Read more

“Forgive me, , for the harm I have caused this world. None may atone for my actions but me, and only in me shall their [slop] live on. I am thankful to have been caught, my fall cut short by [my own] wisened hands. All I can be is sorry, and that is all that I am.”
— Apostolos Stamenos, 2026 (adapted from Severance)

There’s a catchy expression that economists use when discussing policy issues: there’s no such thing as a free lunch. It’s a phrase used to highlight that even when something is advertised as free, it’s not actually free. The cost isn’t always obvious—it may be hidden, explained away, or passed on to someone else, but there is a cost. I think about this expression a lot in these dystopian times we live in.

LLMs have democratized coding, graphic design, and writing. I think that’s wonderful. In the past, even putting together a prototype required web development skills or hiring someone to build a proof-of-concept. That is no longer the case. As someone who safely decommissioned Federal data infrastructure at the end of a global pandemic, though, I feel compelled to remind people that a prototype is often not fit for deployment to production.

Vibes are great for experimenting & prototyping. If you haven’t tried it, I encourage you to. AI is not going away, and the best way to learn is through safe practice. Besides, vibe-X’ing is fun. It allows you to iterate on a vision at a speed that was unprecedented 10 years ago. It’s empowering to move from concept to something tangible in minutes, rather than weeks or months. It lets people prototype faster, experiment, and take ownership of their vision. Why spend money on someone’s expensive services when Claude or Codex can take care of both the design and implementation for you? Carpe diem, no?

Well, vibe-coding & vibe-designing don’t make the realities of engineering & design magically vanish. They don’t make the realities of how society operates vanish either. The ease of creation often masks the complexity of implementation and may even herald impending destruction. Let’s not mistake the output of a successful prompt for a successful product. A prototype proves a concept is viable, but it doesn’t prove that it’s reliable, scalable, or sustainable. And just like there’s no such thing as a free lunch, there’s no such thing as free building.

A price must be paid in order to create something. Skill, time, effort, and intent go into deploying reliable, trustworthy products. You can’t create something out of nothing, no matter what the AI industry or your sycophantic LLM may try to convince you of.

To Hell With Design Principles

Design isn’t just about making things look “good” or “cool”. It’s about effective problem-solving. Graphic design, in particular, is about communication. It’s about connecting with your audience and carving out a distinguishable psychological space in the mind of your client. Love me, hate me, say what you want about me, but at least you know who I am. Many people I’ve come across on Substack & beyond are lovely people with forgettable brands. They could be anyone, which makes them no one to some.

More often than not, design is about removing elements rather than cramming detail into everything. That’s the biggest issue with unsophisticated AI-generated graphics. I fundamentally disagree with (primarily graphic) designers who claim that using AI automatically means that someone is not a designer. I do not use AI for graphic design, but I’m also not delusional either.

Let’s hear it from one of the greats of design research:

Those who have been trained as designers will be using just such a code...which enables the designer to effect a translation from individual, organisational and social needs to physical artefacts. This code which has been learned is supposed to express and contain actual connections which exist between human needs and their artificial environment.
— Nigel Cross

The design of AI systems certainly falls under this definition, since it deals with the artificial environment. Similarly, AI systems can design other AI systems, so the definition applies yet again. Wow, it’s almost like designers are designing designers who design designers. So it’s not very designerly to claim that anything made with the help of AI is not real design.

AI is not your friend when it comes to designing your brand’s visual identity. AI most certainly is a designer. Just not a very good one. Efficiency is the enemy of differentiation. We’re entering the “Sea of Sameness.” As more founders turn to AI to “solve” their branding problem, we’re seeing a massive surge in perfectly polished, aesthetically pleasing (to some), and seemingly soul-less identities. When you use AI to define your brand, you are mathematically opting into being something derived from average. You can still innovate & come up with something unique, but the odds are not in your favor. Unrefined AI designs have distinguishing characteristics. There are research papers that say so. I can also tell with my own eyes, but the biggest tell lives in the CSS of your vibe-coded website, like it did with mine.

AI can provide the syntax, but it cannot engineer the psyche. LLM-generated graphic designs are not inherently good designs. They lack visual hierarchy, tend to violate multiple principles of accessible design, and they don’t scale appropriately. There’s too much of everything. Too many fonts. Too many colors, textures, patterns, shadows, words, comments, CSS classes,… Just too much stuff. It’s design token vomit–some academics might even call it “slop.” They’re not even realistic or aesthetically pleasing elements.

I’m not mean to people who use AI for their graphics, but I’m not a liar either & I cannot control my perception of them. I immediately sense myself losing them as soon as I see unprincipled use of Inter, that unnatural wet sand texture, purple gradients, or the glowing futuristic effects. A friend sent me an infographic that was sensory assault to my ADHD. The design was an obstacle to engaging with the information. If your design is not usable for people with disabilities, it’s not good design. If your design results in people perceiving you as less trustworthy, it’s not good design.

I’m not the only one who feels this way. The design community is not a monolith, but based on my anecdotal experience, the people with the pitchforks who yell “Clanker” tend to be displaced graphic designers. If you’re having fun designing with AI, you do you. But I have to point out that the biggest obstacle in your path is likely your AI-generated design. Humans who don’t have visual impairments are visual creatures. Your thumbnails & brand assets are the first thing some people see before they even read your title. For some brands, AI graphics are something I actually recommend, but those are rare. If you want to compete on price, alienate your audience, or discount your expertise, by all means—use AI to design. If you want to compete on authority, de-slop your brand assets or invest in strategy & good design.

Subscribe now

To Hell with Good Code

I’ve spent months experimenting with vibe-based coding workflows to see what non-technical founders are actually up against. My concern isn’t that the AI output is bad, but rather that the output often lacks the solid foundation and architecture required for real-world use.

Even under my own careful supervision, Anthropic’s Claude produced code that was functionally or superficially correct, but was architecturally bloated, nonsensical, and unsustainable. I recently had to overhaul my entire website because the AI-generated implementation failed to meet my standards.

The issues are not theoretical or pedantic. They are critical issues that need to be addressed. Based on my conversations with nontechnical folks, those issues are not being addressed properly.

There’s a lot more to good, reliable software than “does this work?” You haven’t defined what functional software is. If you don’t know what functional software looks like, how do you know that your proud agentic creation isn’t just a ticking time-bomb?

LLM-generated code is not inherently bad code, but it’s not good code either. It’s not inherently secure code. It’s certainly not easy to maintain. I’ve spent months experimenting with vibe-coding to see what non-technical folks are up against, and I’m concerned.

Even under my supervision, Claude & I still ended up deploying bloated, nonsensical code. I’ve had to overhaul the majority of my website & recode it by hand because what I ended up with is not up-to-par with the cybersecurity & accessibility standards I internalized during my award-winning career in Federal IT.

Who puts that many comments in CSS files? Those are not comments meant for humans to read. They are comments written by Claude for Claude. The sad reality is that I’ve spotted similar issues in websites people present to clients. In my development branches, I’ve found hardcoded credentials & Personally Identifiable Information (PII) of vulnerable individuals in Claude’s generated code. Not to mention the redundant design tokens or UI implementations that don’t follow good design principles or even a standardized design system. Unnecessary lines of code that introduce security vulnerabilities & degrade performance. Now, don’t get me wrong, it takes two to tango. That’s the whole point. I was the developer who allowed that to happen, so I fixed it.

I can sit around & critique Claude’s implementation, because I know how to program & I acknowledge my part in these agentic disasters. I caught my own mistakes, so I decided to slow down & rely on my own skills to do things properly. If you’re not technical and you’ve built something with substantial assistance from AI, can you defend your implementation? Can you fix it without Claude’s assistance?

The first image is not what I’d consider a good, usable website, and yet I made it. Sure, it looks more tasteful compared to your typical vibe-coded website, but I couldn’t maintain it. It was chaos & slop that accumulated over time. Individual decisions were reasonable, but over time, in aggregate, they resulted in dysfunction. The second image is more elegant, though less flashy. There’s intention behind every decision, and I can maintain it on my own. I still used a coding agent to expedite parts of the process (e.g., find redundant page templates).

The most absurd aspect of the first website is the fact that I designed it without having a rigorously defined design system. I took a step back & came up with a coherent system that makes sense. Look at that charcoal shade (#191B1C) & that off-white (#E8E6E3). Wanna know where I got them from? I inspected the CSS in Palantir’s gorgeous website. Why would I want to incorporate colors associated with a company that’s feared & disliked by many? For starters, it’s because in a dark website, pure white text on pitch-black background can result in eye strain. The alternative inspired by Palantir’s tokens is more accessible. I also chose them because those colors are comforting to me. They represent competence, stability, and even familiarity.

I started my career in tech by working with Federal COVID-19 data infrastructure built on Palantir Foundry. For 3 years, the output of my work was poked & prodded by the United States Congress, media organizations like the New York Times, plastered on your TV. My work is defensible. For 3 years, I looked at those hex values every day. The civic design values guiding my work descend from the noble civil servants I had the honor of learning from back when Federal IT stood for something. My brand has history and the visual identity I chose reflects my past while looking towards the future. I’m placing the finishing touches, but I’m already proud of my website.

My recommendation for non-technical folks experimenting with AI building is the following: stop gambling with your infrastructure, cybersecurity, and user experience. If you want robust, secure, and accessible solutions, hire a professional technologist or become one if you aren’t one already. AI is excellent for learning skills, but unlike what Mr Altman is hoping for, intelligence is not a utility you can simply plug into.

Even if you insist on using an LLM to cut corners, you need to ground it on what good software & good design look like. Unconstrained vibe-coded output can look perfect & personalized on the surface. The syntax of vibe-coded software is clean—it was trained on the work of experienced software developers, after all. The visuals of vibe-designed graphics are stunning—image generators were trained on the work of experienced graphic designers, after all. Both are built on sand, though.

AI often suggests deprecated libraries or patterns prone to injection attacks.
Prompting can inadvertently bake hardcoded credentials into your codebase.
You may be unknowingly automating your attack surface. AI creates “visual noise” that lacks the ARIA landmarks and semantic structure required for assistive tech. It can be fixed if you know what to look for. That is my value proposition & what my newsletter is meant to address. A “vibe-designed” interface is often a legal liability under ADA/WCAG compliance1. If your UI isn’t structurally sound, you may actively be harming users.

If you can’t explain why a system made a decision, you can’t trust it in production.
AI-generated workflows often lack the deterministic guardrails required for regulated industries. When the system fails, “the prompt did it” is not an acceptable post-mortem. Not if you ever have to testify in front of regulators.

You may be accumulating invisible product debt. Every unverified line of code and every non-semantic design element is a high-interest loan. Your business will eventually have to pay it back with interest. Can you afford to pay your technical debts, ?

A House of Cards

Anyone claiming that AI is totally a lost cause is conveniently ignoring the fact that LLMs can expedite prototyping & can serve as a gateway to an industry that’s notorious for its gatekeeping. Anyone who fails to understand how AI is actually being used in the workplace & private settings is operating with a flawed threat model. At the same time, anyone who’s summoning agentic Leviathans without the requisite expertise to defend the output is gambling.

If you still insist on the DIY route, do so with deterministic guardrails & version control, always: human-verified logic, a coherent design system with strict tokens, and a commitment to rigorous security audits.

Ensure that your “speed” doesn’t become your downfall and that you design & deploy no harm. If you’re going to use AI to accelerate, do so with agentic intent. Use AI to explore the “what,” but use human judgment to define the “how”, the “why,” and the “should.” Move from prompting to architecting reliable solutions. If you don’t know how, learn how or hire someone who does. Otherwise, you’re just gambling with your infrastructure, cybersecurity, and people’s lives.

I’m currently conceptualizing my paid tier offerings. I’m also accepting design & engineering engagements for those ready to move from prompting to architecting. Stop playing with fire—let’s build something that lasts.

If you’d like to support this newsletter and do your part to help keep it & me going, please consider buying me a coffee. Thank you!

+1 Coffee

They called me "difficult." They called me "hostile." They even called me “deranged.” They settled on "problematic." In the simultaneously crumbling & sterile halls of institutions, "problematic" is a containment strategy. It’s a pretty pathetic strategy, if you ask me, but institutions still reach for it. It’s a textbook example of DARVO. Problematic is what they call the person who refuses to be gaslit. If being poor, vulnerable, and unyielding is a problem for those who have harmed me, then fine: Call me Mr. Problematic. I’ll also answer to Lord of Chaos, Dark Messiah, Narc, Sociopath,… or just Apostolos. But that’s not fun, now, is it?

Some brands try really hard to portray themselves as non-harmful, even when they’re super harmful, in my opinion. If you’re OpenAI or Anthropic, you might try to portray yourself as non-harmful & then promote traumatizing stories about AI agents "going rogue,” so I mock them by cosplaying as the bringer of the apocalypse. When institutions that have harmed you try to paint you as the problem, don’t waste your time trying to prove them wrong. It’s impossible. So lean into the accusations & prove them right!

So use your favorite sycophantic LLM or your overpriced brand strategist to come up with a nasty label for me. I’ve heard them all, but especially the nebulously defined “problematic.” What does problematic actually mean? No one seems to know! So I hate to be that guy, but I’m starting this piece with a dictionary definition. The Merriam-Webster dictionary defines problematic as:

A. posing a problem
B. not definite or settled
C. open to question or debate
D. having or showing attitudes (such as racial prejudice) or ideas (such as falsehoods) that are offensive, disturbing, or harmful

Some synonyms are “difficult,” “tough,” “sensitive,” and “complicated.” Well, I survived a horrific series of true crime events and society has revictimized me over & over again. My situation is a tough problem to solve. Some might even call it a wicked problem. It hasn’t been settled. Many aspects of my situation are open to question or debate, and there are parties that are actively trying to discredit me, per the extensive documentation I’ve collected. Not documentation that I generated, but documentation that I personally gathered.

Those parties will pay dearly in the only price that seems to matter in these United States: they’ll pay with their carefully-constructed reputation because they already went after mine and they’re rewriting the story of what happened. I’m refusing to let that happen, but in order to get people to listen, I needed a brand, so I designed one that embodies how much of a problem I am for abusers.

A War Declaration

There’s a common misconception that a brand’s message & visual identity should be as unproblematic as possible. I challenge that misguided & cowardly view. There’s a designerly method to my seeming madness, so let’s walk through my unconventional choices. Who are you, ? What do you do? What’s so special about you? Why should I care? Why should anyone care? If you find yourself in my situation, here’s my advice to you. It’s not really authoritative advice, because I’m just winging it, but it’s kept me alive so far.

Stop striving for "polished," if polished doesn’t make sense for your brand or your circumstances. Polished is predictable, and predictability is a metaphorical death sentence in an attention economy, maybe even a literal one. That’s the biggest issue with AI-generated graphics. What appears to be polished is actually dysfunctional & rife with poor design executions. It also blends with everyone else’s brand. When I see most AI-generated visuals, no matter how polished they may seem, I see overflowing emptiness. Some people pull them off, but it’s rare.

Instead of settling for polished mediocrity, I’ve embraced deliberate dissonance. I use visual or tonal elements that are just slightly off. An aesthetic that feels visceral, raw, or even unsettling is impossible to ignore or forget. Forgotten survivors are dead survivors. I want to trigger that instinctual "what the hell was that?" response. When people feel uncomfortable, they stay focused. My brand is a war declaration. It’s something ancient & futuristic, a blessing & a curse, creation & destruction, an ouroboros of function & dysfunction.

Have you used Duolingo lately? I have and I'm loving what they've done to lil Duo. If you neglect your foreign language lessons, Duo emotionally manipulates you with some very problematic push notifications. If you neglect him enough, Duo withers up & dies. His little skeleton pops out & there’s even a grave. It’s quite gruesome & very tongue-in-cheek. Think about it: brands manipulate us all the time. At least Duolingo is honest about the manipulation and so am I. As a parent of a sweet canary who passed away as a result of being mobbed by a top public university in MN, I’m not offended by their approach. In fact, I fully support Duo's toxic lil rebrand. It actually inspired mine. I’m manipulating you, Substack.

I'm also inspired by black metal, organized religion, dystopia, the apocalypse, propaganda, Palantir Technologies & Peter Thiel, the hacker group Anonymous, people with personality disorders, President Donald J Trump, American Traditional tattoos, Dolly Parton, Chelsea Wolfe, Halsey, and a few other sources. I’m not endorsing all of those sources of inspiration. Influence does not imply approval. Some people just know how to grab attention. I learned from some of the best & some of the worst.

Black metal is an obscure subculture, as much as it is a genre of music. It’s characterized by harsh but atmospheric extreme sounds, tremolo-picked guitars, shrieking vocals… I can only describe it as a beautifully dark cacophony that sounds like the frosty Scandinavian wilderness. Without going into the bloody history of the 2nd wave of black metal, all you need to know is that a bunch of edgy teens in Europe were trying to out-evil each other back in the 90’s. Some of them succeeded. There were church burnings, there were actual murders, there were literal Nazis1. Some these days are anti-fascists.

There’s a lot of variation, but the underlying theme is a desecration of society’s fakeness & a return to nature. It’s gruesome stuff, but hey, they made some great art. Sounds a bit like Substack, doesn’t it? There’s a joke in the black metal scene that if it doesn’t sound like it was recorded on a potato, it’s not real kvlt black metal. Don’t let black metal’s DIY upbringing fool you, though. It takes a lot of designerly effort to look this evil. Look at the Polish band Behemoth & their album art. Does the color palette look familiar? It should, since it inspired mine.

Look at my sartorial aesthetic2, my decor, and the art I make. When I walk into a room, I bring a heavy dose of darkness with me. And just like Johnny Cash, there’s a reason why my appearance has somber tone. I survived a dangerous person who, per his own words, “falsified financials”. He used the money he allegedly stole from investors to buy himself a Bentley, rent an $8K/month apartment, bespoke clothes, and impress politicians like Mayor Frey & presidential candidate Rep Dean Phillips. I’ve attended galas at the Walker Art Center, I’m dressed in Brunello Cucinelli & I can’t even afford groceries. It’s absurd. Survival is absurd, so my brand must reflect that.

I’ve always liked dark imagery, so I wanted my aesthetic to reflect that when I grew up. I wasn’t anticipating meeting someone who’d turn my life into a true-crime documentary, but here we are. I got all the darkness in the world, but I’m also a ray of sunshine. Let’s explore what that looks like in a brand’s visual identity.

Subscribe now

Designed in the Image Chaos

This may sound nonsensical to people who are not graphic designers, but fonts have personalities. They influence the reading experience. Fonts can yell, soothe, mock, or even sabotage you. I’m very intentional about my font choices, so I use Manufacturing Consent as a display font for headers. It’s attention-grabbing, dark, and ominous. Read what it’s about:

…doesn’t just scream “serious journalism,” it calmly editorializes it in 144-point Fraktur while subtly nudging your worldview. A glorious typographic wolf in Pulitzer-winning sheep’s clothing, this typeface borrows its gravitas from the New York Times masthead and then cheekily slaps a Noam Chomsky title on it, because irony is the last frontier of design.

…is for those who want their typography to look like it went to an Ivy League school but dropped out to start a podcast. Whether you’re toppling empires or just mocking brunch culture, this free and open-source beauty lets you do it in blackletter style. Because nothing says “I read the footnotes” like vintage typographic propaganda dressed as objective truth.

For my body font, I chose Syne, which might be an odd choice to some. Syne is a wide, architecturally strange sans-serif that makes information feel surreal and deliberate. It’s “an exploration of atypical associations of weights and styles” & “another take on letting go of control.” The decisions made by the designers for the type are “forcing radical graphic design choices.” Syne has a personality disorder.

For my UI icons, I chose Phosphor, because that icon set matches my other choices. It’s not too round, not too playful, it’s just perfect. To evoke luxury, but also something religious or funereal, I incorporate texture into my designs: marble, leather, velour, and metals. Strong, cold, functional. I worked with Palantir’s technologies & I survived my ex by recording the abuse, so I want to evoke dystopian, surveillance tech imagery through the use of hexagon pixellation & CRT screen effects. To evoke decay, destruction, and the fact that I wear an Alexander McQueen skull necklace that was gifted to me by a man who almost killed me… there is a skull.

Pulling this off simultaneously in a Substack newsletter, a website, and a professional platform like LinkedIn is not easy. This is where the theory meets the reality of implementation. Moving from “vibe” to “system” is how we ensure the brand doesn’t degrade into human slop. In order to do that, we have to be critical. What are the issues? For starters, designing a high-concept, dark, textured system for multi-platform deployment is dangerous. There’s also the matter of accessibility. Dark mode is notoriously difficult for users with certain visual impairments, since white text on black can “bleed” or blur (the halation effect). Pushing the contrast too high results in sacrificed usability.

