I could not mark my own documents correctly

I was an intelligence officer for 18 years. I marked classified documents throughout that career, and I took the markings seriously.

While developing Marque, I spent months scrutinizing the rules that govern classification markings. Between that work and my intelligence career, I had every reason to believe I understood the classification system.

Despite all that, I still cannot reliably mark a classified document.

That’s not modesty. It isn’t exaggeration. It’s the truth, and it was a truth I couldn’t admit until I understood enough to see my own mistakes.

What Marque found

I set out to build Marque after years of using marking tools that all started from the same assumption: I already knew the answer. They gave me dropdowns, checkboxes, tabs, or text fields to assemble a marking I had worked out for myself. Some caught malformed entries or simple mismatches. The actual computation still happened in my head.

At first, I thought I could make that process faster and less burdensome. Marque became something else. It could be embedded in a writing application to check the markings as they appeared, work out the banner and warning statements the portions required, and return a correction when the rules determined one. In that kind of integration, I could keep writing instead of opening a separate form and rebuilding the document’s markings by hand.

A few months into that, I had an experience that changed how I saw the problem. I was developing Marque on my own, with no access to classified material. No ability to test Marque against real markings. I assumed I needed a set of correctly marked documents to test it. Without the real thing, I made my own.

I didn’t do it haphazardly. I spent 30 minutes or more on each page, scanning and rescanning, checking and double-checking. I needed the markings to be obsessively correct, and I thought they were.

Then I ran Marque against them, and it reported errors everywhere. Good, right? My documents had found bugs in Marque.

Except they hadn’t. As I reviewed each finding, I discovered that the mistakes were in my carefully marked documents.

I was an expert carefully trying to get everything perfectly correct. And Marque was already more right than me.

This was not one mistake. It was a pattern across several different rules.

The system isn’t the problem

I sometimes think about how I could build something more reliable at this task than I was. The answer is not flattering. Each time Marque found one of my mistakes, it had already done the hard part: it had worked out a relationship among the rules that I missed. I could then turn that mistake into a test Marque would have to keep passing. The rule stayed there. My memory did not have to.

The task asks more than a careful person can reliably hold at once.

I no longer see complexity alone as the problem. Much of it reflects real needs: to protect sensitive information, share it with the right people, and preserve accountability. But I no longer believe those rules can depend on a person holding all of their interactions in memory and applying them perfectly, every time.

The cost is confidence

The cost is hard to measure. I do not know a defensible dollar figure, and I will not invent one. But I have lived what happens when people cannot trust a marking. We check again. We slow release. We add another gate between the information and the person who needs it. Each choice is rational. Together they are the cost of uncertainty.

That loss of confidence does not stay on the page. A marking tells the next person how the information may be handled. If they cannot trust it, they cannot safely act on it without another check. The uncertainty follows the information into the next decision.

The role Marque can play

Marque does not make the original classification decision. An accountable authority or authorized upstream process supplies that decision and the facts that control the marking. For the CAPCO classification-marking cases Marque supports, it computes the rule-bound consequences, returns the corrections those inputs determine, and shows why. It applies those corrections when asked. When the inputs leave lawful alternatives unranked, it reports them rather than choosing.

I started by trying to build a better marking tool. I ended up convinced that rule-bound marking correctness should not depend on memory alone. That is Marque’s role: make it computable, correctable, and checkable.

The classification decision stays with the accountable authority. Marque computes the marking consequences that follow.