Israeli statute sections. I found 563 errors across 69 legal clusters - all because the system "remembered" sections from prompts instead of verifying them against the registry. Since then, no component in the system writes a statute citation without checking the verified registry first. AI memory is not a legal source of truth.

ORCA - not "another AI tool for law". Dozens of those exist. Most built by programmers who read about contracts. ORCA was built by someone who wrote briefs. Examined witnesses. Stood before judges.

The difference isn't marketing - structural. Programmer builds legal system = "document generator". Litigator builds legal system = decision system. Facts establish cause? Which is stronger? What will defendant argue? Where are gaps? What's the risk? Not questions for a text generator. Questions that require years in the field.

Let's not fake a story. ORCA isn't finished. But it works - and improves every day. Built from a belief: Lawyers focus on what only humans can - judgment, empathy, strategy, presence. The rest - the machine does. Faster. More precise. With consistency humans can't maintain across 40 pages.

You discover he cut corners in places an experienced lawyer knows not to cut. That he invents sections when he can't remember the right ones. That he doesn't distinguish between "this could be true" and "this is true".

The rules are the discipline that turns a junior associate into a system you can file with in court.