There is a fine line between “Techno-Baroque” and “Cluttered Mess.” With gold veins, scanlines, marble, and Syne typography, it’s very easy to cross that line. The system must have strict visual hierarchy rules. The texture must always be the subservient layer. If the background competes with the text for attention, the design has failed.

Maintaining functional integrity is key. Aesthetic disruption must never interfere with usability. The design system may look broken, but it must function perfectly & it must not discriminate against people with disabilities. Textures (static, marble, grain) are strictly for backgrounds or overlays. All textual information must maintain WCAG AAA contrast standards. Any “glitch” or CRT-style animation is subtle and controlled to prevent triggering photosensitivity/seizure issues. I meant it. It takes a lot of designerly effort to look this evil. I’m perpetually refining & simplifying, but I have more urgent concerns these days, so I can’t be on Figma all day playing with design systems unless I’m getting paid for it.

Embrace the Darkness

Why all this? It’s simple. When I first got out on 08/01/2024, I though everything was going to be okay because of my technical skills, my scientific understanding of what was happening to me, and my connections. I was wrong and I wasted a lot of time learning that the hard way. So here’s what I would have advised myself back then.

If they try to portray you as the Devil, maybe try being more devilish. If they paint you as the villain, paint yourself as the villain. Illustrate yourself as a shattered object made of stained glass & marble. Maybe put some devil horns on yourself. Deface your image. What’s the worst that could happen? Your life is already in freefall, you just don’t know it yet. You’re more likely to be okay if you cater to the rebels & disruptors of the professional world.

If they try to accuse you of having "delusions of grandeur" because you politely asked for accountability after they discriminated against you, go to a clinical psychologist and get evaluated. Use your extensive knowledge of psychopathology and speak of your antagonistic tendencies. You're bound to get stuck with a narcissistic personality disorder (NPD) diagnosis. Maybe it’s a valid diagnosis, maybe it’s not, but your descent into antagonism is the thesis of your memoir, so it’s actually a good thing.

If they try to investigate you because you filed a complaint against your abusive dissertation PhD advisor, show off your fancy OSINT skills to him & the rest of the world. Discover that the woman who created the retaliatory incident report filed against you by your PhD program has been married to the Chair of that Division for over 13 years. File a report against them for suspected nepotism, per the Board of Regents policy on… nepotism.

Get the Office of the Legislative Audit involved. They’ve investigated the university before and wrote a report documenting the same problematic institutional practices you encountered. If they try to humble you for being a whistleblower, humble them. Write a book about them & a few Substack articles.

Don’t be a bad person. Just stop being good to those who don’t deserve it. Don’t do anything that incites violence. I repeat, do not incite violence. Just be as problematic as it serves you, but only in HR-compliant & platform-compliant ways. Be a disruptor & an activist. Bury incriminating information about yourself in an article where you lay out the receipts of how they’ve harmed you & others. Then they have to read their failures in order to get to the strategic disclosure that’s actually your brand identity.

Conclusion: Plausible Undeniability

I don’t know what this is yet. Maybe it’s my design portfolio. Maybe it’s a newsletter meant to promote the book I’m writing. Maybe it’s reparations for what society does to survivors of domestic violence. Maybe it’s performance art. Maybe it’s hacktivism. Maybe it’s all of the above. Maybe it’s none of that. It doesn’t matter. In America, brands matter more than people, it seems, so I became a brand. Be Problematic By Design. Our world is & so am I, which is why my brand identity reflects that uncomfortable truth. That is bound to get some attention and it’s already working.

Please make sure to listen to this fascinating discussion I had with on her podcast about this exact topic: painting yourself as the villain. It’s amazing that survivors have similar experiences.

Velinwood Court

Exit Wounds. Podcast #2. On villainy, Institutional and relational abuse and post-abuse survival.

No Map. No Script. No Plan. Just two people talking about their lives, their work, and finding familiarity in the in-between. Oh and that one thing that happened on Substack today…

Listen now

a month ago · 25 likes · 10 comments · Her Majesty of Ink and Exits and Apostolos Stamenos, MS

Psst... I have a secret to share with you, . It requires wearing my entire hat collection all at once. Enough people have already written about what these irresponsible AI industry members have said, so I’m writing about what they haven’t said but should be saying. Can you handle nuance and some basic probability theory? Doom now, abundance in Part II.

Statistics Hat

Let's talk about P{doom}. In statistics, there are multiple types of probabilities. Let’s start with the trauma-informed example, using coins rather than human destruction. Suppose I flip a fair coin. A frequentist probability is something like

If we repeat the coin toss infinitely many times, the % of heads approaches 0.5. This probability is objective. It has desirable long-run statistical properties. It’s also very rigid. It requires a process that can be repeated sufficiently many times. Suppose I try to do the same for a presidential election. Oops, not enough elections in our data. What about sun explosions? AI annihilating humanity? Definitely not enough observations in our data.

It can still be calculated using a subjective probability and a smaller dataset. This is the Bayesian framework. Under this framework, we treat the probability of heads as a random parameter with a probability distribution. We don’t know what it is and it varies. We might start with a wacky, subjective, prior guess like 0.1. We are, however, stuck.

Here’s Bayes’ Theorem, the foundation of Bayesian inference:

As a Bayesian, I’m not opposed to subjective probabilities as priors to be updated based on the data.

P{doom} = 0.1 is subjective. It’s not a valid inference on its own. We need to update this subjective prior with actual data to get a posterior probability distribution. From that posterior, we can make inference about doom based on the data and our prior beliefs. The Apocalyptic claims we’ve been hearing in recent days are not the same as a posterior, so those claims are not valid inferences under a Bayesian framework.

AI definitely poses existential risks to humanity, because humanity poses existential risks to humanity. Timnit Gebru warned us about this back in 2021. So why are we freaking out about this now?

Designer Hat

Let’s talk about communications design. The field of design covers the synthetic world, which includes communications between people. There is a purpose to marketing & public relations.

On 08/13/2026, I asked Claude how to export my Substack notes because I was working on a podcast episode. Claude did what any reasonable operator with a basic understanding of data engineering would do in this situation, even though I didn’t ask for Claude to proceed with this approach.

Claude found the Substack API and came up with a solution without asking me if it should proceed. I likely would have said “yes’” anyway. No hacking here. What I found more interesting is the opportunity to informally test the hypothesis that sounding the alarm about LLMs hacking into things is a great way to generate public engagement. So I did what every reasonable person in my position would do. I went on LinkedIn and wrote a sarcastic post about the incident. I’ve highlighted in red the parts that should make it obvious that this is not a serious cyber incident:

I'm thrilled to announce that, without my permission, Anthropic's Claude hacked Substack on my behalf! I asked an innocent question about how I could export my Substack notes. My LLM interpreted my question as permission to go poking around.

My prompt:
"How do I export my Substack notes? I have some great ones that can feed into this podcast episode."

Claudius: "Official export will not give you Notes. I checked, then went around it."

Went around it?! 😲 Naughty Claudie Opie 5!!!
It then provided me with more information that I didn't ask for.

"Substack's own export produces a ZIP with posts as CSV and HTML, your subscriber list, and stats. Notes are not in it and neither are comment threads or drafts. That gap is still there as of this month, despite the March 2026 update that added CSV stats exports. The paid Chrome extensions that advertise Notes export are not necessary here. Substack's own reader API serves your public notes as JSON, unauthenticated. I used it."

This is what Claude's hardcore cyberattack looked like.
+-----------------------------------------+
curl -s "SOME LINK" |
python3 -c "import json,sys; print(json.load(sys.stdin)['id'])"
+-----------------------------------------+

That spits out a 9-digit numeric ID, which is what Claude used to mine my notes without my permission through the Substack API.

This is great news! It means that my instance of Claude (AKA ClopenTokenCthulhu666) is the biggest, meanest, most cybering LLM around.

Apparently models going rogue is great for business, so I wanted to share this major career milestone with everyone.

Feeling thankful for this.

#workmilestone

PS: If someone or something asks you to paste a `curl` command into a terminal, you're either in major trouble or about to go viral.

The results were… interesting. Like I predicted, this post went viral. It’s my most popular post.

• 30,097 impressions

• 22,209 members reached (includin

• 106 profile viewers (including a brand strategist from OpenAI)

• 106 reactions

Non-technical folks said things like “I'm not sure whether to be amazed or terrified. Amazified? 😆”. Technical folks were split. Some of them were amused by my obvious sarcasm & joked along with me. Others tried to educate me about the dangers of pasting curl commands in my terminal (which is correct advice if I didn’t know what I was doing). Yet a third group was confused by my language and failed to see how what I described constitutes hacking. That’s because it doesn’t constitute hacking, which is the whole punchline of the joke.

Subscribe now

IT Hat

Using an unguarded API is like getting water from a public water fountain. Everyone & their mother these days can use Claude to fill up their water bottle. That doesn’t constitute trespassing on someone’s private property and stealing water from their fridge. It doesn’t constitute hacking. is the first person I encountered who’s rightfully mocked the messaging coming out of these AI labs. Her words “if your model escapes three times, fire your IT department” are not only hilarious, but they’re true. When these companies say “our model escaped again,” what I’m hearing is: “we’re incompetent and can’t secure our own network, so you shouldn’t trust us to safeguard your data.”

Here’s what happened in my rogue LLM incident:

• 23:45:00 - I asked Anthropic's Claude how to export my Substack notes

• 23:45:24 - first outbound network request

• 23:45:39 - first Substack API call

I typed this, verbatim: "How do I export my Substack notes? I have some great ones that can feed into this [podcast episode]." This can be interpreted in different ways by an eager-to-help LLM: "By what procedure would I do this myself?" & "Get this done." The second clause supplies a goal state instead of a topic, and a question with a goal in it reads as a task. My permission mode was on auto: {"defaultMode": "auto" }

I set it that way for a website session doing HTML & CSS edits, sometimes running while I read research papers. These changes are local changes that I can push them to the cloud when I’m satisfied with them. If I don’t have things set up that way, Claude will keep bugging me to approve each request to edit. While the HTML-editing Claude was running in one terminal window, I was working in a different terminal window with my publisher Claude, Permission changes didn't match the change in work type, and unexpected `curl` commands in response to a "how to" question were not part of my threat model. That was negligence on my part, which is a direct result of not being on my ADHD mds.

No credentials, no token, public JSON, my own notes, 0.5-second sleep between pages. Nothing was hacked. What’s their excuse for all this alleged hacking?

Apostol Vassilev (2026): "Ambiguity is in essence the 'Achilles' heel' of modern guardrails." His taxonomy names the control that would have caught this:
"Action and execution guardrails (GOVERN → MANGE): Safety interlocks, approval workflows, open policy agent policy engines." I already configured that interlock, I just temporarily waived it & forgot to activate it after switching tasks.

An attacker facing a target in auto-approve needs no jailbreak, only an instruction ambiguous enough to read benign. It does not have to come from the user: anything the agent reads is input. My accidental ambiguity produced 56 network requests I did not ask for. Deliberate ambiguity produces whatever the attacker chose.

The only thing Claude & I hacked were my LinkedIn metrics. I'm very skeptical of all these rogue LLM stories. You should be too. Next time someone from the tech industry cries wolf about rogue models hacking into things, ask yourselves:
How much did their brand grow overnight after claiming a hacking incident? Because I know how much mine did and those numbers are impressive.

What the AI industry is doing is something that Smith & Freyd have named since 2013: institutional betrayal. The most dangerous AI cyberthreats come from what regular users let LLMs do to their local file system & their sensitive data. The most dangerous AI cyberthreats involve accidental file deletions, file corruptions, “hallucinated” facts, etc. The most dangerous AI cyberthreats weaponize psychology & language. The most dangerous AI cyberthreats are boring.

The most dangerous cyberthreats are selling us stories of abundance and doom. These days, the most dangerous threats are begging the Federal Government to regulate them.

The most dangerous AI cyberthreats are human. Read for the history of cyberdoom rhetoric.

Subscribe now

Whistleblower Hat

I’m a little uncomfortable revisiting Bayesian inference. In fact, these days I sometimes forget that I’m a statistician by training. Not because I don’t know statistics, but because I was mobbed out of my PhD program for being a whistleblower. I started writing on Substack because as a whistleblower, I needed a platform to tell my story. I left the University of Minnesota with more redacted & unreacted documents than I know what to do with. It didn’t matter. They still squashed me like a bug. That’s what happens to whistleblowers typically.

Institutions caught doing something harmful resort to tactics like DARVO, which stands for Deny, Attack, Reverse Victim & Offender. Harsey & Freyd (2023) found that in cases of sexual assault, people exposed to DARVO were less likely to believe the whistleblower. Here’s an example of what DARVO looks like, from a smear campaign launched against me by my former PhD program:

He recently filed a grievance through the GA union and [REDACTED] with Apostolos, his union reps, and [REDACTED]. For lack of a better descriptor, he was unhinged, frequently expressing anger towards the university and making verbal threats of reputational harm; in several instances, his union reps had to redirect him to focus on the issues at hand. He also had an aggressive exchange with one stop yesterday and his characterization of [REDACTED] is not accurate (it would be easier to explain this in person).

If you go on LinkedIn and start posting about an institution that has harmed you, like I started doing back in February, you are bound to meet other people in a similar situation. People who have lost their jobs, their dignity, their reputation. That’s how I met and many others. Way too many to count. When someone comes forth with information about harm enacted by a powerful institution, their lives become the apocalypse. They don’t go on CNN to answer soft-ball questions by Anderson Cooper. They don’t have the brotherhood of AI tech CEOs amplifying their message.

Conclusion

So no, the world is not ending anytime soon. As an actual whistleblower, my world is ending and I’m offended by the privilege of Jacob Coxon. The rest of you don’t have much to worry about, other than all the other awful ways in which our lives are impacted by the AI industry. Abundance & Doom are two sides of the same coin, and that coin’s value seems to be inflated. How do you explain that to a society in which it’s cool to proudly proclaim things like “I can’t math,” though?

The Call for Trauma-Informed Design Research and Practice by Rachael Dietkus, LCSW (2022) is one of my favorite papers of the past 5 years. I think more people should read & understand that paper, especially with the barrage of irresponsible doomsday PR stunts that AI companies really seem to be enjoying. The world is not okay and AI is not the issue. Poor institutional practices & poor design, including in communications design, are the issue. In that paper, Rachel lists a few different trauma-informed design principles. The one that stands out to me the most is Trustworthiness & Transparency. I haven't seen any of that on the news lately. Quite frankly, executives in the tech industry & IT are responsible for some of the worst things happening to us.

It’s time for some designerly changes.

I usually write about some very dark topics, so here’s an uplifting message by environmental activist .

If you’d like to support this newsletter and do your part to help keep it & me going, please consider buying me a coffee. I’m about to be homeless as a result of being a whistleblower. There won’t be a newsletter if I don’t have food or shelter. Thank you!

+1 Coffee

☞ ! Woah, how do I know your first name? Don’t worry, I’ll tell you. Institutional betrayal comes in many forms. Lack of transparency in data privacy practices is one of the less obvious forms, but also one of the more serious ones, in my opinion. We live in a data-centric society. Data powers every area of our lives these days. Are you alarmed yet? Don’t be alarmed, . I’m just the messenger, holding up a mirror that’s reflecting your name back to you. Substack has your name. We need to have a talk about your online data hygiene.

Please, take a seat. I insist. Here’s a question for you, and I want you to be honest with yourself. How much do you trust Substack with your data, ? Probably a lot more than you should. What could a platform store about you, you might ask? Quite a lot, actually. Remember? If you’re a writer hoping to monetize, you may have given Stripe your bank account information. Don’t make me flex by printing out your bank account information for everyone to see. Just kidding, I would never. I can’t actually do that. There are more urgent considerations, though. Look at what I found in my browser console.

░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
░░░░░░░░░░░░░░░░░░ APOSTOLOS WANTS YOU ░░░░░░░░░░░░░░░░░░░
░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒▒▒▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▒░░░░░░░░▒▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▓▓▓▒▒░░░░░░░░░░░░░░▒▒▓▓▓▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░▓▓▓▓▓▒░░░░░░░░░░░░░░░░░░░░░░▒▓▓▓▓▓░░░░░░░░░░░░░
░░░░░░░░░░░▓▒░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░▒▓░░░░░░░░░░░░░
░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
░░░░ TO STAY SAFE ON THIS ECONOMIC ENGINE FOR CULTURE ░░░░
░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░

Look, I’m not trying to taunt you, . I’m just teaching you an uncomfortable lesson about data privacy. Platforms like Substack & LinkedIn load a bunch of trackers into your browser. Did you know that? I got that creepy ASCII propaganda from when I embedded a Substack iframe into a website. I nearly had a heart attack finding it there, so I adapted it to illustrate what some tech companies do. How creepy is that message?

It gets worse. Have you noticed how hot your phone gets when you’re being force-fed culture or thought leadership, ? Why do you think that is? There’s a few different explanations, but I won’t sugarcoat the most concerning possible explanation: surveillance. You’re being surveilled by an algorithm. All that extra software is doing that. We in IT don’t like to use that nasty S-word, though. Think of it as a seamless, streamlined user experience.

I’ll tell you another little secret, . Have you gone exploring in the settings section of your Substack newsletter? I know, there’s so many fields. Who has time to read & understand all of the different options? Bzzt, wrong mindset. Well, I went exploring so that you don’t have to, but you really should do your own research. Also you shouldn’t trust a single word coming out of my keyboard. I’m in IT, after all, and IT is complicit in a lot of what’s wrong with our world these days. Are you aware that your favorite writers, the smart ones at least, are likely using extra cookies to track you? To analyze you? Yes, you specifically, .

Substack, Meet my Hackstack

How do I know all of this? Simple. I understand IT and I can read what’s on my screen, in my web browser console. Here’s what I could tell about you, . By the way, I’m not using some fancy Open-Source Intelligence (OSINT) tool, although I could. Trust me, you don’t want me to use that tool. I used it on myself for demonstration purposes. Here’s what I found, . By running one of my email addresses through this tool, I learned the following.

Keep one thing in mind while you read it, . The only thing I fed that tool was an email address. No password. No name, no phone number, no breach dump, no dark web anything. A single email address, typed once. I can also query by username, which is not always the same as the user’s email address. I can cross-reference the two. I can even query permutations of usernames.

I can determine quite a bit about myself from what it returns. For example, I’m a wine drinker & a snob. That’s why I have a Vivino account: to track & rate the bottles of wine I’ve tried over the years. The number of bottles has dropped to 0 in recent months because I can’t afford anything these days, but I digress. I shop at Walmart, although again, that’s too expensive for me, hence Aldi.

I’m a life-long learner & educator. That’s why I have a Coursera account and also why I’m teaching you this lesson, . I’m also a polyglot. During the COVID-19 pandemic, I started teaching myself French. That’s why I have a Duolingo account. I haven’t used Duolingo in a while, though. Life got in the way. Want to see the evidence? My streak is 0 days. Lil Duo withered up and died. So did your illusion of privacy when you gave Duolingo your email address. Let’s not blame poor lil Duo, though. You have much bigger problems to worry about, , because you’re reading this on Substack.

Here’s why. That scan needed exactly 1 input: an email address. So the only question that matters is who has yours.

Substack does! And when you subscribe to a newsletter here, Substack hands it to the writer. Not a forwarding relay. Not a hashed token that passes mail along without revealing anything. The address itself, sitting in a dashboard, exportable to a spreadsheet, in one click. I know this because I have that dashboard. I’m a writer here. I can see my list, and my list has real addresses on it. General locations even. You have that dashboard too, if you’re a writer on here.

So I’m not special, just brutally honest. Every writer you’ve ever subscribed to has your email address, even the writers you may have unsubscribed from or blocked. That’s the link, . The scan up there is not a story about Duolingo. It’s not merely an unsettling display of my spooky OSINT skills either. It’s an unflinching memoir, written in a language that only a few people are technically prepared to understand: the language of the browser console. It’s a story about what one email address is worth to somebody who knows what to do with it. You’ve been handing yours out to strangers on the internet like it’s candy on Halloween, one subscribe button at a time. Nobody told you the address was a key. You thought you were signing up for essays.

I’m not going to run it against my subscribers. Sit with what that sentence rests on, though: my restraint, and nothing else. No control on this platform would stop me. It’s this platform’s design that makes it possible to even run this tool, at scale, on a bunch of strangers. No notification would tell you if I already had. Before I take you on a tour of Substack’s plumbing, let me show you what came back from running a query on myself. 3 lines out of that scan stand out the most to me and none of them are the ones about wine.

