ACCESS

The only light in this room comes from the monitor and the shelf behind me, where a row of human skulls sits in the semidarkness, keeping me company. I talk to them sometimes. They never interrupt, and they've already been through the worst thing that could happen to them, which gives every conversation a certain calm. Every function I write tonight begins with the same question. The question isn't what the thing is supposed to do, but what happens when someone uses it in the dumbest way physically possible.

This question has a name, and almost everything people think they know about that name is wrong.

Everyone knows the everyday version. Something goes wrong, someone shrugs, says "Murphy's law," and that's that. The line was never meant that way. It comes from a 1949 U.S. Air Force program in which Dr. John Paul Stapp strapped himself to a rocket sled to find out how much braking force a human body could take. When a reporter asked why nobody had been seriously injured in all this, Stapp answered that they had always accounted for Murphy's law: before every run, think through everything that could go wrong and prevent every single one of those things. So Murphy's law doesn't predict a catastrophe; it explains why one never happened.

That's how I work. Before I act, I run things through the machine in my head, which I call the Bullet Test, and watch what moves when I pull on a single point. That goes for code and servers, contracts, an email to a client, and a skull on the table. Every system has a path to disaster, and my only job is to know it before anyone else finds it.

I pass this mindset on to everyone I work for. And it also determines what I don't do: I turn down jobs, almost never because the work would be hard, but because I've run the machine and, 4 moves ahead, seen a bullet I don't want to stand next to. The full story is in the post "Murphy's law was the answer to the question of why nobody died."

I read people, not in the sense of a motivational coach interpreting body language but in the forensic sense. I hear a sentence and register what lies beneath it at the same time: the point where someone turns away from part of what they know, the half second that isn't hesitation but concealment, the phrasing that's factually correct and, in practice, a lie.

For 25 years, I did this in courtrooms, across from people who had very serious reasons to obscure the truth and under conditions where a mistake could cost a life or lose a case. Today I do it in board meetings, crisis meetings, and the first 5 minutes with a client who has already decided what they want to hear. The contradiction is always there. It isn't a gift. It's more of an occupational hazard that has turned out to be useful.

Large organizations gradually go blind to their own dysfunctions. The chains of responsibility grow longer and thinner until the person authorized to fix a problem is no longer close enough to see it, and eventually the process documentation describes a company that no longer exists in that form. That costs money and piles up liability, and you see it in the numbers but never in the minutes.

I usually spot the pattern in the first conversation. It doesn't take weeks of analysis; it takes the questions an internal review can't ask because its people are standing in the same blind spot they're supposed to examine.

Since August 1, 2026, my focus has been on cybersecurity: server hardening, secure architecture, and software development, all built by someone who has spent decades on the forensic side of what happens when security fails. I approach a system the way someone who wants in would, with authorization and a written agreement, and then I close up whatever I find. Vibe coding isn't the problem here. Vibe coding without a threat model and without a single person in the process who has ever looked at the system from the outside is a problem, and when it fails, it usually goes out with a bang.

That doesn't stop the forensic work. I still write expert reports on bones and skulls, worldwide, and I give the unidentified dead their faces back: from 3D skull scans, from photographs that could never be published in that form, and by aging faces or making them younger. After 25 years of reconstructing how something went wrong, you see sooner where a live system is going wrong right now.

A criminal case file can quickly run to several thousand pages spread across hundreds of PDFs: interviews conducted months apart, police reports, expert reports, cell phone and computer analyses, photographs, and timelines that only fit together as long as nobody looks closely. A criminal defense attorney rarely has the time to read all of it closely enough and almost never has the forensic background to see where a finding doesn't hold up. The errors that later get turned against the client are sitting right there: the timestamp that contradicts a statement, the lead that was never investigated, the expert report resting on a method that can't support it, the detail in a photograph that nobody connected to anything else.

I help lawyers analyze the file. I show them what doesn't fit, translate forensic and technical material into plain language, and flag where a follow-up question, a motion to take evidence, or an independent expert report might be appropriate. Legal judgment and defense strategy remain entirely with the attorney. My part is making sure nothing in the file gets overlooked.

My name doesn't appear in any case file. To the lawyers I work with, I'm "The Systems Ghost," the invisible hand in the background. The impact comes from the findings the attorney uses in their own brief, not from me personally, and anyone who doesn't appear in the file presents no attack surface for the other side either.

Over the past few years, I've built my own AI systems for this work. CrimeBot is the most important. It's based on a large language model with around 753 billion parameters, a mixture-of-experts architecture, and a context window of around 1 million tokens. I used fine-tuning to specialize it in law, criminalistics, and analysis. It reviews thousands of PDFs, reads well over 1,000 pages of a case file in a single pass, and finds the cross-references between volume 3 and volume 17 that nobody can still keep straight by the time they reach volume 17.

CrimeBot doesn't work as a single model but as a team of several agents, each with its own profession and specially fine-tuned for it. One reviews and organizes the file. One does nothing but look for errors. One is an investigator with 30 years of experience and asks what he would have done differently at that crime scene. One thinks like a prosecutor and builds the prosecution's case as strongly as possible so the defense knows what it's up against. And at the end, one reads the entire file through the judge's eyes and says what the judge will make of it. The agents don't play their roles; they inhabit them. They think and feel their way into those positions, have that profession's expertise, and were trained to weigh options and make decisions with a certain degree of emotional intelligence. For every case, I configure the agents specifically for the field involved, which takes several hours. Every source is checked multiple times and independently, and every quotation from the case file and case law is checked against the original before it goes into a document.

I actually built CrimeBot for cold cases; that's what it was made for. It knows the laws of whichever country is involved and lays every page alongside every other until the contradiction nobody saw years ago stands out. Then it suggests what was overlooked back then. It was intended for the justice system and law enforcement. But I no longer work for the German justice system or German law enforcement, and there are reasons for that. So CrimeBot runs locked down, on my own hardware, and I am the only one who can and will use it. No case file goes to the cloud or an outside AI provider, nothing leaves my systems, and when the engagement is over, all data is immediately wiped to military standards. We agree on confidentiality in writing before I open the first page.

I don't sell packages. There's no price list, no tiers, and no menu of services, because every problem is built differently and a package price reliably solves the wrong problem at the wrong level. So my first question is what the actual problem is, plainly stated before it gets dressed up to sound better. If I can't help, I say so in my first reply, without softening it. If I can help, things usually move along faster than you'd expect.

I expressly accept anonymous inquiries and ask no questions about identity or background. Anyone who needs discretion gets it by default. The problem doesn't have to be neatly worded, by the way; it just has to be there.