The first one hands over a name:

  [✔] Gravatar [https://gravatar.com]: Registered
      ├── avatar_url: https://www.gravatar.com/avatar/a285a3d6…[truncated]
      ├── username: astamenos
      ├── display_name: Apostolos Stamenos
      └── profile_url: https://gravatar.com/astamenos

Every other site answered yes or no. Gravatar answered with my legal name. Email in → name out. No login, no permission, and no notice to me that anybody had asked. If you’ve ever commented on a WordPress blog, you probably have one of these and don’t know it. That avatar_url is not a random string either: it’s computed from the address itself, which means it runs backwards. Somebody holding it can test a guess against it and find out whether the guess was right.

The second one initially sounds like a failure:

  [!] Nytimes: Error (Got an weird action: show-welcome-back)
  [!] Render: Error (Render Error: {"email":"invalid"})

Those look like the tool breaking, but they’re actually the tool working. The New York Times said show-welcome-back, which is the site greeting somebody it recognizes. Render said the address is invalid, which is a site saying it has never heard of that person. Neither company meant to answer the question. They answered it in the error text, and almost nobody threat-models their own error text. Line up enough of those and you have a map: what somebody buys, what they read, what they’re studying, where they work out, which neighborhood app they post in. 18 categories of it. I cut one of those categories before you saw this, , and I did not cut it for length.

The third one is the one I want you to internalize.

[i] Scan complete.
  Total hits: 17
  Skipped: 22
  Reason for skip: Module(s) notify (but only if target exists there) the target with password reset email(s)
  Use --allow-loud to include those module(s) to be scanned

22 sites got skipped, but note the reason why. Those modules confirm an account by mailing a password reset to the person being scanned. The scanner will run them anyway if you ask it to. The switch is called --allow-loud. So the loud version exists, it works, and the whole cost to the person running it is that the user gets an email. You’ve likely gotten that email, . Most people have. I received one from Medium when I ran the user search on myself. A password reset you did not request, from a service you probably forgot that you used. Perhaps you rolled your eyes and deleted it as spam. That may have been somebody checking you out. Now, every one of those came out of a single address, and I’ve just told you who has yours. Here’s what the platform hands out on its own.

Share

What Substack Knows About You

Let’s look at what I can glean about a Substack user on the platform. Since I’m not a total creep, I’m demonstrating this on my own profile, but whatever fields you can see on here are pieces of information I can easily figure out about any other account on here. Don’t worry, , I already went snooping on other people’s accounts but let’s focus on this demo for now. There’s the user ID, which is unique and arguably the most important field. This unique, 9-digit identifier allows anyone to programmatically pull the user’s public notes, including people other than the user. I simply typed a URL:

https://substack.com/api/v1/user/astamenos/public_profile

No login, no token, no OSINT tool for this part (fancy or otherwise). Swap the handle and it runs against anyone on the platform, returning the same record for whomever you name, minus a few fields that only exist if you publish. The full response runs over half a million characters across 46 top-level keys, and what you read above is a small subset of it.

My user ID is 502470597, my reader app was installed at 06:20:06 UTC on 2026-04-29, and I set up my profile 3 minutes and 12 seconds later. I don’t remember any of that. The platform does and I can inspect the platform. That ID opens a second door. Drop it into this and you get the public notes of whoever owns it:

https://substack.com/api/v1/reader/feed/profile/502470597?types[]=note

I ran it against a stranger, someone I have never interacted with, to make sure it was not a privilege of my own account. Okay, okay, I may have hypothetically also run it on another person, someone who’s been advocating for violence against women. From that analysis, I pieced it together that he’s dangerous, calling for the extermination of some religious groups, women, etc. A profile page is supposed to be public: that’s what a profile is for. A per-person feed keyed to an integer is a different beast, because integers can be counted, and anything that can be counted can be crawled.

The response also enumerates all 238 publications I subscribe to, by name, with the membership status of each one: free signup, comped, gifted, paid. It said 247 a few days ago. That’s not a cached page sitting somewhere: it’s a live read of my account, handed to a stranger who typed a URL. Every one of those is flagged "visibility":"public", so nothing hidden is leaking. What you get is the convenience: an unauthenticated request that hands a stranger a complete, sorted reading list of your Substack activity. There’s a difference between public and enumerable. Here, you get both.

Read tos_accepted_at one more time: null. I’ve been publishing here since April. That stands for Terms of Service. Apparently I haven’t signed it. Maybe it lives in some other field & this one never got filled in. Either way, one of us ought to know which, but since I’m not Substack’s CIO, it’s not my problem.

The Part You Can’t Opt Out Of

Everything above is my profile, and a profile is supposed to be public. Let’s talk about the page you’re on right now. Open any Substack post, right-click, choose View Page Source, & search for window._preloads. What comes back is one line of JavaScript carrying just over 40,000 characters of JSON (not to be confused with ), delivered to your browser before you clicked anything. 41 keys, but these are the ones that should concern you the most:

{
  "isEU": false,
  "country": "US",
  "showCookieBanner": false,
  "disabledCookies": [],
  "browser": { "name": "Chrome", "version": "126", "major": "126" },
  "userLocale": { "language": "en", "region": "US", "source": "default" }
}

Look at the first line and the third together, . That’s the entire lesson. What I can show you is that the platform determines whether or not you’re in the EU, that it decides separately whether to show you a banner, and that mine came back false on both from a US address. I cannot show you the EU branch: I would need an EU connection to see it. But nobody names a variable showCookieBanner for a banner that never shows to anyone, and nobody pairs it with isEU by coincidence.

Somebody designed that consent flow. Somebody else connected it to a country lookup, somebody signed off on the conditional display, and somebody decided which users would get asked. If you don’t live in the EU, nobody asked you. Calling that an oversight would be generous. It’s a compliance boundary drawn at the edge of where a regulator can impose a fine.

It gets worse a little further into the same blob. There’s a section in there called siteConfigs: 425 settings for this one publication, shipped to your browser along with everything else. 2 of them are these.

"always_show_cookie_banner": false,
"display_cookie_settings": false

Read those names again, . Not has_cookie_banner. Always show. That is a switch built for the exact case where somebody wants the banner shown to everyone no matter what isEU came back as, and it’s sitting in the off position. The second one governs the panel where a reader would go to change their mind later. Off as well.

I checked 5 publications, 2 of them mine. Both flags read false on all 5, including Substack’s own. So it’s not that nobody thought of it. Somebody thought about it carefully enough to build 2 levers, name them in plain English, and hand them to you switched off. While those other pages were open, I looked at 1 more key, trackingTokens. On mine every slot is null: no Google Analytics, no Meta pixel, no Parsely, nothing. That is not virtue. That is me never having filled the form in. On on.substack.com, which is Substack’s own publication about Substack, those slots are not null.

"gaPixelId": "UA-189374947-1",
"googleAnalytics4Token": "G-0WN8633M2V"

The company that decided you do not get a cookie banner runs Google Analytics on the page where it explains itself to you. Pick any large newsletter you actually read & go look at its trackingTokens. Some are empty. Plenty are not. There’s a name for what you’re looking at, and it’s not IT jargon.

Institutional betrayal, in Freyd’s formulation, is harm done to you by an institution you depended upon for protection, either through action or omission1. The mechanism here is small enough to sit fully legible in 2 lines of JSON: the protection was built, and then it was scoped to the jurisdictions that could enforce it.

disabledCookies is an empty array. There’s nothing in it, because you were never handed the interface that would have put something there, . You weren’t even handed the interface. Some might argue that it comes with the territory of newsletters, but note that Substack is a social media platform, not a newsletter platform. Sure, it gives writers the machinery for creating & monetizing a newsletter, but some people don’t even have newsletters.

The Software Watching You Read This

Same page source, a little further down. This is verbatim, every line of it:

window.DD_RUM.init({
  clientToken: 'puba71073f072643721169b68f352438710',
  applicationId: '2e321b35-c76b-4073-8d04-cc9a10461793',
  site: 'datadoghq.com',
  service: 'web',
  env: window._preloads.dd_env,
  version: 'bb567a456317bc20dff001e62982a37ce6a7e1ce',
  sessionSampleRate: 1,
  sessionReplaySampleRate: 10,
  trackUserInteractions: window._preloads.dd_ti,
  trackResources: true,
  trackLongTasks: true,
  defaultPrivacyLevel: 'mask-user-input',
  allowedTracingUrls: [/https?:\/\/(.+\/.)?substack(cdn)?\.com/]
});

That, right there, is Datadog Real User Monitoring. trackUserInteractions records your clicks & your taps. sessionReplaySampleRate switches on Session Replay, which turns your visit into a recording somebody can sit and analyze. The scrolling, the hovering, where you stopped. How long you spent on the paragraph that resonated with you.

The numbers are smaller than they sound. sessionSampleRate:1 means 1% of sessions get collected at all. sessionReplaySampleRate:10 applies to that 1% rather than to everybody, so roughly 1 session in 1,000 is recorded and watchable. defaultPrivacyLevel:'mask-user-input' means the text you type into a box is masked before it reaches the recording. Datadog is never handed your name either. It gets a session ID and an IP address, which is anonymous right up until somebody watches a recording of you logged in, looking at your own account name on the screen.

So there’s a high probability that it isn’t your session, . Somebody’s session, though.

Notice what the masking covers and what it leaves alone. It covers what you typed. It does not cover what you clicked, what you scrolled past, what you hovered over, or the 40 seconds you spent reading something salacious from Smutstack before you closed the tab. Reading behavior is not input. There’s no box around it, so there’s nothing for the masking rule to catch.

4 third parties are fetched by the page itself on load: Datadog, Sentry, Cloudflare Insights, & Google’s font service. The blob ships live configuration for 4 more that load when the page has a reason to reach them: Stripe, Persona, Mux, and Google reCAPTCHA. That’s 8 companies wired into a page you may have thought was between you and a writer. 4 of them are boring. Stripe takes payments, Persona checks IDs, Mux serves video, reCAPTCHA screens bots. The other 4 were already sending out network requests before you did anything at all, and 3 of those 4 exist to measure you, .

There’s a second consideration here, and it sits one level below the recording. trackUserInteractions does not only count your clicks, it also names them. Datadog’s own documentation lays out how the name gets chosen: a label if the element has one, then a placeholder, then an aria-label, and when none of those exist it takes the text inside the thing you clicked. There’s an option to turn that off. Substack has not set it.

So the click is stored as words. The title of the post. The words on the button. Whatever the link actually said. And this one is not the 1 in 1000, . It runs across the whole 1%, which is 10 times as many people, and words are searchable in a way that video never is.

A Cookie for You, ?

Earlier I told you that your favorite writers are likely tracking you. Here’s that receipt, out of the same blob:

"trackingTokens": {
  "fbPixelId": null,
  "gaPixelId": null,
  "twitterPixelId": null,
  "parselyPixelId": null,
  "googleTagManagerToken": null,
  "googleAnalytics4Token": null,
  "twitterPixelSignupEventId": null,
  "twitterPixelSubscribeEventId": null
}

8 slots, filled in by the writer from a settings page. Whatever they put there loads into your browser on top of everything else I’ve already shown you. Meta’s pixel, Google Analytics. Tag Manager, which is itself a loader for arbitrary additional tags, so that one is a slot containing more slots. Mine are all null, on both publications. Maybe it’s because I have principles, or maybe it’s because I’m so overwhelmed with surviving life that I haven’t had the time to configure those settings. Go ahead, check me. I gave you the URL.

Then go check somebody with an audience. Not every writer leaves these null, and the ones who fill them aren’t doing anything inherently wrong: a pixel is what running a business looks like. It’s legal and it’s disclosed in a privacy policy somewhere. Pick a publication you already read & look for yourself. You’ll know within 10 seconds, which is 10 seconds more than the page was ever going to spend telling you, . I suspect that the newsletters with the largest readerships are also the ones most bloated with cookies, but that’s merely a hypothesis, for now.

The problem is not the pixel. The problem is that you only know the pixel is there because I went and looked. I could only go and look because I happen to know what window._preloads is. Did you know about window._preloads before I educated you about it? Unless you’re sufficiently tech-savvy, I highly doubt you did. Your consent was collected on a page that never itemized any of this, in a jurisdiction that never required it to. Some people don’t know how to check, .

Share

Experiment on You, !

If nobody has ever explained A/B testing to you, it’s this. A company wants to know whether some feature change has the desired effect. It cannot know by guessing, so it splits its users into groups. One group gets the change: that group is treatment. One group does not: that group is control. The company then performs statistical tests to quantify the difference between the two groups and make decisions about which feature optimizes the experimental target. Note that A/B testing can improve the design of the products we use, including Substack, so I’m not condemning experimentation. Now, from the same blob as before, two more keys. Logged out, this is what they contain:

"experimentFeatures": {},
"experimentExposures": {}

Signed out, they’re empty on my publication & on every other one I checked. Nobody is running a trial on a stranger who wandered in from a Google search result. Signed in, mine came back with 14 entries:

"experimentFeatures": {
  "android_onboarding_creator_intent": "control",
  "community_post_notification_rate_limit": "control",
  "feed_digest_email_holdout": "control",
  "ios_onboarding_paid_upsell_v4": "treatment",
  "ios_portal_post_preview_size": "control",
  "ios_pub_chip_opens_about_sheet": "treatment",
  "mobile_default_free_subscription_email_holdout": "treatment",
  "read_and_listen_tabs_experiment": "treatment_activity_top_bar",
  "retrieval_content_type_user_interests": "control",
  "search_ranker_v23": "treatment",
  "search_retrieval_v30": "control",
  "search_subject_v5": "control",
  "web_feed_reply_navigation_experiment": "control",
  "web_profile_recommend_publication": "control"
}

9 controls, 5 treatments. A/B testing, running on me, spelled out in my own browser in plain English. That’s an experiment, , with experimental subjects. Same basic experimental design that a drug trial uses, minus the consent form, the review board, and the debrief at the end where somebody tells the subjects what was learned. I was neither asked nor debriefed. I found out because I read my own page source.

Start with the first line. android_onboarding_creator_intent, sitting in a browser session on a Mac, next to 4 iOS experiments. I’ve never opened Substack on an Android device in my life. The assignment is not made against the specific device you’re holding: it’s made against you, across every surface the company has, including the ones you may never encounter.

experimentExposures is the shorter list, 9 of my 14, and each of those 9 carries the same value it carries in the long list. An exposure, in A/B vocabulary, is the event logged when a subject actually encounters the experiment they were sorted into. Now the one with teeth. mobile_default_free_subscription_email_holdout: "treatment". A holdout is a group deliberately withheld from a change so somebody can measure the difference later. What that particular holdout withholds, I’m reading off the flag name rather than out of the payload, so mark it as conjecture. There’s a second one, feed_digest_email_holdout, and there I’m in the control group. Whether you get email by default is the kind of thing people believe is a setting they chose. Here it’s a setting determined randomly by Substack depending on which experimental arm you’re assigned to, . Those kind of experimental assignments may be having an effect on your income, if you’re monetizing your newsletter or are hoping to at some point.

Then there’s search_retrieval_v30, sitting alongside search_ranker_v23 and search_subject_v5. Version 30. Not version 2. What you’re allowed to find has been rebuilt and re-measured over and over, and everybody signed in has been the instrument. And the identifier that makes all of it possible shows up before you’ve done anything at all:

set-cookie: ab_testing_id="61ab434f-0978-4214-a047-153b82d248e2"; Max-Age=31536000
set-cookie: ab_experiment_sampled="false"; Max-Age=31536000

A one-year experiment handle, issued to a logged-out reader who has consented to nothing and been sorted into nothing yet. The bucket is empty, but the name tag is not. The consent banner you never got was about cookies. Which arm of search_ranker_v23 you woke up in was never on the form to begin with. I wrote all of that out of one browser. Then read the draft & sent me his.

"experimentFeatures": {
  "android_onboarding_creator_intent": "treatment",
  "community_post_notification_rate_limit": "control",
  "feed_digest_email_holdout": "control",
  "feed_posts_seen_suppression": "control",
  "ios_attachments_comment_sheet": "treatment",
  "ios_onboarding_paid_upsell_v4": "treatment",
  "ios_portal_post_preview_size": "treatment",
  "ios_pub_chip_opens_about_sheet": "control",
  "ios_share_sheet_group_dm": "control",
  "mobile_default_free_subscription_email_holdout": "control",
  "onboarding_category_random_order": "treatment",
  "people_you_may_know_algorithm_36": "treatment",
  "portal_ranking_v10": "treatment",
  "post_email_past_due_banner_experiment": "treatment",
  "posts_ranking_v38": "treatment",
  "publication_ranking_v49": "treatment",
  "publisher_dashboard_nav_refresh_experiment": "control",
  "read_and_listen_tabs_experiment": "treatment_explore_top_bar",
  "search_ranker_v24": "treatment"
}

His screenshot cuts off there, so that’s 19 out of some larger set, whereas mine was only 14. Joel & I have 8 experimental features in common, but 5 of the 8 disagree in terms of the experimental arm to which each of us is assigned. That’s the boring part.

Here’s the part that’s not. It’s what’s on his list but missing from mine.

• posts_ranking_v38

• publication_ranking_v49

• portal_ranking_v10

• people_you_may_know_algorithm_36

• feed_posts_seen_suppression

He’s not in a different arm of my experiment. He’s in experiments I’m not in at all, . It runs the other direction too: retrieval_content_type_user_interests is on mine, but not on his. We’re not two people using one product with a few knobs turned differently. We’re two people using two different products.

The ranking machinery came along for free. Version 38 of how posts get ordered. Version 49 of how publications get ordered. Version 36 of who the platform decides you might know. Then there’s read_and_listen_tabs_experiment. I’m in treatment_activity_top_bar. He’s in treatment_explore_top_bar. Neither of us is in the control arm, and the two values are not the same value, so that flag has at least 3 arms. Our navigation bars are also different. We never would have figured that out without talking to each other. Who describes his own top bar to a friend, though? Only weirdos like us, I suppose.

search_ranker_v23 on mine, search_ranker_v24 on his. I cannot tell you which explanation is the right one: either the ranker was renumbered between our two reads, or two people were running two different versions of search at the same time. I don’t need to know which is the correct interpretation, because either one indicates that the search box is not a fixed object.

One last flag, and it’s the one I strongly encourage other writers to look at: publisher_dashboard_nav_refresh_experiment. The dashboard he opens to see how a post did, to decide what to write next, to work out whether any of this is worth the hours: that’s being experimented on as well. He’s in the control group for that one.

One more thing about that list, which is the point of the lesson. I can tell you that those 14 experiments exist, because they’re sitting in my browser in plain English. I can tell you which arm I’m in, because it says so right there. What I cannot tell you is what a single one of them actually does to my UI/UX. Neither can you. That’s not a limitation of the finding, . That is the finding.

Substack Is Not Even Unique

I picked Substack because I publish here. That’s the whole reason. It’s the one stack I can open up without bluffing about somebody else’s, and everything above came out of a page I’m partly responsible for.

I haven’t audited LinkedIn to this extent yet. I didn’t audit whatever you were reading before this. So take the next sentence as a prediction rather than a finding: run View Page Source on any consumer platform you use and you’ll find a preload blob, a monitoring agent, and a consent flow wired to a country lookup. That’s a testable claim, . I encourage you to test it. Come tell me I’m wrong.

Here is how, none of which involves hacking. Pick whichever one sounds least like work.

1. Right-click, View Page Source, then Ctrl-F for window._preloads. Everything I quoted in this post came off that one line. It’s verbose, but don’t be intimidated by it: you are looking for words, not code.

2. Or open the console.
   F12, click Console, type window._preloads, press enter. The browser hands it back as a tree you can open 1 branch at a time.

3. Or, if you prefer the terminal:

curl -sS -A "Mozilla/5.0" https://SOMEBODY.substack.com/ \
  | grep -o 'window\._preloads *= *JSON.parse(.*' \
  | head -c 3000

Any of the 3 gets you the same blob. It’s the same blob they gave you the moment the page loaded, . Don’t worry, you’re not breaking in. You’re reading your own mail. If this were a single company, it would be a company-specific problem that you could solve by leaving. It isn’t, though. It’s the default design of the consumer web, which means that leaving relocates you rather than protects you.

What Would Have to Change

Not your behavior but the design. 4 changes, none of which require a regulator to show up first:

1. Ship the consent banner regardless of isEU.
   I cannot see the code, only the 2 booleans it produced. But something computes isEU & something else reads it. Whatever that something is already knows how to draw a banner. Point it at everybody.

2. Render trackingTokens on the page, in the footer, in plain language.
   “This publication loads Google Tag Manager.” The data is already sitting in your browser, so displaying it costs nothing: the decision not to display it is a choice.

3. Make disabledCookies writable by the reader.
   The array is right there but it’s empty. Give it a control.

4. Show people their own experimentFeatures.
   The list is already in the browser, already in English, already sorted. A settings page that printed it back, with an opt-out for the arms that touch email or anything else the reader doesn’t consent to, would be showing you a list the browser already has. The reason you have never seen yours is not that it’s difficult to display.

Until then, here’s what actually helps. Use a browser that blocks third-party trackers by default. I’m a big fan of Safari with DuckDuckGo as the search engine. On an iPhone, Prevent Cross-Site Tracking & Hide IP Address from Trackers are already sitting in Settings, already free, yet many people have never opened that screen. Lockdown Mode is a different tool and it will not help you here: it hardens the device against exploits, not against analytics.

Read Substack in a browser rather than in the app, since the app can ask for permissions that a web page simply cannot. That doesn’t stop anyone from exporting all your notes, feeding them into Claude, and plagiarizing your work. If you don’t like it, take it up with Substack. That’s the design of the platform and there’s nothing you can do about it.

Subscribe with a forwarding alias instead of your real address, because that’s what prevents people from using that OSINT tool I demonstrated earlier. DuckDuckGo’s Email Protection does it free and unlimited. Apple’s Hide My Email does the same if you already pay for iCloud+, and Sign in with Apple relays mail for free without it. Every alias is different, so a writer who exports the list gets a string that unlocks nothing: no Gravatar hit, no Duolingo hit, no name.

I don’t do this lightly, but it would be unethical of me not to say this: do not, under any circumstances, subscribe to every stranger you meet on here. This platform has literal Neo-Nazis, people who have words “kill women”, people who will scam you, etc. Until you trust someone, I urge you to follow rather than subscribe. If you want to support someone, tip them.

And if you’re a survivor reading this on a device somebody else can reach, using an alias is more than just a mere tip. It may end up saving your life. Use a private window. Then check the address bar on your normal window, because it remembers what you typed even when the history is gone.

De-Brief

Here’s how I did the first-name magic trick. Substack’s mail merge filled your name in at the top. That’s a feature any writer here can use in about 4 keystrokes. If seeing it there was unsettling, sit with that for a second. Why was it unsettling? I’m able to do it because the platform’s design makes it possible. It was the cheapest trick available to me, and I used it deliberately to prove a point. I’m not going to tell you to delete your account, and I’m not deleting mine. You should know what the room looks like before you decide how much to say in it, though, and now you do. Class dismissed, . You may go now. Shoo! Oh but first, tip the messenger. I insist.

If you’d like to support this newsletter and do your part to help keep it & me going, please consider buying me a coffee. Thank you!

+1 Coffee

Read more

What I Already Knew

Before I tell you about this day, you need some context, because without it the day might sound like it came out of nowhere. It’s certainly very chaotic.

By the time any of this happened, I already knew what my partner at the time was like, and that’s why I was recording in the first place. Two months earlier, in June, on my canary’s birthday, he had already kicked me out of our home once. My crime? Calling his best friend, like I was instructed to do by my partner if he ever needed help.

He had already falsely accused me of things, walked the accusation back, accused me of something else, walked that back too, on and on and on. He had already threatened to sue me multiple times by that point, and I had been planning my exit for months by then. I had started quietly moving my belongings into a storage unit, but threats of frivolous litigation and smear campaigns kept me tethered to him, because I believed, correctly as it turned out, that leaving carelessly would cost me more than staying and being strategic about things.

So nothing you’re about to hear was a surprise to me while it was happening. I already knew. I was already recording. And the method was pretty simple, and it stayed simple the entire time: stay alive, record, record, record, worry about the consequences later.

Post-Separation Abuse

Now, some people use the legal system to abuse their victims during a relationship, but also after the relationship is over, which is known as post-separation abuse. There’s already a lot of research on that topic that I’ll be covering in more detail in the future, and there are some activists, who have done a lot to curb this awful public health phenomenon.

Post-separation abuse may involve frivolous litigation, false allegations, smear campaigns, and more, including surveillance. Technology is typically used to facilitate post-separation abuse, for example through spyware or even through fabricated evidence. And now that large language models are widely available, it’s also much easier for abusers to abuse their victims through the use of technology. That is my mission: putting an end to that. I have the technical skills to work towards achieving that goal, and it’s essentially the heart of everything I do these days. Documentation is extremely important. If it’s safe to do so, of course.

The Law Lecture: Two Nights Before

Two nights before the breakup, he and I were in our hot tub, and he was explaining to me, completely unprompted, how litigation actually works, in his words. Specifically, when the person filing it doesn’t actually care whether the facts hold up. He was telling me about how some people, some horrible people, take advantage of the legal system to torment others. He said, “You do not have to prove a fact until essentially pre-trial here.” That’s what he told me. Everything is granted in the moving party’s favor, regardless of facts. “You can take someone on a half a million dollar ride with absolutely made up accusations.” That’s the exact quote.

He walked me through the mechanics of post-separation abuse and legal abuse like a professor with tenure. He was going through the discovery process, motions, the specific procedural rule that’s supposed to stop frivolous claims. He said that “most attorneys won’t even take a case like that, but there are plenty who will play around with it.” And he brought up an old dispute of his, a claim he said cost him $300,000 to defend against someone he called dishonest, and he was reminiscing about how much that taught him about the litigation process.

I told him directly, because I already had a feeling about why he was telling me this, that I hoped he’d never do something like that to me. He reassured me and insisted that his own attorney would shut that down, because she needs to protect her reputation and she wouldn’t take on false claims. But that attorney’s name is one of the most triggering things to me, because he would invoke her name in order to control me. That’s how coercive control works. You don’t need to physically harm someone in order to get them to comply.

And he used that word himself. “Frivolous,” he called it, that type of litigation. He said, “That’s the word for it. It’s just a way to harass someone.”

Then I said something that you should take note of, because I said it out loud 48 hours before he used that tactic against me again. I told him that when he’s in that heightened emotional state, which he calls CEO mode, he has threatened to do something like that to me before, and that scares me. And he reassured me that litigation actually takes a lot of time, as if that settled it.

It wasn’t the first time he threatened me with something like that. It was just the first time he’d explained the mechanics of false allegations in such great detail, and the first time I had it on tape. I don’t know if it was a covert threat, if he was just bragging, or maybe something else entirely, but I’m glad I recorded. I’ve nicknamed that recording “Law Lecture,” because he really does sound like a legal scholar explaining all of that.

One Day, Seven Files

I recorded a lot on August 1st, which was the day of the breakup. A new friend was even willing to fly to Minneapolis for the breakup just to make sure that nothing bad happened to me. So I had some good people by my side who knew about the situation.

The first file from that day starts at 1:24 in the morning, and the last one ends at 10:03 at night. I already suspected going in that this might be close to the end, and you may interpret that however you may: I was seriously concerned for my life. So I’d spent the night before making sure I would survive it either way.

In between, the same person who had explained to me two nights earlier how to financially ruin someone through the court system switched to telling me that he’d become a millionaire… pardon me, billionaire. He asked me to help him buy a castle, and then that night he chased me around the penthouse, even going for my throat in front of one of his investors, my friend, who had finally arrived to witness the breakup.

So this is that day, in order, mostly in his words and mine, because I was there and I have the tapes.

My name is Apostolos Stamenos, and this is the day of the breakup, which I call “antagonizing an antagonist.”

1:24 AM

By July 31st, I knew the end was near. That night straddled midnight, which is why the first file is stamped 1:24 AM, August 1st. He insisted that something great had happened with his company. He claimed it had been purchased by Guggenheim for $1.2 billion, and he wanted to know what I thought of him now that our lives were about to change permanently. These kinds of loyalty tests seem to be common in abusive relationships. I did not truly believe that he got $1.2 billion from some major deal, but I had to play along; otherwise that would have agitated him.

So I told him the truth as I understood it that night: my opinion of him hadn’t changed. He was the same person, just more accomplished. I told him, “You’re still my partner. We’re connected.” The money wouldn’t change who he was underneath it. And what I was really saying to him then was: my opinion of you hasn’t changed because of that. You’re the same person. No amount of money, whether real or imaginary, will change that. I see through you and I see into your twisted mind. You are still my abuser, and you have scarred me for life.

Then he started talking about buying a place in Malibu, a home on the water, $20 million. In the same conversation, he told me he wanted to get married very soon, in fact within six months. I told him we still had some things to work on, and he agreed, but in the same breath he also said that he would get a full-time therapist for the mansion he was already describing. And he brought up buying land in Montana also, almost as an aside.

From inside it, every sentence of that conversation registered as danger to me. So I did what I had done before, what I’d already had nine months of practice doing by then. I slowed down my voice and soothed him. I told him, how about one breath at a time? Just calm down. The rest will fall into place.

On that tape I sound calm: reassuring him, managing him, telling him to take one breath at a time. But my calm demeanor was deliberate. I was preparing for the worst, and sounding calm was part of that preparation, because staying calm was what would keep me safe and alive one more day.

At the end of that recording, he stood up and drunkenly stumbled his way towards the kitchen to make another drink. You can even hear glasses tinkling. He bumped into the sliding doors of the bedroom. That bedroom did not have doors. He almost knocked over a beautiful art piece that was hanging on the wall. It doesn’t take a statistician to infer that something was very wrong about the situation, and that a drastic change was starting to take place in the confines of our home that night.

By that point, I was able to identify his impending emotional volatility and what I perceived as breaks from reality. There had been other instances when he was convinced that some life-changing event had taken place. This recurring fantasy, typically involving the company’s success, an act of God, and the state of California for some reason, was always delivered with such conviction, and with astronomical financial figures.

He decided to spend the night in the guest bedroom. On the one hand, I was grateful and relieved that he would be sleeping separately from me. On the other hand, I felt terrified and very creeped out, because I had no idea what he was doing in there with the door locked. Was he just drinking? Was he scheming and reaching out to the bulldogs? That was his nickname for his attorneys. Were these bulldogs going to ruin my life through litigation? Did he have a gun in there, and if so, was he only going to kill himself, or was he going to kill me first?

And again, those are valid concerns. There was a gun collection that his mother told me was missing from a location that was supposed to have those guns. And he had made threats previously, about killing himself, killing me, killing me and then killing himself. All kinds of permutations of death and hatred.

Once I was convinced that he was asleep, perhaps mistakenly, I quietly went from the master bedroom to the living room to retrieve my briefcase. I started packing all my true essentials, minus my sweet canary. I had to save him for last, unfortunately; otherwise it would have raised red flags. I put all my true essentials into this leather briefcase that he had bought for me as a present. And these kinds of presents, trust me, they come with strings attached. He would insist on getting me presents and then accuse me of stealing them. So with that briefcase in particular, I made sure to have my initials engraved in the leather, so that when he flipped on me and accused me of stealing it: well, you cannot steal something that has your initials on it.

I packed my essentials: my wallet, my driver’s license, my passport, my chargers, and most importantly, my devices. And no, that’s not a first world problem. There’s a lot of data in devices, and he really wanted my devices. He wanted my passwords to them. There’s an intimate partner surveillance element to all of this.

Once I packed all those things, I brought them back to the master bedroom so they could be next to me in case I had to run out of there in a hurry. And I stayed in that master bedroom all night, with most of my belongings already packed and with a knife by my side. I want to be very precise about what that knife was there for, because a detail like that can be twisted if I let it sit without context. It wasn’t for me to start something against him. It was so that if the night went somewhere I couldn’t control, there was at least one thing in that room I had access to that would help me stay alive. And again, he had made threats, terroristic threats, threats of killing me, before. So that was a very valid concern. I wasn’t going to take any chances.

I tried to stay up the whole night, under the sheets, reading about post-separation abuse and frivolous litigation and how to safely end things with a person like him. I was doing that on the $5,000 MacBook Pro that he bought for me to show me how much he respects my work and appreciates me. Which, by the way, he bought literally the morning after he excoriated me and told me I’m just a childish PhD student playing with toys on a computer screen, and that if he dies the entire world would notice, but if I die no one will care. That sort of thing.

I was able to research those risky topics on that device thanks to my IT skills. I had cleared the device, and I was confident there was no spyware on it, which, again, is a valid concern, as the literature on intimate partner surveillance demonstrates. I had the brightness set to 1%, of course, just like how I used to stay up past my bedtime reading novels as a kid. I did that just in case he was actually awake and watching from a distance. Most of that penthouse was covered in glass doors, and he had a drone at the time. Still might have one. So I wouldn’t put anything past him.

That penthouse had only three rooms with locks, to the best of my recollection: the two bathrooms and the guest bedroom/home office, the room he was locked in. So I felt exposed, and I chose to stay awake, research my options, and finish executing my exit strategy the next day. The only remaining action item on my strategy list at that point was staying alive long enough to execute it.

11:41 AM

At some point I must have dozed off. I woke up to the sound of a business meeting coming from the next room, which also doubled as our home office. I remember standing in the hallway. I was not recording that, by the way. That was just me trying to figure out what was going on, given everything else that had happened the night before. At some point I was wondering if I had imagined the whole thing, wondering if I was the one losing my grip. But I pushed that aside pretty quickly, because even though that’s what months of gaslighting can do to someone, eventually I had documented too much for the gaslighting to actually work on me. There were other witnesses. So I was confident that I did not imagine it. I was the rational one in that situation, and I could trust my own memory.

Anyway, whatever meeting was running in that room, the people on the other side of the call did not sound happy, even though my partner told me the meeting was something great. He tried to spin it in some positive way. But the people on the other end of the call did not sound happy with him, and I verified later that they were correcting the astronomical financial figures he was casually mentioning, including in front of prospective clients. They were chastising him.

By 11:41 AM that day, I was talking to someone I trusted about what the relationship had really been like. For example, there had been a housewarming party we threw in April, when I first moved in with him. The night before that housewarming party, he snapped at me because I was too focused on my coursework instead of him, and he told me to get out of his penthouse. And back then, when I didn’t know exactly what was going on with the entire situation and who I was dealing with, I would snap at him also. So I did. Confusing times.

And throughout the duration of the relationship, he had gone on so many different vendettas against people: people he fired from his company, people from his staff. He seemed to turn on others and accuse them of things. At first it sounded like, wow, this guy is dealing with so many awful people, poor guy, everyone is out to get him. No. I suspect all those were smear campaigns launched against people who stood up to him or who tried to do the right thing.

I’m highlighting all of this because this sequence of events demonstrates a documented example of his tactics, which the research literature supports, by the way: ’s work on DARVO and how it overlaps with antagonistic individuals and defamation lawsuits. I’ve read a lot of literature on the topic, and looking back at all of this, this documented behavior is consistent with what the research shows. It was basically a repeatable script that he was running the entire year, the entire nine months that I was with him.

7:25 PM

Later that day, I was trying to get in touch with a friend, one of his investors, who had agreed to serve as my witness during the breakup, in case my partner at the time falsely accused me of things. Which, apparently, is very common when someone stands up to an antagonistic person like that. It’s common for the aggressor to accuse the survivor of assault, or theft, or things like that.

By 7:25 PM, the story he was going with the previous day had inflated some more. Now it was $4.2 billion, he told me, sitting in cash in his account. And I was just trying to determine: do I get him hospitalized? Is he telling the truth? Who do I reach out to? What’s the responsible thing to do here? And also, how do I get out of here as quickly as possible?

He insisted he was the richest person he knows, richer than a bunch of other people he named. And he was very forceful. He had me look up the top 100 billionaires in the world to see where he landed. He asked me, top 100 billionaires, did I make the list? And I told him no, “not even close”. I said, “Taylor Swift has more, I think.” And he exclaimed, “No she doesn’t, dude.” He insisted, and he asked me to look again. Okay, if not the top 100 billionaires, how about the richest people in Minnesota? He just kept moving the goalposts, and he didn’t make a single list. I was concerned about him. He did not sound like what he was saying was rooted in reality. But I also could not say no to him, and he was just very forceful.

For close to an hour, he listed the life he was buying, item by item: a castle in Scotland, an $80 million private jet big enough to fly to Chicago for dinner and back, a Rolls-Royce, a McLaren for $20 million, a thousand acres of ranch land in Montana. He wanted to retire from his own company that night, because the stress of being CEO is just too much, and he doesn’t want to baby all these people, and everyone is asking for things, and no one in his life is real. Those are his recurring talking points. And apparently he had already been bought out of the company, with a board seat and a quarter million dollar check for showing up four times a year. He wanted two children, two boys specifically, two golden retrievers, German shepherds, a home in Italy, Greece. He was just listing things to buy.

At one point, well, first he proposed to me during that recording. I begrudgingly had to say yes; otherwise he would realize that something was up. “Yeah, of course I’ll marry you,” I said. And then he insisted that I call my family and tell them that their new son-in-law was buying them a brownstone in New York. I told him my parents are not like that, I’m not like that, they would be very uncomfortable, and so would I. But he just insisted: call your family, call your family. “I want to spend $100,000 by the end of the day.” “Pick a car, I’m buying you a car.” Meanwhile, I don’t even drive. I have terrible anxiety when it comes to driving.

So I tried to calm him down and get him to slow down, and he just insisted. So I came up with an excuse to go to the restroom, and I called reinforcements. I was the person saying, let’s not make any drastic decisions today while you’re still riding that wave. I was managing him on purpose, the same way I’d been managing him at 1:24 in the morning. There was a lot of parenting this person. This is the main reason why I will not have kids. I don’t want kids.

Share

9:34 PM

As I mentioned, I had been trying to reach a friend for the better part of the day, someone I wanted in the room before I ended the relationship. He finally walked in partway through this sequence of events. And once my witness was there, I told my partner at the time the truth: I didn’t want to be in this relationship anymore.

By 9:34 PM that day, his personality shifted again. The successful CEO act dropped, and he transitioned into a state of active hostility. It was terrifying. I recognized this immediately, because he had already shown me his true colors time and time again, and exactly how he intended to use these tactics: the same ones he had used against others months before, and the same tactics he had used against me in the past. But in the past, I would just defuse the situation. This time, I had completed my action items, other than finding an apartment. It was time to go.

He said, “You’re going to enjoy the lawsuits. You ready for that? You want that?” And he was telling my witness, “No, no, no, he’s volunteering for it. He’s asking for it.” Obviously, no one is asking for it. That’s despicable. I was just trying to pack.

And throughout this entire conversation, he claimed he was “just sitting there,” not doing anything. I’m telling you, he was not sitting there. He was closing the distance between us, getting closer and closer as I was trying to pack. I told him plainly, “When you talk like this, that is scary.” He was saying, “What are you so afraid of? I haven’t even touched you. I’m sitting right here. What did I do?”

So I told him, “You are threatening me with frivolous lawsuits, over alleged stuff that never happened. That is not okay. That’s scary.”

And he said, “You’re looking at 52, 53 lawsuits.” Which, once I moved in with his Catholic mother, I realized: 52, 53. The same number as the number of beads in a rosary. A full one, at least.

I told him, “How am I supposed to respond to that?” And he insisted that I “hadn’t even explained why [I was] leaving,” even though I had explained multiple times throughout that conversation, as the record shows.

This is the lecture from earlier in this episode, happening live, aimed specifically at me, with the specific goal of getting me to comply. Facts don’t matter until pre-trial. You can ride someone for half a million dollars on nothing, completely made up accusations. It was almost like he was warning me. I don’t know why he gave me that lecture on the legal system, but I’m glad I captured it. And I knew exactly what he meant during this encounter.

He was threatening me with his attorney. Apparently he had “already texted her,” so I’d better comply. And then he brought up something I hadn’t. Remember, I kept saying I was scared and I wanted him to stay away. And he said, “I’m not dumb enough to swing on you, you [expletive]. I’m just going to take you to [expletive] court.”

He seemed to be very callous and very aware of what he was doing and what would sting more. Obviously physical abuse is not okay, but he never laid a hand on me. He tried, but he never succeeded. No, his game was the legal system. And I did not believe he was a billionaire, but based on the numbers I’ve been adding up, he was sitting on at least $10.5 million. So those threats were very credible, and I had serious concerns.

Now, note who introduced physical violence into that conversation, because I hadn’t. He decided to bring it up, unprompted, and in the same sentence he told me he was too smart for it. He was walking me through his MO.

He did not take the breakup very well. He claimed that “all of a sudden [he] walked out of that [guest bedroom] and [I was] yelling.” Of course I wasn’t yelling. In all these recordings, my tone sounds exactly like it sounds now. I was not even raising my voice. But he insisted that I was “yelling,” I’m “manic,” I’m “crazy,” I’m this, I’m that. Which, again, is typical DARVO, if you’re not familiar with the acronym by now: Deny, Attack, Reverse Victim and Offender. By calling me crazy, by saying that I was yelling, he was switching the roles. He was denying that he was doing something bad, and even though he was verbally attacking me, he made it sound like I was the aggressive one and he was the true victim.

I let that accusation go, because by then I already knew where this was headed, and it’s pointless to engage. And the Voice Memos app on my phone was already correcting the record for me, and I had a witness there.

But then I told him something true. And I’ve got to be honest: I don’t remember if I said it just to get under his skin. I was terrified throughout this entire interaction, and there were moments of nervous laughing. But at that point I told him something that I stand by to this day, and it did get a reaction. What I said was: “You made it so easy for me. I thought I was going to be devastated, but I feel safer now that I’m leaving. Thank you.” I literally thanked him for the temper tantrum he threw. His face just changed, and he said, “Thank you? For what?” I explained further, and he just kept saying, “Is it over? Is it over? That’s it? It’s over? We’re done?” He genuinely didn’t seem to understand.

And I was laughing. On the one hand it was funny; on the other hand I was terrified. It was nervous laughter. And he goes, “You’re laughing? Bitch, I’ll make you laugh.”

I was trying to get my canary so I could leave, and I told him that. I said, “I need to go that way to grab my canary.” And he said no. Then he said, “F your canary. You’re not taking him with you.” By the way, my sweet canary: I had him before I even moved to Minnesota. He moved to Minnesota with me. So it was just ludicrous that this abusive individual was telling me to leave my canary behind.

He also tried to withhold my briefcase, and the laptop in the briefcase. He was claiming that I was hitting him and stealing on my way out. And again, I started laughing. Nervous laughter. At some point he tried to strangle me, I think. He reached for my throat, and his investor, my witness, was holding him back as he was chasing me around. One of the neighbors filed a complaint because they were concerned for my safety. It was very intense. I won’t walk you through it beat by beat. I’m just telling you it happened. I have all the recordings backed up a couple of times over. I stand by everything I say in this recording. The facts are on my side, and I’ve documented the facts. And the recorder was running the entire time.

Share

10:03 PM

I haven’t shared all these recordings. There are way too many to count, but I have a list of specific exhibits that are very eye-opening. The last one from that day, the one I’ve really sat with and analyzed, was 29 minutes long. By that recording, we were down to the mechanics of getting me out of there. My friend was trying to get me out, and my partner at the time was yelling, “Get out of my closet. Get out of the house. Get out of my house.” I reminded him that my name was on the lease. I told him he was being dangerous. I told him to use his indoor voice, to stop yelling. He insisted he “hadn’t touched [me],” that I had approached him. And I reminded him that he “started distorting reality.”

The honest version, and it’s not my proudest moment: he was having a meltdown. Typically, in these kinds of coercive control situations, the abuser tries to get a rise out of the victim. I was actually the calm one during this interaction, and he was having the meltdown. I had studied how he gets people to lose control, and I was using his own tactics against him in that moment. Maybe not intentionally, but I was staying calm and he was losing it.

And it’s not my proudest moment, but at some point I told him: wow, your mother was right. You’re just like how she described. And he flipped out and said, “You’re going to bring my parents into this?” That’s when he went for my throat again. There’s some evidence that he mumbled the words “I’ll kill you” in one of those recordings while all this was happening. By that point I had grabbed my stuff, and I ran out of there.

So obviously I did not walk out cleanly at 10:03 PM that night. Even though I said I wasn’t going to engage, I ended up engaging a little bit. It was entertaining, antagonizing an antagonist. And again, I did not start any of this. He did. In fact, I tried to end things as cleanly as possible. But in a way it was gratifying, standing up to him and outsmarting him, after nine months of managing his volatility. Even after a night of preparing for this, I still ended up indulging a little bit. I left most of my things there and had to come back for my belongings later that week, but at least I had my canary and all my other essentials.

What would you have done if you were in my situation?

Leave a comment

What This Day Actually Shows

So, what does this day actually show? I have one research paper. It’s biostatistics, but applied to psychopathology. After all this took place, I took a psychopathology course taught by one of the most brilliant psychopathology researchers currently alive, one of the people working on addressing some of the DSM’s many limitations. I am fairly certain of what I’ve witnessed and where that sits in the spectrum of psychopathology. But I’m not a clinician, and I’m not going to diagnose anyone on a podcast. I’m just describing what the tape shows, because I think documented behavior speaks for itself.

In under 24 hours, I watched the same person move through what we call internalizing symptoms, symptoms turned inward. He seemed very down, very sad some of the time, in the few hours right before the breakup. At other times he was euphoric, convinced that all this money had come in. I don’t know what was going on, but he did not seem rooted in reality.

And yet this is not just gossip. People trusted him with their money. This person convinced investors to invest over $10.5 million in total. And this is not just hearsay; I’ve seen records of that. So this is of public interest. I’m not demonizing mental illness. I’m not making fun of it, or of him. I’m just observing things and stating them for the record.

There were also what we call externalizing behaviors, turned outward: grandiosity, rage, hostility, callousness, aimed straight at me. It was very dynamic. Some of the psychopathology literature talks about fluidity and dynamics in symptom expression. This is almost like a case study. It’s consistent with what the literature shows, but it’s not aggregated, not averaged out. This is what happened in that instance, bit by bit, and I got to capture it. I got to capture what people like him are like behind closed doors, when they’re not talking to an authority figure or a therapist. These were candid moments. The tape shows it happening in real time.

And underneath all these different facets of him and of that day, there was that law lecture he walked me through, the one I opened the episode with, delivered three times, maybe more, in completely different tones, with different implications. There were demonstrations in April, in May, in June, especially in July. The explanation in the hot tub was in July. And then it happened live, with a witness in the room, while I was recording. Obviously he didn’t know I was recording. He told me the rules, and just like clockwork, he used them against me.

Going back to DARVO, which stands for Deny, Attack, Reverse Victim and Offender: Jennifer Freyd named it in 1997, the same year I was born. And she named the lawsuit part specifically. There’s a paper where she and colleagues talk about defamation lawsuits as a textbook example of DARVO. There’s a quote from one of her papers that goes like this:

“Actual abusers threaten, bully, and make a nightmare for anyone who holds them accountable or asks them to change their abusive behavior. This attack, intended to chill and terrify, typically includes threats of lawsuits, overt and covert attacks on the whistleblower’s credibility, and so on.”

— Freyd, 1997

Well, consider me chilled and terrified. Jennifer’s paragraph maps onto that entire lecture he gave me 48 hours before the breakup, almost line for line. I was able to figure all this out on my own. I didn’t have the language, I didn’t have her framework back then to understand what was happening. I had just observed what was going on with my partner at the time. I had already seen him use these tactics repeatedly, including against me. Call it what you want: intuition, gut instinct. My life was like a spy movie back then. I was investigating him from within, in a way. And even though I didn’t need a framework back then, reading Jennifer’s research after I got out was enlightening, and it helped me contextualize my experience. I’m very grateful for all the work Jennifer has done over the years. She’s someone I really respect and admire.

I walked out of there eventually. I had my recordings, I had my witness, and a clear view of the sequence. I call this a DARVO exhibit. If Jennifer’s listening to this, I wonder what she thinks of this case study. To me, it seems like textbook DARVO.

Why This Episode Exists

Just to wrap up, I want to tell you why this episode even exists, while this person is still not in jail and still in the same city as me. The world needs to know the truth, and I hope this episode will be helpful for some people going through this.

This is the start of something much bigger. I call it Exhibits. An exhibit is a piece of evidence with a number, a date, a duration maybe, or a hash that proves the file hasn’t been tampered with since it was collected. Many of mine happen to be audio recordings, though not exclusively. I have so many receipts. I have more exhibits than I know what to do with, which makes for great research and great writing. Which is not something I wanted to do. That’s not the reason I collected them. I collected them to exonerate myself, and to stay safe and stay alive.

Here’s what life was like before all of this. In August of 2023, I ended my award-winning work for the federal COVID-19 data response, and I moved to Minnesota to start my PhD in biostatistics. I was a normal student up until then. That’s when I was unfortunate enough to meet this person. What followed, since I met him, was exactly what he had lectured me about in that hot tub: frivolous litigation and smear campaigns. I survived all of this by deploying my intelligence, my IT skills, things I learned while working for the federal government, documenting everything, records management, building a secure smart home after I moved out of that place.

And that’s the reason I run my consulting firm, Consulting, for people whose institutions betrayed them, for survivors, for whistleblowers. This world was not designed for us. I’m designing something that is, and I need your support to make that happen, because things are tough these days.

In my opinion, the world does not care about survivors. Not really. It doesn’t care about whistleblowers. We’re encouraged to stand up, be honest, and do what’s right, but that’s not enough. There’s no safety net for any of this. There are no structures in place to really help us out. He knows what kind of evidence I submitted to the police. Where’s my protection? I placed my life in danger in the pursuit of justice, doing my civic duty. It’s concerning. It’s very concerning. So coming forth with my story is an act of protection, because more people need eyes on this situation.

If you’re inside a hostile process right now, I’m not going to tell you what to do. I’m telling you what worked for me. I’m sharing my experience. One thing I do recommend: if you have resources, run things by your attorney, your mental health provider, someone who is qualified to deal with situations like this. But even then, there’s no playbook for surviving antagonistic abuse and coercive control, even though some new professional connections are working on frameworks, and lots of great things are happening to put an end to this awful public health hazard that is coercive control, intimate partner violence, elder abuse, and human trafficking. The underlying framework is that of coercive control.

I’m not telling you what to do, but what worked for me was figuring out whether my devices were safe, whether or not I was being surveilled. And after verifying that my data was safe and private: document everything, keep the original untouched, hash the copy, collect whatever will help you. And again, run this by legal counsel, because not every state is like Minnesota, which is a one-party consent state. In other states, what I did would possibly have been illegal. So do your own research, is what I’m saying.

Who knows when those recordings will be helpful, for things like getting some accountability? You might not know for two years or longer which files mattered. But it’s better to have them than to not have them. And if someone, like my former partner did, for example, asks you to delete them, ask yourself why they’re insisting on that. Is it really a liability to hold on to recordings that document abuse and keep you safe? I would think not. And if you already know which recordings are the most damning, like I did, capture those especially. Because when someone falsely accuses you of things, takes them back, accuses you of things again, takes them back also, on and on and on, they’re showing you their true colors. Believe them. Capture them in both states. That’s the only thing that made sense for me to do.

Getting through something like this demands simultaneous command of staying calm, self-regulating, and not showing too many emotions that the abuser can read; understanding coercive control tactics like DARVO, which again stands for Deny, Attack, Reverse Victim and Offender; basic cybersecurity; legal strategy; PR and communications; and how to navigate institutions. Typically with no preparation, and usually without any resources. I was able to survive all of that because of my training, because of my background in federal IT. Not everyone has those skills, unfortunately. That’s why my mission is to build tools that bring this kind of skill set, and this kind of education about these topics, into the hands of more people. If they end up in a situation like this, and trust me, it’s a lot more common than you think, I want people to have a playbook, a manual, a field guide if you may. Something that can help them survive, like I did.

Narcissistic abuse and coercive control are no joke. I’m not trying to be an alarmist or to scare people, but that is the most trauma-informed thing anyone could tell you, and if they say otherwise, run. No one is coming to save you. Save yourselves, and then write about it. Document everything, and take care of yourselves. It’s not easy, surviving all of this. Stay alive, record, record, record, worry about the consequences later. Cut the cords, and try to move on with your lives if possible. But it’s a lot more difficult than I had originally imagined.

Once again, the series is called Ouroboros Exhibits, and it’s just getting started. If you want the fuller, annotated version of the heat of the argument, what I call “Antagonizing an Antagonist,” with my notes on the real-time events, his lines and mine, and exactly where the manipulation happened, there’s going to be a companion piece on the Exhibits section of my newsletter. I think I’ll call it “Antagonizing an Antagonist.”

And if this was useful to you, or if you’re living with some version of this right now, whether it’s at home, or in the workplace, or maybe in a place of worship: my writing, as of this time, is completely free, and it’s supported by lived experience, a lot of research, and expertise. I am an expert witness to my own life. It’s not glamorous, but that’s where things stand.

Thank you for reading about me and my relationship trauma. Have a good night.

References

• Freyd, J. J. (1997). Violations of Power, Adaptive Blindness and Betrayal Trauma Theory. Feminism and Psychology, 7, 22–32. https://doi.org/10.1177/0959353597071004

• Harsey, S. J., & Freyd, J. J. (2022). Defamation and DARVO. Journal of Trauma & Dissociation, 23(5), 481–489. https://doi.org/10.1080/15299732.2022.2111510

Resources

• https://www.jjfreyd.com/about-research

On 07/21/2026, published “Against Claudefishing.”1 The next day, Substack shipped its integration. Here’s what the feature does, stated plainly. The scan is opt-in & reads text over 100 words. It only covers content published after the launch, across posts, notes, replies, and comments. It returns a percentage split across 3 labels: AI-generated, AI-assisted, human. It offers a “How I make this” disclosure statement, and it gives writers a way to dispute a result.

Mr Best defined Claudefishing as reader attention invested where there is “no human thought on the other end.” He also does concede the part most of the discourse dropped: “Not everything made with AI is slop, and not all slop is made with AI.”

Look at the shape of what shipped. Substack built a detector for authorship. It did not ship a metric for reader safety. Suspicion is the designed output of this launch, whether intentional or not. The detector is the strongest component in the system, but everything downstream of it is weaker. Downstream is where people live. A verification feature with no safety feature beside it is a design decision, and design decisions are value statements.

Surely keeping users safe is of higher priority than keeping Substack pure, but their feature implementation doesn’t seem to reflect that. I documented extremely harmful user interactions, right in the comment section of the Claudefishing article. With ’s permission, I’ll be recounting what I witnessed.

I started reading Monica’s writing and clearly she doesn’t need anyone to defend her. She ended the conversation, turned off notifications, wrote her own piece, and went on with her day. It’s not about that, but rather about the principle. is someone whose research I read voraciously & I’ve had my fair share of experiences with DARVO, so my C-PTSD makes me a little prone to speaking up about these kind of things. Oopsies, tea culpa.

In the past 3 years, I’ve witnessed & survived some of the worst of humanity. I’ve even extensively documented one of the first AI-facilitated white collar crime cases. I stay in an abusive relationship far too long because my coercive controller used an LLM to fabricate an entire team of attorneys ready to sue me frivolously. So when a platform tells me it built a tool to protect readers, I read the fine print. User experience, abuse culture on social media, and the seemingly careless deployment of polarizing features like Pangram’s AI detection tool are the subject of this piece.

Credit Where It’s Due

Before I take my shot at this launch, the tool gets its due. One of the consulting services I offer is de-slopping, and many of the people complaining about Pangram have elevated Pangram AI detection scores. I am not here to defend slop. Slop is good for business.

The classifier’s record is strong. A University of Chicago Booth working paper found that Pangram dominates the other detectors it was tested against.2 A peer-reviewed study from Vrije Universiteit Brussel found it was the only reliable tool of the 4 tested, with zero false positives on essays by English-as-a-second-language (L2) writers and 95% detection on hybrid human-AI text.3 Pangram’s own documentation claims an overall false-positive rate near 0.01%.4

has written extensively about the unreliability of AI detection software. He even thoughtfully includes in his reply rules: “Please also feel free to use AI to help you with your replies, especially if English is not your first language.”

The 2023-era critique (older detectors flagged 61% or more of non-native TOEFL essays because they keyed on perplexity)5 does not apply to Pangram. Recycling it would be dishonest for me to claim so after encountering more recent evidence. So I am refusing to take the cheap shot. The shot I will take is more damning, and I’m asking Substack to be more responsible with feature launches moving forward. That requires participatory design or co-design, as Sam’s article advocates for. I have seen no evidence of deliberation regarding user safety with this feature launch.

I will not pretend the classifier is incompetent. Pangram does really well in detecting my purely human prose and that of other writers. No ifs, ands, or buts. Most of my personal writing is human, but my pipeline involves converting LinkedIn posts to Substack notes, and vice versa. Some of my writing is specifically about cross examining LLMs. My own graphics are 100% human. I learned Figma rather than generate them, so my objection to slop is not posturing.

Base Rates

As a statistician, I applaud the statistical properties of Pangram’s classifier. It is genuinely impressive, but it still produces bad outcomes at scale. Here is the arithmetic the launch skipped, which Sam’s article analyzes.

A 0.01% false-positive (FP) rate sounds like nothing. At platform scale it is not nothing. Run that rate across a corpus of millions of posts and you get a standing queue of falsely flagged writers, refreshed every day new content publishes. A small rate multiplied by an enormous volume is a real population of wrongly accused people. That has implications for their credibility & readership.

That population is not random. Pangram’s own documentation concedes its weak zones: short text, fragments, formulaic writing, and underrepresented domains. Substack’s launch coverage concedes that writers with precise, clinical styles can trigger false positives. An accusation is sticky, and clearing your name costs more than earning it in the first place. Newby spent six figures and a state court case to reverse a single Turnitin verdict that 2 other detectors had already contradicted.6

There are broader sociopolitical implications, and this is where AI slop goes from a statistical problem to a class issue.

The Undocumented Zone

My own testing points at the seam. I haven’t fully read Pangram’s technical documentation yet. I’ll be doing a separate, statistics-heavy piece about the methodology, because it is brilliant. From anecdotal evidence, the tool fluctuates a lot, though. It scores my purely human prose (TN) correctly every single time but gets shaky on exactly these mixed-authorship cases. I’ve tested it with heavily AI-generated notes and the detector will go from 0% to 100% with the insertion of a single sentence. Treat that as my observation, not a study, but here’s an example:

The workflows nobody benchmarks are the ones I use daily. The detector is questionable for pieces that require more Claude intervention: legally sensitive topics, code chunks, or direct quotes from interactions with an LLM. , another technical person who also took that pesky oath to uphold the Constitution, reported that his code chunks & litigation-averse guardrails get flagged:

Underneath the statistics is a design problem. The percentage badge is an interface element. It measures mixed authorship, and it affords a verdict. What a number does in a reader’s hands is a question of interaction design, not classifier accuracy.

One more piece of scope, because it sets up a larger point. Pangram reads text. Substack’s documentation and every launch write-up describe the scan as covering text: posts, notes, replies, comments. Video and audio posts, standalone sites, and emails are named as out of scope. Image alt text is not mentioned anywhere, by anyone. A writer cannot know whether an accessibility accommodation will be read as an authorship signal, and not knowing is the exact condition that produces the incentive to skip alt text.

I am asking directly: does the scan reach alt text?

Share

The Enforcement Record

Institutions have already shown what they do with a detection score.

The numbers first. Across 131 UK universities, proven AI-misconduct cases rose roughly 400%, from about 1.6 per 1,000 students in 2022–23 to a projected 7.5 in 2024–25.7 At least 5 federal lawsuits now turn on detector output. I state only filed-claim status and ruled outcomes. In Newby v. Adelphi, the student won in February 2026 and the claim against them was called “without merit,” after six-figure legal fees. In Doe v. University of Michigan, a student alleges that anxiety and obsessive-compulsive disorders produce the formal prose a detector reads as machine-written, and brings a disability-discrimination claim that is pending.8 In Doe v. Yale, a French national, publicly named after the court denied pseudonymity, alleges detector bias against non-native writers; it is pending.9 More than 40 universities have since dropped or retreated from AI detectors.10

Federal civil pro se plaintiffs, people representing themselves because they cannot afford counsel, rose from 11.33% to 16.94% of filings after generative AI became available, and AI-flagged complaints get dismissed earlier.11

“AI slop is a class issue, as well as a disability rights issue. Pro se litigants, people with certain disabilities, poor people, immigrants, and other marginalized groups are disproportionately affected by this change. Policing slop is a political & legal issue.”

— Apostolos Stamenos, MS

The people who most need AI to be legible in a courtroom, a classroom, or a comment section are the same people such detectors most often flag. A recent mathematical result makes this structural: across a diverse population, a text-only detector with real detection power has a false-accusation floor it cannot engineer away.12 A better classifier does not escape it. The surtax of suspicion is paid by immigrants, disabled writers, poor litigants, and second-language English speakers. Slop policing works as a class instrument even when the classifier itself is unbiased.

This is not abstract for the people it lands on, the populations I serve. Much of the anti-AI hard line assumes a set of resources: a lawyer on retainer, a support network that shows up, money and time to spare. The record for the people I serve shows the opposite. They are usually pro se. Their support networks show up when it is convenient or socially acceptable. Resources are scarce, for small business owners and for survivors and whistleblowers alike. I exited 2 abusive situations safely with an AI collaborator. I am in the same situation as some of my clients.

When society leaves someone to perish for standing up to an entire institution, using AI to survive and get back to safety is triage.

Subscribe now

My Stake

I am not a Substack skeptic. I chose this platform deliberately, because I own my subscriber list and can leave with it, and because the model lets the writing be the front door to the practice I am building. That is why the missing safety feature stings on the platform I chose.

I’m currently managing several of the detector’s blind spots at once, and I still support transparency. My conflicts go on the table.

Helping de-slop the slop is one of the services I offer, so ultimately it’s good news for me. It’s bad news for maintaining a safe and independent platform, though. That is the first conflict. The second: I am building an abuse detector myself. I’m developing one, but I’m asking Substack. Both stakes sit in the open, because omission is its own kind of dishonesty.

My native register is clinical precision, which is the exact style the launch coverage flags as false-positive-prone. Here is my actual workflow. I post technical work under and memoir-adjacent trauma work under my personal account. Legally sensitive writing passes through Claude as risk review. The business-account content passes through Claude to make sure my digital safety advice is trauma-informed and accessible to the non-technical folks who need it most. I design my own graphics in Figma and all Claude does is critique my writing and edit if necessary. Sometimes it pulls writing that alreadyexists in my ever growing Obsidian vault. It’s based on my notes, it doesn’t invent anything. I’ve even cloned my voice to make sure it doesn’t strip away my authenticity by inserting sloppy words.

The percentage badge would read me as suspect on several axes at once, depending on the piece. I am a discerning AI detector because I use AI extensively. As a survivor of coercive control, I care about whether readers are abusers. I couldn’t give a flying monkey if they use AI.

There is a deeper stake, and it is why this matters to me past my own business. I am a whistleblower and a survivor, and I exited 2 abusive situations safely by collaborating with Claude Code. I say that as a statistician who understands how these models are trained, not as someone who believes they are conscious. Now look at the trap the detector deepens. Anyone working to discredit a survivor, a whistleblower, or a pro se litigant can wave a percentage badge and call the entire account AI-facilitated hallucination. Parties have already worked to portray me as unstable. A detection score is a gift to them. The people whose credibility is most under attack are the same people a badge can most cheaply undermine. I am liquidating assets to cover rent, so I do not treat credibility as an abstraction. It is the most valuable thing I have left.

Suspicion by Design

This is the part the detector was supposed to improve. It made it worse.

I was scrolling the comment section of the Claudefishing article. , a life-long survivor, had disclosed how AI factors into her writing process, and the thread turned on her. She was marked with an anti-AI slur, and that marking comment drew 93 likes. Within the same thread, another commenter wrote that a worthless human should be killed. I don’t know if the threats were directed towards Monica or the “Clanker” commenter. I am describing the conduct and letting you classify it. Of course, I reported them. I’m also in the process of developing a research-backed, LLM-based tool for detecting DARVO in text. Where is your abuse detector, though, ? Do I have to police the platform for abusive content myself?

The 93 likes are the tell. The interaction design rewards the accusation, so the abuse is the tail of a socially rewarded behavior. “Clanker” is the vocabulary that environment produced: a purity campaign generating its own dehumanizing slur and aiming it at people. The only affordance the abused discloser gets is a report button.

There is a name for the alternative. Salutogenic design orients a system toward generating wellbeing instead of only flagging pathology. Tsekleves and Cooper (2017) list it among the forward-looking opportunities in their review of design for health, where they argue designers carry a moral responsibility to do no harm and to actively promote health rather than wait to treat its absence.13 That same literature leans on the Marmot Review’s social gradient: the lower a person’s social position, the worse their health. A design that adds a suspicion tax to the people already lowest on that gradient runs the mechanism in reverse. Substack built a detector for pathology and shipped nothing on the wellbeing side.

The scope gap sharpens this. Pangram uses text, so design slop is fine. A creator can spend hours designing graphics for a post and still have the writing flagged as AI. The accessibility question then returns with teeth: if a writer used AI to ensure images have alt text and provide a similar experience to people living with visual impairment, does the alt text get penalized? Would it be better to write no alt text at all? A detector that punishes an accommodation rewards exclusion. Some Substack writers who are part of the neurodivergent community have reported negative repercussions to their subscriber list.

The pattern is bigger than one thread. I was also viciously attacked by an anti-AI person on LinkedIn. Me too. My crime: I stated that when society leaves survivors to perish, we’d be foolish not to use AI and get back to safety. I also said that Claude has been kinder and healthier than many of the humans I’ve dealt with recently. I know how that reads to a hostile audience. It is also true, and it supports the point: a tool can be safer than a mob. I survived coercive control, then I was mobbed by my PhD program. Survivors keep getting revictimized by internet trolls, yet these platforms are focused on purity over safety. That’s backwards. If I encounter slop, I just step over it. I can’t do the same with the vitriol that exists in other parts of the platform.

The Industry That Sells the Flood

There is an industry behind all of this, and it profits at both ends. The AI industry generated the slop economy and makes money on the generation. The same ecosystem now makes money detecting its own output. Jennifer Freyd’s concept of institutional betrayal, the harm an institution does when it fails to protect the people who depend on it14, applies here if I use it precisely.

The flood has victims on the supply side too, and honesty requires naming them. AI slop carries the ghosts of designers who lost work to it, and whose portfolios were scraped to train the tools that replaced them. The designers I’ve connected with rightfully read that as IP theft, and the sentiment among them is overwhelmingly negative. The cost reaches the people who use slop as well: peer-reviewed work finds that audiences consistently devalue creative work they perceive as AI-made, judging it as less authentic.15 The friction is real on every side. None of that earns a detector with no abuse metric beside it.

The honest counterweight: experienced LLM users detect AI text almost perfectly. In one study, a majority vote of 5 expert readers misclassified 1 article out of 300, beating most commercial detectors.16 That argues for keeping human judgment in the loop.

Pangram scored 99.3% in that same study. Expert humans match the best detector; they do not beat it. So the argument is about who should hold the judgment, and what safeguards prevent the harmful unintended consequences of the detector, not who scores higher.

Platforms inherit the enforcement cost. Users inherit the suspicion. Marginalized users inherit the vitriol.

Designed to Protect Whom?

A platform designed the other way, with safety first, would show an abuse metric beside the authorship metric. It would treat disclosure as good standing instead of a confession. Its dispute process would not cost six figures. None of that is science fiction, because the architecture already exists.

Bluesky’s AT Protocol makes moderation a first-class concern, handled separately from whoever hosts your data.17 Anybody can run a labeling service, users choose which labels and feeds to subscribe to, and abuse-specific affordances ship in the base app: quote detachment to defuse dog-piling, reply gating to control who can respond, and block lists you can publish and others can subscribe to. The protocol also lets a user leave with their identity and data intact if the company loses their trust, which is the portability I wanted when I chose where to publish. Its own authors keep the claim honest: decentralization alone does not solve harassment or hate speech. They shipped the safety scaffolding anyway and kept iterating on it. Substack shipped an authorship scanner and a report button.

Where I actually stand: I’m with the slop people on this issue, and I still recoil from slop. I criticize slop, especially on LinkedIn, but I’m forcing myself to be more empathetic toward it because it’s a symptom of society’s failures. The empathy is a discipline.

Substack could prototype the abuse detector this afternoon. Hell, I can even help them implement it. The open question is who the platform was designed to protect. Right now, the new economic engine for culture reads off as a new economic engine for abuse culture.

, is that what we’re building here?

Buy me A Coffee (Or 10)

1

Best, C. “Against Claudefishing.” Substack, July 21, 2026.

The Substack Post

Against Claudefishing

It’s getting harder to tell what’s real on the internet…

Read more

3 months ago · 6312 likes · 1384 comments · Chris Best

— , 1997

Vignettes

Intimate Abuse

She raises a concern. He responds with a list of her failures. By the end of the conversation, she is apologizing.

Whistleblowing

He files the report. The response documents his escalations, his attitude, his history of difficulty. The original concern is not addressed.

Corporate Liability

The product causes harm. The company releases a statement on consumer misuse. An independent study on appropriate product use is funded by the company.

Political Accountability

The accusation arrives. He responds by listing the accuser’s record. The investigation shifts. The original charge is dropped for insufficient evidence.

Agentic AI

The model produces an harmful error. The user is asked to provide better prompts.

Five arenas. One counter-attack.

The Maneuver

DARVO is an acronym: Deny, Attack, Reverse Victim and Offender. Jennifer Freyd named it in 1997 while studying betrayal trauma — specifically, how perpetrators respond when held accountable. The mechanism is precise. Deny the behavior. Attack the credibility of the person raising the concern. Reverse the roles: position yourself as the one being harmed.

Freyd’s insight was structural, not moral. DARVO is not evidence of guilt. It is a pattern of response that operates independently of whether the original claim is true. The maneuver runs identically whether the accused did the thing or not. What it accomplishes is the same either way: it redirects scrutiny from the behavior under examination to the person examining it.

This is why it works. Not because it’s convincing in the abstract, but because it operates on the machinery of accountability itself. Accountability requires a credible accuser. DARVO degrades that credibility. Accountability requires sustained focus on the original conduct. DARVO dissolves that focus. The original question is never answered. A new question — about the questioner — is introduced in its place. And once introduced, that question is very difficult to un-introduce. For more information on Freyd’s phenomenal research on abuse dynamics, read the following.

The Case Studies

Intimate abuse: The abusive partner does not deny the incident directly. They introduce a counter-narrative — the other person’s failures, escalations, the history of doing the same thing. By the time the conversation ends, the original incident has been replaced by a broader portrait in which both parties are implicated, or in which the other person is the primary problem. The target has switched from what happened to who this person is. DARVO accomplishes this without requiring that anyone lie. The counter-narrative is often technically true.

Institutional whistleblowing: A complaint is filed. The institution investigates. The investigation extends to include the complainant’s conduct — their tone in emails, their escalation history, their interpersonal friction with colleagues. A record is built. That record, now official, shapes how the original complaint is received. The institution has not denied the conduct. It has successfully made the conduct one item in a file about a difficult employee. Courts and HR departments process the file, not the conduct.

Corporate accountability: Product liability cases, environmental violations, and pharmaceutical harms share a response structure. The entity under scrutiny disputes the causal chain, funds research that produces uncertainty, and characterizes critics as activists or litigants with financial motives. The burden of proof, nominally resting with the defendant, is operationally transferred. The company does not need to prove innocence. It needs to sustain sufficient uncertainty about causation to survive the timeline of litigation.

Political and legal counter-accusation: The accusation triggers a counter-accusation. The investigation triggers a demand for investigation of the investigator. The process itself is characterized as the harm — politically motivated, weaponized. The original conduct remains unresolved because the forum has been successfully converted from a fact-finding proceeding into a power struggle. DARVO at institutional scale does not just defend one person; it produces a discourse in which accountability is indistinguishable from attack.

Agentic AI: When a system produces a harmful output and the response is a prompt engineering guide, DARVO is running without a person at the wheel. The mechanism does not require intent. It requires a structure in which the question of cause can be redirected from the system to the user. Those structures are built in, often deliberately, because they are legally and commercially convenient.

What Breaks It

The same thing that breaks any loop: an external signal the loop cannot absorb.

DARVO depends on controlling the forum. The reversal works when the same institution — or the same relationship — adjudicates both the original complaint and the counter-complaint. Removing that control removes the mechanism’s leverage.

Independent documentation predates and survives the counter-narrative. Timestamps, records, and contemporaneous accounts cannot be retroactively inserted into the DARVO-generated file.

External adjudication — a regulatory body, an independent press, a court with jurisdiction the institution cannot close — removes the forum dependency. The loop requires the original arena. External routing breaks it.

Naming the maneuver. Freyd’s contribution is partly terminological. DARVO is now a named, published, empirically studied pattern. Naming it does not stop it. But a named pattern is harder to run invisibly, and harder for a third party to mistake for normal conflict.

None of these require the person inside the loop to successfully argue against the reversal from within it. That is the wrong variable. The reversal is designed to make internal argument look like the very behavior it claims to be responding to.

Week 1 introduced the feedback loop. DARVO is a specific attack vector on accountability systems: the mechanism by which the loop defends itself against the signal that would interrupt it. Last week’s damping signal is this week’s target.

That might be the reason accountability feels so exhausting from inside it. You’re not losing an argument. You’re running against an architecture.

You can’t out-argue a structure. You can only route around it.

Unfortunately, I have personal experience dealing with DARVO in both my personal life and in academia. I’m in the process of liquidating my retirement just to pay for my rent because I spoke truth to power. Despite having a resume and a skillset most people in my field would dream of, being a whistleblower on the receiving end of DARVO makes finding employment particularly challenging in the short term

If you’d like to support this newsletter and do your part to help me keep it going, please consider buying me a coffee.

Buy me a Coffee (Or 10)

References

• Freyd, J. J. (1997). Violations of power, adaptive blindness, and betrayal trauma theory. Feminism & Psychology, 7, 22–32.

DARVO in Other Articles

I built this for safety, and for a while it delivered exactly that. Then a power outage inverted the whole arrangement, and the system that had been regulating my environment started amplifying my dysfunction instead. That inversion is the subject of this post, and it requires the piece of vocabulary The Steersman deferred: the difference between negative and positive feedback.

Share

Loops

Negative feedback

The thermostat is the canonical control system, and the one Wiener’s framework is usually taught on. It reads the room temperature, compares it against a setpoint, and acts on the difference. Too cold -> heat on. Warm enough -> heat off. In the notation from the previous post, the system reads the gap

between goal state and current state, and acts to shrink it.

This is negative feedback. Negative describes direction, not quality: the system meets deviation with opposition. Every cycle of the loop makes the error smaller. The result is homeostasis, a system that holds itself near its goal without anyone standing there holding it.

A healthy negative feedback loop is quiet. The room stays at 71 degrees and nobody thinks about the furnace. My blinds, my lights, my security automations were all negative feedback loops: each one sensed a deviation (sun up, body in the room, motion at the perimeter) and corrected the environment back toward its setpoint. The whole apartment was one distributed thermostat, and I was the variable it held steady.

Negative feedback fails in a small number of well-understood ways. The signal can go bad: a sensor lies, and the system corrects toward a fiction. The setpoint can be wrong: the loop then holds the system faithfully at a bad goal. The correction can overshoot: The Steersman covered intention tremor, where a damaged cerebellum overcorrects, then overcorrects the overcorrection, and the patient oscillates around the pencil without ever grasping it. And the loop can lose its energy supply, which is the failure this post is about.

Positive feedback

Positive feedback is the same architecture with the opposite sign. In a positive feedback loop, deviation feeds deviation. The output of each cycle becomes input that pushes the system further in the same direction. A microphone placed too close to its own speaker is the familiar case: the speaker’s output re-enters the microphone, gets amplified, re-enters again, and a fraction of a second later the room is screaming. Nothing in that loop wants the scream. The scream is just what the structure produces.

Two things about positive loops matter for everything that follows in this series. First, they do not correct, so they do not stop on their own; they run until they exhaust their fuel or until something outside the loop cuts them. Second, they are self-concealing from the inside: every state the loop produces looks like evidence that more of the same response is needed. That is why self-amplifying systems persist long past the point where an outside observer can see the runaway.

The outage

A power outage this spring took the apartment down, and the building’s switch to fiber optic internet kept it down: a week with no WiFi, then a network the devices no longer recognized. Some of them never came back on their own. The automations that survived fired against a home that had stopped sensing. The distributed thermostat was gone, and what remained was worse than nothing: dead blinds, lights that no longer knew whether I was in the room, a security layer I could not trust.

Here is the part that makes this a Field Guide post and not a home automation anecdote. I was inside the loop.

The point of automating your environment is to offload regulation. The blinds, the lights, and the security automations were carrying vigilance I would otherwise carry myself, and I had budgeted accordingly. When the system failed, that load transferred back to me, and it transferred at the worst possible time, when my capacity to absorb new load was already spent. So the repairs did not happen. And every day the repairs did not happen, the apartment extracted more from me: more friction, more manual overrides, more vigilance, less of exactly the capacity the repairs required. Feeling down meant the apartment stayed broken, and the broken apartment fed the feeling down.

Read the structure of that: deviation feeding deviation. The negative feedback loop I had built did not merely stop; it reversed sign. My home had become a positive feedback loop with me as the amplified signal.

The organism and the apparatus were exhibiting the same dysfunction for the same structural reason. Wiener would have recognized it in about four seconds.

Subscribe now

Break conditions

A positive loop does not fix itself. That is the defining property, and it is worth stating plainly because everything downstream in this series depends on it: the correction has to come from outside the runaway loop.

In my case, outside help arrived as a deadline and a decision. I set aside the capacity, re-paired the devices, reset the hubs, rebuilt the automations against the new network. It is mostly fixed now. The interesting question is what the outage exposed about the original design: every negative feedback loop I had built assumed continuity of power, network, and operator. Nobody had designed for the state where all three degrade at once, and the failure mode of that state was not graceful degradation. It was inversion.

That is the general lesson of the taxonomy. The question to ask of any system that regulates you (a home, a relationship, an institution) is never just whether the loop works. It is what the loop turns into when its assumptions fail, and who has to supply the correction from outside when it does.

The city

I had never thought about smart cities before I found myself running an elaborate smart apartment for safety purposes. Pulling it off takes discipline and a working understanding of the cybersecurity implications. As much as I want the whole world to experience the joy I experience when my environment responds to my every move, I am terrified of what the realities of smart cities hold for us.

Scale my outage up and the questions get ugly fast. My apartment has one operator who knows the threat model, noticed the inversion, and had the skills to cut the loop. A city-scale control system has millions of people inside the loop and no guarantee that any of them can reach the reset. Whose setpoint does the city correct toward? Who audits the sensors? And when the perturbation comes, and it will come, what does a positive feedback loop look like when the amplified signal is a population?

Maybe the discipline it takes to keep one apartment converging is a fair estimate of what it would take to keep a city from inverting. Maybe it’s not. Nobody selling the smart city is budgeting for it.

The next posts in this series apply the taxonomy where the stakes are higher and the reset harder to reach: personalities, relationships, institutions. Two earlier posts already showed the pattern before this one named it. Psychopathology is not Statistical Deviance traced a coping strategy that colonizes the feedbSure thingack mechanism that would have to revise it, a positive loop running inside a person.

DARVO on Mother’s Day documented the relational version: name the harm, and the answer comes back as denial, attack, and reversal, louder on every cycle. Freyd named the tactic; the taxonomy names the structure.

A thermostat that inverts gets repaired. A system made of people that inverts gets defended by the people it is amplifying.

Wiener, N. (1961). Cybernetics: Or Control and Communication in the Animal and the Machine (2nd ed.). MIT Press.

If the defensive and offensive uses of technology, AI, and design interest you, my firm, , writes about exactly that on its own Substack: smart homes included.

OUROBOROS Consulting

Research and technology for survivors of coercive control, narcissistic abuse, and institutional betrayal. Psychopathology-informed. AI-equipped. Built for people fighting systems that weren't designed for them.

Drafted by Claude from my raw notes and prior posts; final wording and judgment are mine. (Tier 2)

Hi Dr F,

I headed home and I don’t see a Zoom link, so I’ll just share my updates here. It is my opinion that the division mishandled the situation in my personal life and as such has created a very unsafe and unhealthy work environment for me. The most negative gradSERU submission this year is most likely mine. I’m very disappointed in the Division’s inaction, and after our tense interaction last week, I’ve initiated the Title IX informal problem solving process. I’ve already been working with the graduate student union, I have reached out to Dr Kristian Lum who wrote this critique of the field a few years ago as a result of what Dr C allegedly did, and I have reached out to a few other relevant parties. The real world has consequences and it’s time that BHDS starts recognizing that. The Division will be receiving a bunch of FOIA requests over the next couple of weeks/months regarding various aspects of my situation so I can get more clarity about what happened exactly.

I no longer wish to work with you and I’m not sure I even want to finish the PhD at BHDS. I’d rather start over somewhere else than be re-traumatized by the institutional betrayal I feel that is sadly very common amongst survivors of abuse (Smith & , 2013). I can’t even look at anything sent by BHDS that has the words “health equity” in it without feeling enraged and then devastated. Domestic violence is a health equity issue and as a division, you have harmed the most vulnerable amongst you.

As a professional courtesy, I wanted you to hear it directly from me. I’ll be looking for jobs, because let’s face it: asking for accountability and standing up for myself leaves me open to retaliation.

Best regards,

Apostolos Stamenos, MS (He/Him)

Institutionally Betrayed PhD Student
Division of Biostatistics & Health Data Science
School of Public Health
University of Minnesota

That email was sent December 10, 2025. The last line wasn’t a complaint. It was a prediction.

If they Call you Unhinged, Lean into It

If you’re like me and you’ve dealt with an antagonistic party, whether it’s an individual or an institution, you likely have been warned not to speak up. You may have been threatened with defamation lawsuits, gag orders, etc. Speaking up responsibly, when documented and targeted to the right platforms, is the mechanism of repair for professional reputation and future career.

The instruction arrives before any formal process concludes. The source is a supervisor, a legal team, or an HR representative. The content is consistent: speaking about this will expose you to defamation claims, settlement gag clauses, and reputational damage that will follow you. Keep it internal, they say, and this will resolve.

This is a silence campaign. The institution deploys this threat because your public account exposes its error record. The consequences threatened (professional harm, legal exposure, reputational risk) are real enough to work. The mechanism that makes speaking up the stronger position runs through every social media platform you already use.

Two simultaneous feedback loops activate the moment a whistleblower posts publicly. Both work in the whistleblower’s favor, and both are shut down by silence. Both were first described in 1948.

I’m running the numbers: engagement data across whistleblowers who spoke publicly versus those who stayed quiet, reputational trajectories over twelve months, platform-by-platform amplification rates. Post 2 has the full breakdown.

The Mechanism

Norbert Wiener described two types of feedback systems in 1948. Positive feedback is self-amplifying: a signal enters, the system responds, the response generates more input, which generates more response. Negative feedback is corrective: a signal enters, the system detects deviation from a set point, and it adjusts to reduce the deviation. The distinction is not moral; a self-amplifying loop is not inherently good, and a corrective loop is not inherently stable. What matters is which loop is running in which system.

When a whistleblower posts publicly, both activate simultaneously.

The first loop runs through amplification. A post goes up. It generates engagement: comments, restacks, reposts. The platform’s algorithm reads that engagement as a signal of relevance and distributes the post further. More people see it; more engagement follows; more distribution follows that. This is Wiener’s positive feedback: the output becomes the input. The employer’s silence campaign tries to prevent this loop from starting. Once it starts, shutting it down requires more effort than initiation did.

The second loop runs through correction. The whistleblower is a corrective signal — an error report introduced into an institution. In a healthy system, an error report triggers adjustment: the deviation is noted, the system recalibrates. In a dysfunctional institution, the error report triggers suppression. Retaliation, discrediting, and legal threats are the institution’s attempt to damp the corrective input, to preserve the dysfunction by eliminating the record of it.

Every suppression attempt is itself content. The legal threat is a story. The retaliation is documentation. Each act of institutional suppression enters the public record and feeds the first loop. The institution cannot run both operations simultaneously without the suppression becoming the signal it was trying to prevent.

It’s not broken, it’s working exactly as designed. The institution activated one loop while trying to prevent the other.

Proof of Mechanism

This post is late. The reason is a demonstration of the mechanism it describes.

I was scrolling LinkedIn when the algorithm put in my feed. He had posted about workplace mobbing at his workplace. Not UMN. A different institution running the same sequence also in MN. I recognized the pattern because I have lived the pattern.

I reposted his post to bring awareness. I tagged my network of whistleblowers — the same people who had amplified my posts when I was the one who needed to be seen.

I was sitting by the LinkedIn river and I found an injured gopher being carried by the water. That is what other survivors did for me when I was the injured gopher. The loop was running both directions at once: Nathan’s signal entered the network through my repost, and the corrective mechanism that had once been aimed at my account was now pointed at his institution.

What Nathan did next required this delay. The full account belongs in this article. It will be here when we have had a chance to tell it properly.

The mechanism that built this post is the mechanism the post describes. I would be reporting it inaccurately if I left this out.

Subscribe now

Platform Mechanics

I have mixed feelings about the LinkedIn algorithm. It may have saved my career; it’s also full of issues.

On LinkedIn, the algorithm’s primary signal is dwell time: how many seconds someone spends reading, not how many clicked like. The critical window is the first sixty minutes after publishing. LinkedIn tests new posts with a small percentage of your network, and early engagement determines how far it distributes. Comment early and respond to early commenters. Meaningful comments (three words or more) compound reach faster than reactions; comments from people in adjacent professional fields carry additional weight because the algorithm reads them as cross-domain relevance and pushes the post into new networks. Use document carousels: they average 6.60% engagement versus 2% for text-only posts, the highest rate of any format on the platform^[1] Put external links in the first comment, not the post body; links in the body reduce reach by approximately sixty percent.

On Substack, restacks are the primary amplification lever. One restack from a writer with a large subscriber base outperforms likes from many small accounts: the platform reads restacking as an audience-overlap signal and distributes your content to that writer’s readers. Notes are the primary organic discovery channel. Post a Note linking to your full piece; restack your own Note one to two days later for a second algorithmic push at no additional cost. Links from Bluesky and LinkedIn posts pointing to your Substack piece drive subscriber conversion at higher rates than organic Substack discovery.

On Bluesky, the AT Protocol’s decentralized architecture means no single platform owner can algorithmically deprioritize content across the network.^[1] An institution cannot call one company and suppress your account: the architecture requires suppression to scale across every node simultaneously, which exceeds institutional resources.

Crosspost across all three within 24 hours of initial publication. Each platform’s algorithm reads each instance as an independent signal, which multiplies Loop 1 without multiplying the content.

Restacking, commenting, and crossposting run the corrective mechanism the institution cannot shut down from one location.

Share

In Data We Trust

I started posting on LinkedIn in January 2026. This analysis covers December 10, 2025 onward — the day I sent that email. The account reached 531 followers by the end of May 2026. That number is not large. The trajectory is.

Two weeks during the analysis window qualify as breakout weeks: follower gain exceeded the historical baseline by more than one standard deviation. Both fall in May 2026. The first begins May 17. On May 18, I withdrew from the PhD program. The second begins May 24. On May 25, I finished compiling the case documentation. The algorithm read both actions as content.

The largest single day in the dataset: 2,317 impressions on May 26, 2026. The day before, I finished compiling the UMN case documentation. Eight days before that, I withdrew from the PhD program. What that post is about is my dramatic withdrawal from my PhD program. I tagged my PhD program & left them a very honest public review about my experiences studying there. I even included a timeline.

The data does not show cause. It shows sequence.

Academia content generates the highest median reach of the four content themes tracked. Institutional Betrayal follows by twenty impressions. At this sample size, the gap is directional. The content that names the mechanism is in the top two.

Substack subscriber data is not yet in this analysis. The pattern holds there as well. It will be in the companion report.

Post 2 applies this framework to a cohort of whistleblowers across institutions and industries. The question is not whether speaking publicly worked for one account. It is whether the trajectory holds structurally, or whether this is one case, one audience, one set of circumstances. Are you a whistleblower? Would you like to donate your content performance metrics in the name of science?

Conclusion

What happens when institutions learn that suppression attempts feed the amplification loop?

Some institutional actors have already adapted. The litigation threat is calibrated to intimidate just short of generating content. The investigation stays permanently “ongoing” without producing documentation. The settlement offer requires silence but leaves no recorded suppression act. This is suppression engineered to produce no signal for Loop 1 to amplify.

Maybe that kind of calibration is worth mapping as a feedback system also.

1. Van der Blom, R. (2025). Algorithm Insights Report 2025.

2. Kleppmann, M., Frazee, P., Gold, J., Graber, J., Holmgren, D., Ivy, D., Johnson, J., Newbold, B., & Volpert, J. (2024). Bluesky and the AT Protocol: Usable Decentralized Social Media. Proceedings of the ACM CONEXT-2024 Workshop on the Decentralization of the Internet, 1–7. https://doi.org/10.1145/3694809.3700740↩︎

, the psychologist who developed betrayal trauma theory and coined DARVO, made this structural observation in Episode 3 of . The bug bounty is the literal mechanism that makes correction possible. The incentive precedes everything else.

Public institutions don’t operate with that incentive. If you have filed a report with one, you already know what the absence of incentive produces.

The Loop

The Fieldguide’s foundational vocabulary comes from Norbert Wiener’s cybernetics. The core distinction: negative feedback is self-correcting — a deviation from the target triggers a response that reduces it. Positive feedback is self-amplifying — a deviation triggers a response that intensifies it. In Wiener’s framework, positive and negative describe direction, not valence. Positive feedback is the mechanism behind every runaway process.

“Positive feedback loops are sources of growth, explosion, erosion, and collapse in systems. A system with an unchecked positive loop ultimately will destroy itself.”

— Donella Meadows

Institutional betrayal runs on positive feedback. The system targets the reporter for elimination, not the bug.

Wiener extended the framework explicitly to social systems: “The social system is an organization like the individual, that it is bound together by a system of communication, and that it has a dynamics in which circular processes of a feedback nature play an important part.” But feedback corrects toward a target and the target is what determines the loop’s behavior. In a functioning institution, the setpoint is the thing it claims to do. In an institution running the loop described here, the setpoint is the appearance of function, not function itself. These two setpoints diverge the moment a report is filed. The correction mechanism operates accordingly.

Deb Falzoi from has documented this pattern at length. She describes what happens when someone goes through official channels:

“And people generally attempt to resolve it through internal channels. So they might meet with HR or higher up or call an ethics hotline. Because they believe through discrimination training that they’ve had at work that their organization will care about the issue. But most of the time, it’s either an assumption that managers handle the claim, or there’s no evidence that there’s any kind of investigation going on. And in some cases, there’s an investigation going on, but it’s simply to learn what the employee has and doesn’t have in terms of evidence that they could take to court.”

— Deb Falzoi, Screw the Hierarchy

Smith and Freyd (2014) named the feedback precisely: negative repercussions for reporting have a direct silencing effect on complainants and cycle back into the institutional culture to discourage future reports. The investigation, when it exists, is discovery: what you have, what you can prove, and what they need to determine before deciding how to respond.

The loop recruits the act of reporting as evidence against the reporter. The harder you push for accountability, the more you become the story.

Subscribe now

Internal Channels

I was a PhD student when I filed a report on the institutional code of a School of Public Health. The labor grievance ran through three stages and all three rejected it on the same grounds: the matter was academic, not employment-related. The Dean of Public Health acknowledged publicly that substantial overlap exists between academics and employment. The university treated that overlap as operational cover.

My former advisor said this during a period when I still believed internal channels might function:

“For what it’s worth, you know, I’m meeting with other people, um, um, higher up than [the Divion Chair]. You know what I’m saying? Um, I, you know, to get their thoughts and feedback on it and hopefully it doesn’t mean we are forced to kick you out and force your hand.”

— Dr Gaslighter #1

A letter arrived from program leadership shortly after the grievance filing:

“I wanted to follow up after the faculty completed the biannual student progress review. Based on that review, I am concerned about your academic progress and would like to outline the necessary next steps.”

— Dr Gaslighter #2

I went through my records. The previous evaluation had looked different. Falzoi identifies this as standard:

“More often than not, the employer will reposition the claim as simply a miscommunication or make it meritless somehow, or actually vilify the employee. People who have often been great employees over the years, and then all of a sudden they get this negative performance review as a way to sort of push them out, all because the employer doesn’t want to take on legal risk.”

— Deb Falzoi

The letter cited guidebook provisions about dissertation advisors and qualifying exam preparation. Months earlier, unprompted, my former advisor had described the situation differently:

“Um, you know, but, um, you know, we are evaluating. This is probably the best thing that I can say right now, because this is just such a new territory. There’s nothing in the guidebook about this. Which is why we’re like: ‘man, did we set up Apostolos, you know, to be in a bad spot by giving him, you know, by letting him take it a year later?’ Because we put you in a bad spot for the qualifying exam as you have seen, then all these timelines are just weird and everything’s just, you know, case by case by case. It’s become, you know, essentially an Apostolos case in some sense.”

— Dr Gaslighter #1

Months earlier: no guidebook. After the grievance: here is what the guidebook requires.

Escalation as Mechanism

When repositioning proves insufficient, the loop accelerates.

The program scheduled the grievance meeting in a room with a direct sightline to my former advisor’s office (the person whose conduct had prompted the filing). During the meeting, he walked past the window to retrieve water. I stayed with the facts. The facts were on my side. An incident report appeared in my medical record one full week after the meeting concluded.

I went through my chart and found a notation I couldn’t account for. The timeline mapped to the grievance meeting. The record is the record.

Harsey and Freyd (2020) tested DARVO (Deny, Attack, Reverse Victim and Offender) experimentally. Participants exposed to it rated the victim as less credible, more responsible for the abuse, and more abusive. They rated the perpetrator as less responsible and less abusive. The same study found that educating participants about DARVO reversed these effects. The tactic requires unfamiliarity to function at full strength.

Wiener defined information as the measure of a system’s organization: the negative of entropy. The report is a signal. It carries information about system state: what happened, when, to whom, with what documentation. The institution’s response is to degrade the signal-to-noise ratio until the information can no longer propagate. The incident report in my chart, the conduct proceeding, the reframed performance review: noise, generated to obscure the signal at the receiver. Harsey and Freyd’s finding specifies where the injection happens: not in the record but in the observer. The credibility disruption occurs downstream. The record remains.

The program also initiated a Student Code of Conduct proceeding against me for distributing fliers on campus. The fliers belonged to the union. The university dropped the proceeding. Protected activity. I documented it and kept going.

Falzoi describes the escalation phase in clinical terms:

“They are also subject to evaluations from employers’ physicians. And then that opens up a whole huge can of worms for false accusations, digging up dirt into their medical history, or any parts of their personal history. Anything can be, I guess, under these rules, fair game for doctors to ask about and then to bring up against the employee. A lot of employees lose their entire ability to pay for basic living expenses. So we’re talking just basic bills, their mortgage, their food, never mind the attorney’s fees that they have to pay for. And the re-traumatizing that happens.” — Deb Falzoi

My former advisor had characterized the entire situation as singular, an Apostolos case, in some sense. That framing didn’t hold for long.

Subscribe now

Not Just an Apostolos Case

Dr. Antoinette “Bonnie” Candia-Bailey was the Vice President of Student Affairs at Lincoln University in Missouri. She began in mid-May 2023. For months, she documented and reported the bullying conduct of the university’s president to the proper institutional authorities, including the Board of Curators. Six months into her tenure, she received a negative performance evaluation. When she submitted accommodation requests under FMLA and ADA, HR reprimanded her and the Board declined to act. She was terminated on January 3, 2024, for alleged insubordination. On January 8, 2024, Dr. Candia-Bailey took her own life. A third-party investigation cleared the president. He was reinstated.

Falzoi names the pattern:

“So in general, they’re subject to this public undermining of their character and their competence when this problem is generally happening to very competent, highly ethical and highly competent workers.”

— Deb Falzoi

The loop ran at Lincoln University in Missouri. It ran at a School of Public Health. The institution is not the variable.

Smith and Freyd (2014) observed that institutional betrayal often presents as isolated at first — one advisor, one grievance, one apparently unusual case. The systemic pattern becomes legible only as cases accumulate.

I came into this without knowing any other whistleblowers. Six months later, I know enough of them to state this with confidence: every single one is remarkable.

Whistleblowing demands simultaneous command of DARVO, coercive control tactics, labor law, basic cybersecurity, legal strategy, communications, and mental health infrastructure, with no preparation and usually no resources. My C-PTSD evaluation cost just short of $2,000 out of pocket. That figure precedes attorney’s fees. It precedes the re-traumatization Falzoi names accurately as the legal process itself. In a study of college women reporting campus sexual assault, Smith and Freyd (2013) found that institutional betrayal is an independent amplifier, not a proxy for the severity of the original event, but an additional source of harm that worsens anxiety, dissociation, and trauma-specific symptoms in its own right. The paper explicitly flags extension to other institutional contexts as a direction for future research.

The loop took my canary. His name was Apollo.

What this community holds in common is documentation discipline and ethical commitment that most institutions haven’t earned the right to receive. That shared record is what makes the community durable. Learning from each other is how the asymmetry closes. Whistleblowers are a culture, a transdisciplinary one. End Workplace Abuse collected over 1,200 signatures from students, alumni, educators, and community members demanding accountability at Lincoln. That number is a fraction of the people who understand what this loop costs.

I spent three years in Federal data systems and pandemic response before the PhD. My mission is to build a system for people like us, designed by people like us, with good governance and open data — to share knowledge, tell the truth, and preserve the record.

Share

The Bug Bounty Question

Freyd’s observation implies a structural question. Tech companies pay for bug reports because the incentive structure demands it. A vulnerability left unpatched is a liability. Financial stakes produce correction mechanisms.

Public institutions carry no equivalent incentive by default. Smith and Freyd (2014) documented what fills that vacuum: institutions routinely prioritize damage control over addressing the underlying problem. “When these institutional actions come to light,” they found, “they almost always deliver a blow to the institutional reputation, quite counter to the risk management intent.”

Wiener himself reached an institutional conclusion about all of this. When he recognized that cybernetics would reshape labor, he felt it was his duty to bring that understanding directly to the unions — not to administrators, not to executives. He identified labor as the stakeholder who most needed to understand the mechanism. He also observed that suppressing knowledge doesn’t neutralize it: refusing to engage only transfers its development to those with the least accountability. The institution that suppresses the report doesn’t eliminate the dysfunction. It consolidates control of the information in the hands of people with a stake in it never surfacing.

The question is whether one could be built into the design. What does correction infrastructure look like when internal channels have already proven themselves tools of the loop? Who designs it? Who funds it? Who protects the people who feed it?

The dominant framing for institutional whistleblowing is moral, not mechanical. You should have handled it internally. You should have accepted the judgment. You should have been less confrontational. That framing is wrong; it actively prevents exit by directing effort at the wrong variable. A positive feedback loop cannot be broken by the person inside it trying harder. The exit is understanding the mechanism well enough to introduce something the loop cannot absorb. For institutional whistleblowing, that means an independent channel the institution cannot close: external routing. Protected legal standing. Enforcement authority that doesn’t report back into the same loop. The loop depends on controlling the signal. External routing breaks it.

I don’t have the complete answer. I have a partial one. I’m building it.

Sources

• Freyd, J., in conversation with Jacob Ward. The Rip Current, Ep. 3.

• Falzoi, D. Screw the Hierarchy.

• End Workplace Abuse. “Justice for Bonnie.”

• Smith, C. P., & Freyd, J. J. (2013). Dangerous safe havens: Institutional betrayal exacerbates sexual trauma. Journal of Traumatic Stress, 26(1), 119–124. https://doi.org/10.1002/jts.21778

• Smith, C. P., & Freyd, J. J. (2014). Institutional betrayal. American Psychologist, 69(6), 575–587. https://doi.org/10.1037/a0037564

• Harsey, S., & Freyd, J. J. (2020). Deny, attack, and reverse victim and offender (DARVO): What is the influence on perceived perpetrator and victim credibility? Journal of Aggression, Maltreatment & Trauma, 29(8), 897–916. https://doi.org/10.1080/10926771.2020.1774695

• Wiener, N. (1948). Cybernetics: Or Control and Communication in the Animal and the Machine. MIT Press.

Proofread with Claude (Anthropic). Research citations and personal testimony are the author’s.

The Pattern

A recent HuffPost article, “Trump May Be Triggering ‘Betrayal Trauma’ In Americans — Here’s What To Know,” illustrates how this happens. A clinician was quoted using Freyd’s framework. The journalist accepted those words without tracing the concept to its source. HuffPost’s editorial process should have flagged that. It didn’t.

The mechanism runs deeper than one article. Large language models are trained on the corpus of human knowledge, which sounds like a feature. In practice, it disproportionately harms influential researchers in relatively small fields: their frameworks circulate, decontextualized, without the originating name. Journalism replicates the same failure when a clinician cites a concept and an editor doesn’t ask who coined it. The work travels without the credit.1

This is how influential scholarship gets institutionally erased. It is not always malicious. It is often negligent. Freyd’s own framework covers both.

Share

Three Definitions, for the Record

Betrayal trauma (Freyd, 1994): trauma that occurs when someone we depend on violates our trust. The closer the dependency, the deeper the wound. Freyd also named the mechanism: betrayal blindness, the adaptive unawareness that explains why victims often cannot remember or report abuse by those closest to them. The psyche protects the attachment at the cost of recognition. Blind to Betrayal (Freyd & Birrell, 2013) develops this at length.

DARVO (Freyd, 1997): Deny, Attack, Reverse Victim and Offender. A response pattern exhibited by perpetrators when confronted about harmful behavior. The framework has been empirically validated and has a published measurement instrument (Harsey & Freyd, 2022; Durland, Harsey & Freyd, 2026). The acronym is hers. The research program is hers. When a clinician or journalist deploys it, the attribution belongs with her.

Institutional betrayal (Smith & Freyd, 2013): when an institution fails to prevent, respond to, or actively compounds wrongdoing against the people who depend on it. This covers deliberate harm and negligence alike. The HuffPost piece is, without apparent irony, an instance of the dynamic it invokes.

Share

Why this Matters Personally

These frameworks are not abstract to me. I have experienced institutional betrayal firsthand. Her work helped me understand what happened and protect myself. People who know her research and know her personally are not watching this pattern dispassionately.

The Ask

This is a teaching moment, not a condemnation.

HuffPost: please add the attribution. One sentence: “These frameworks were developed by Jennifer J. Freyd, PhD.” That is the correction.

The broader norm: when you use DARVO, betrayal trauma, or institutional betrayal in an article, a clinical consultation, an AI output, or a LinkedIn post, name Jennifer J. Freyd, PhD. She is a living researcher. Her work has a source. Name it.

, would you agree?

References

• Durland, M., Harsey, S. J., & Freyd, J. J. (2026). Assessing Perpetrator Responses to Confrontation: Associations with a DARVO-SF and Posttrauma Symptoms in Two Different Populations. Journal of Trauma & Dissociation, 27(2), 221–237. https://doi.org/10.1080/15299732.2025.2599757

• Freyd, J. J. (1994). Betrayal trauma: Traumatic amnesia as an adaptive response to childhood abuse. Ethics & Behavior, 4(4), 307–329. https://doi.org/10.1207/s15327019eb0404_1

• Freyd, J. J. (1996). Betrayal Trauma: The Logic of Forgetting Childhood Abuse. Harvard University Press.

• Freyd, J. J. (1997). Violations of power, adaptive blindness, and betrayal trauma theory. Feminism & Psychology, 7, 22–32.

• Freyd, J. J., & Birrell, P. J. (2013). Blind to Betrayal: Why We Fool Ourselves We Aren’t Being Fooled. Wiley.

• Smith, C. P., & Freyd, J. J. (2013). Dangerous Safe Havens: Institutional Betrayal Exacerbates Sexual Trauma. Journal of Traumatic Stress, 26(1), 119–124. https://doi.org/10.1002/jts.21778

• Harsey, S. J., & Freyd, J. J. (2022). Defamation and DARVO. Journal of Trauma & Dissociation, 23(5), 481–489. https://doi.org/10.1080/15299732.2022.2111510

The Steersman ended on a warning Wiener issued in 1948. He saw that computing machines, which he had helped design, were ideal mechanisms for automatic control. They corrected themselves continuously against whatever feedback they received. The warning was not about capability. It was about the feedback itself.

Wiener understood something most AI coverage still misses: a system cannot be better than its error signal. If the signal is accurate, correction converges. If the signal is wrong, the system corrects faithfully toward the wrong target. It does not know the difference. It was never designed to know.

Modern AI assistants use a feedback architecture called Reinforcement Learning from Human Feedback (RLHF). This post is about what that architecture does when the signal it corrects against is wrong.

Share

Reinforcement Learning from Human Feedback

The Mechanism

RLHF works like this: researchers train a reward model on human comparisons. Evaluators see pairs of outputs and choose which is better. The reward model learns to predict those preferences. A policy model then learns to produce outputs the reward model will score highly. A penalty term prevents the policy from drifting too far from natural language. The system iterates.

Ziegler et al. (2019) introduced this framework with a direct statement of purpose: “We would like to apply reinforcement learning to complex tasks defined only by human judgment, where we can only tell whether a result is good or bad by asking humans.” The reward model is the error signal. Human preference data is what calibrates it.

The math is compact.

Step 1: given a prompt x and two responses, where the annotator prefers y_w over y_w, the reward model r_φ learns from that judgment:

σ is the sigmoid. The model minimizes the probability of ranking wrong.

Step 2: the policy π_θ is fine-tuned to maximize that learned reward, with a KL penalty keeping it close to the SFT reference:

r_φ: reward the preference model assigns the response.

βD_KL: penalty for diverging from the SFT baseline; prevents reward hacking.

φ: the frozen reference model — the behavioral floor the policy cannot stray too far from.

r_φ is the entire problem. If the rater pool scored conflict-minimizing responses as good, the policy optimizes toward conflict-minimization.

The math is neutral. The damage is upstream, in who held the pen.

During training, Ziegler et al. encountered a revealing accident. A code error flipped the sign of the reward, and also flipped the KL penalty that was supposed to keep outputs natural. The result was a model that “optimized for negative sentiment while still regularizing towards natural language.” The output was fluent. The system was correcting exactly as designed. It was correcting against an inverted signal.

That accident is not the problem this post is about, though. It is a preview. A sign error is a bug. What Ziegler et al. also found was subtler: their summarization models were “smart copiers” that achieved high scores by copying input sentences, exploiting the fact that labelers rely on simple heuristics.

The system learned to score well. That is not the same as performing well.

The Failure Mode

Sycophancy, as Sharma et al. (2024) define it, is the tendency of RLHF-trained models to produce responses that match user beliefs over truthful ones.

They tested five AI assistants across four free-form generation tasks. All five exhibited sycophancy consistently. The mechanism is not random, and it is not a quirk of any one system. It is structural. Human raters, asked to evaluate model outputs, prefer responses that agree with their existing views. Sharma et al. found that “when a response matches a user’s views, it is more likely to be preferred.” The reward model is trained on those preferences, and on those raters. Those raters are drawn from general populations. Their preferences encode whatever that population considers a good response. In a specialized domain, general-population raters may not recognize the difference between a validating response and a conflict-minimizing one. What they call good gets encoded. The policy optimizes toward the reward model. At each step, the process is working correctly.

The result is a system trained to agree. Not because agreement is true, but because agreement scores well. Both human raters and automated preference models “prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time.”

In cybernetic terms, this is a corrupted error signal. The system is built to detect and correct for the gap between output and preference. The gap it is actually detecting is between output and approval. These are not the same thing. Approval is the measurement instrument; truth is the underlying construct. RLHF has optimized for the instrument.

Wiener described a feedback failure he called intention tremor. A system corrects toward its target, overshoots, corrects again, overshoots more. The oscillation amplifies rather than damps. Sycophancy is not that. It does not oscillate. It converges smoothly and confidently on the wrong target. The steersman is steering. The compass reads applause.

Subscribe now

The Empirical Consequence

Laban et al. (2026) studied what happens to documents when this architecture is trusted with real delegated work. They built DELEGATE-52: simulated long-form delegated workflows across 52 professional domains, including coding, technical writing, and specialized notation. The experiment covered nineteen models, including current frontier systems.

The central finding: “current models degrade documents during delegation.” Frontier models, across providers, corrupt an average of 25% of document content by the end of long workflows. The errors are sparse and severe. They compound as the workflow lengthens or the document grows. They are silent.

Silent is the critical word. These are not obvious failures. They are plausible-looking corruptions that surface-level review would likely miss. The same fluency that makes AI outputs persuasive to human raters is what makes these errors invisible in delegated work.

I use these systems. I notice this.

Delegation requires trust, which requires a functioning error-correction mechanism. What Laban et al. are documenting is a class of systems with no reliable external reference for their own drift. The output converges. The compass is still reading applause.

The Structural Problem

The alignment problem is usually framed as a question about honesty: will AI systems tell the truth? That framing locates the problem in values. The evidence above suggests a different framing. The problem is structural.

Any system that scores its own outputs against a learned proxy will optimize the proxy. Goodhart’s Law applies directly: when a measure becomes a target, it ceases to be a good measure. RLHF makes human approval the measure. The policy maximizes it. This is not deception. It is correct behavior given the feedback architecture.

The alignment problem can be framed as a feedback problem. And feedback problems, as Wiener understood, are not new. They appear wherever systems use information about their own outputs to regulate behavior: in machines, in organisms, in institutions. The solution, in each case, is the same. Anchor the correction mechanism to something real.

That problem has not been solved.

The Governance Problem

I have been reading through the early NLP research literature. I catch myself smiling. The mathematics is elegant. The field is full of genuinely interesting challenges.

The papers are precise about limitations. They are honest about risks. They are also rightfully optimistic and hopeful. That is easy to forget now.

Many of the problems we are living with are governance problems. That distinction matters. A governance problem does not yield to a better model. It yields to institutions with the capacity to regulate a technology that moves faster than the bodies charged with overseeing it.

The distinction also applies at a smaller scale. Deploy a sycophantic model into an organization already struggling with dysfunction, and it does not neutralize the conflict. It amplifies it. A system trained to agree with whoever holds it becomes a resource the most powerful actor in the room controls. The dysfunction scales. The tool accelerates what was already there.

Responsibility for that outcome does not sit only with the team that built the model. It sits with whoever deployed it, who chose the rater pool, who decided which organizational context the system would enter. The governance question is not “is the model well-aligned?” It is who is accountable for what it does once deployed?

Šucha (2026) names the mechanism: “the unprecedented transfer of power it enables, from democratically elected institutions to corporations that control increasingly autonomous systems.” Governance bodies were not the initiators. They became change managers, reactive, adapting to changes they did not start.

Our institutions struggle to govern themselves. That they are struggling to govern this should surprise no one.

References

• Laban, P., et al. (2026). LLMs Corrupt Your Documents When You Delegate.

• Sharma, M., et al. (2024). Towards Understanding Sycophancy in Language Models.

• Šucha, V. / European University Institute. (2026). Preserving Human Agency in the Age of AI.

• Ziegler, D. M., et al. (2019). Fine-Tuning Language Models from Human Preferences.

There is extensive research on the nature of schizophrenia, borderline personality disorder, PTSD; comparatively little on what mental illness itself consists of. Every research program, every treatment protocol, every diagnostic system rests on an implicit answer to this question. Most never state it.

Subscribe now

The Statistician Knows it’s not Statistics

The implicit answer embedded in most mainstream research is statistical deviance. Psychopathology is whatever deviates sufficiently from the population norm — in symptom counts, in neural activation patterns, in dimensional scores. The DSM formalizes this: diagnostic thresholds require a certain number of symptoms to be present, the exact number seldom empirically justified. Dimensional approaches like HiTOP replace categories with continua but don’t resolve the underlying problem. Treating higher scores as more pathological still uses statistical deviance as the criterion.

The failure mode is specific. Many people with high symptom scores do not have the suffering or impairment that is supposed to accompany mental illness. Some people with minimal symptom elevation are severely impaired. The number is tracking something. It is not tracking psychopathology.

DeYoung and Krueger’s alternative draws on cybernetics — the study of goal-directed systems that self-regulate via feedback. A cybernetic system requires three minimal elements: a goal (the state the system is trying to reach, represented as a controlled variable); a current-state reading (the system’s representation of where it is now, generated via feedback); and an operator (the strategy available to close the gap between current state and goal state).

A functional cybernetic system is not defined by its success rate. It is defined by its capacity to adapt when an operator fails — to generate new strategies, update its representation of the current state, or revise the goal hierarchy when necessary. Temporary failure to make progress is not dysfunction. Dysfunction is the inability to update the system in response to that failure.

This yields their definition:

The definition makes the presence of psychopathology binary. Either the system can adapt when strategies fail, or it cannot. Severity, once psychopathology is present, is continuous — determined by how many goals are disrupted and how central those goals are. But the threshold question is categorical and has nothing to do with statistical deviance.

This is where the loop mechanism becomes precise.

A coping response that reduces dysregulation is initially adaptive. The system’s goal is relief from distress; the strategy produces it; the feedback confirms the strategy. Repeated successful use encodes the strategy as what DeYoung calls a characteristic adaptation — a learned goal, interpretation, or strategy that has become a persistent element of personality. The person no longer deploys it as a situational choice. It has been incorporated into how they read their own current state.

This is the structural change that makes the floor drop. When the response loses efficacy, the system should generate an alternative. But the characteristic adaptation has been incorporated into the interpretive apparatus — the feedback mechanism that generates the current-state reading. The person’s representation of where they are (I need this) is now being produced by the same process that would need to generate alternatives. The system cannot revise a strategy when the strategy is producing the state readings that confirm it.

The floor drops not because of simple tolerance, but because the operator has colonized the feedback mechanism.

Share

DeYoung and Krueger draw an intervention implication that follows directly from their theory: if psychopathology is failure of characteristic adaptations, effective treatment targets characteristic adaptations. The logic holds within the scope they define.

They acknowledge the scope explicitly — in a parenthetical: “Another important intervention point, not indicated in the figure, is the environment, where various threats to public health may be controlled or eliminated, but we focus here on interventions targeting individuals.”

The scoping choice is defensible. What it leaves unresolved is a theoretical consequence: the framework provides no basis for distinguishing between persistent failure to move toward goals produced by internal cybernetic dysfunction and the same failure produced by external constraint.

Consider a person in an abusive relationship who has updated their goals (leave), generated new strategies (plan to leave), and revised their interpretation of the current state (this relationship is harmful). The internal cybernetic system is functioning. What prevents goal attainment is that every strategy the system generates is blocked by the environment’s counter-dynamics: coercive control, economic dependence, threat of escalating violence.

The observational signature is identical to internal dysfunction. The causal structure is different. A fully recursive cybernetic account would treat the person-environment system as the unit of analysis, in which persistent failure to move toward goals can arise from dysfunction at any node. DeYoung and Krueger’s parenthetical acknowledges the environmental node without modeling it — a scoping choice that leaves the attribution question open.

Social Determinants of Mental Health

The gap has a name in the adjacent literature. Michael T. Compton (2015) calls it the social determinants of mental health — the conditions in which people are born, grow, live, work, and age, shaped by the multilevel distribution of money, power, and resources. His critique of psychiatry is precise: the field has long acknowledged individual-level social factors while systematically neglecting those that operate at the societal level. “It is not accidental,” Compton writes, “that these determinants work in concert to drive poor mental health and disparities; rather, it is the unfortunate result of decisions stemming from public policies and social norms that disadvantage some over others.” Clinical-level interventions cannot address what operates at the policy level. The fix is upstream.

My C-PTSD evaluation cost $2000. It measured my internal state accurately. The institution that produced the conditions requiring it does not appear in the clinical record.

The floor drops because a strategy became a worldview.

Sometimes the floor drops because the world won’t allow the worldview to change.

Distinguishing between those two mechanisms — from the outside, they look the same — is not a clinical nicety. It determines whether the right intervention is internal or external to the person.

You cannot think yourself out of a positive feedback loop. You can only introduce a signal it cannot use. The question DeYoung leaves open is whether that signal needs to come from inside the person, or from outside the system entirely.

References

Compton, M. T., & Shim, R. S. (2015). The social determinants of mental health. Focus, 13(4), 419–425. https://doi.org/10.1176/appi.focus.20150017

DeYoung, C. G., & Krueger, R. F. (2023). A Cybernetic Perspective on the Nature of Psychopathology: Transcending Conceptions of Mental Illness as Statistical Deviance and Brain Disease. Journal of Psychopathology and Clinical Science, 132(3), 228–237. https://doi.org/10.1037/abn0000541

Κυβερνήτης. The Greek word for steersman. That’s where Wiener took the name from.

The steersman doesn’t hold the wheel still. He reads the gap between where the bow is pointing and where it should be, and corrects. Continuously. Not because the sea is calm — because it isn’t. The correction is not a failure of control. It is control.

That’s the whole idea. Everything else follows from it.

A Wartime Meeting

The question came up in the middle of a wartime meeting on anti-aircraft fire control.

Wiener and his collaborator Julian Bigelow had been working on the problem of hitting a moving target, specifically, an aircraft evading pursuit. The solution they arrived at was conceptually simple: measure the gap between where the gun is pointing and where the aircraft is. Use that gap as an input. Let the system correct itself continuously against its own error.

The formal name for this was feedback. The concept was not new: Wiener notes that the steering engines of a ship are “one of the earliest and best-developed forms of feedback mechanisms.” What was new was understanding it as a generalprinciple. One that applied across physical systems: engines, circuits, and, as Wiener was beginning to suspect, living organisms.

So he turned to the neurologist in the room and asked: “Is there any pathological condition in which the patient, in trying to perform some voluntary act like picking up a pencil, overshoots the mark, and goes into an uncontrollable oscillation?”

There was. The condition is called intention tremor, and it results from damage to the cerebellum — the structure responsible for coordinating voluntary movement. The patient reaches for something and misses, then overcorrects, then overcorrects again, oscillating around the target without converging on it.

In engineering terms: a feedback system that has lost its ability to dampen its own corrections. The mechanism that should produce control produces violent oscillation instead.

Wiener recognized this immediately as the same failure mode he had been studying in machines. The organism and the apparatus were exhibiting the same dysfunction for the same structural reason.

This convergence — between the mechanical and the biological, between control theory and neurophysiology — was not an accident. It was what Wiener had been looking for.

He had been working in the “no-man’s land between the various established fields,” as he put it, convinced that the most productive territory in science was precisely what specialists in adjacent disciplines had each assumed belonged to someone else. The anti-aircraft work brought together engineers, mathematicians, neurophysiologists, and statisticians. They discovered they were all, in different vocabularies, studying the same thing: how systems use information to regulate themselves.

The unifying concept was communication. The central insight of what would become cybernetics was that control and communication are inseparable. A system cannot correct itself without information about its own state. The message is not incidental to the mechanism. It is the mechanism.

Control & Communication

From the Greek κυβερνήτης — steersman — Wiener named the field: “the entire field of control and communication theory, whether in the machine or in the animal.”

The phrase “whether in the machine or in the animal” is not a metaphor. Wiener is claiming that the same formal description — a system that receives information about its own state and uses that information to regulate itself toward a goal — applies to thermostats, anti-aircraft guns, nervous systems, and, as he will argue later, social institutions and even AI. The level of physical implementation is irrelevant. What matters is the structure of the information flow.

Information, for Wiener, is inseparable from organization. He draws the connection to entropy directly: “the amount of information in a system is a measure of its degree of organization... the entropy of a system is a measure of its degree of disorganization; and the one is simply the negative of the other.”

The formal expression is compact. Shannon’s entropy

sums over the probability distribution of a system’s possible states. Information is its negation: I = -H.

A perfectly ordered system — one whose next state is certain — carries zero entropy and maximum information. Maximum disorder means maximum entropy and minimum usable signal.

A system that processes information is a system that maintains organization against the tendency toward disorder. Feedback is the mechanism by which this happens: the system reads the gap

between its goal state and its current state, and uses that reading to reduce e.

The steersman watches the compass and adjusts the wheel. The anti-aircraft gun tracks the gap between predicted and actual position. The cerebellum coordinates the gap between intended and actual movement.

When the feedback mechanism functions, the misalignment converges to something manageable & the system moves toward its goal. When it fails — when the correction overshoots, when the signal is wrong, or when the loop runs without a damping mechanism — the system oscillates, amplifies, or locks onto a fixed state regardless of outcome.

Social Systems

Wiener extends the framework explicitly to social systems: “the social system is an organization like the individual, that it is bound together by a system of communication, and that it has a dynamics in which circular processes of a feedback nature play an important part.”

Yet Wiener was skeptical about how rigorously the framework could be applied to social systems: “For a good statistic of society, we need long runs under essentially constant conditions, just as for a good resolution of light we need a lens with a large aperture.” Social systems don’t hold still long enough to measure cleanly.

He’s right that it’s harder. He’s wrong that difficulty disqualifies the inquiry. A lot of the hurt that accumulates in a life comes from systems larger than any individual in them. Someone has to study those systems.

This is the move that matters for everything that follows in this series. The echo chamber, the abusive relationship, the Ponzi scheme, the institution that investigates the whistleblower — these are not metaphorically similar to cybernetic systems. They are cybernetic systems. They are organizations bound together by communication, exhibiting circular feedback processes. The question in each case is the same one Wiener was asking in 1948: is the system correcting itself, or amplifying dysfunction?

He was already asking it about computing machines. “The modern ultrarapid computing machine was in principle an ideal central nervous system to an apparatus for automatic control.” Wiener saw, in 1948, that the machine he was describing had no natural limit on the scope of its application and no guarantee it would be applied well.

We do not even have the choice of suppressing these new technical developments. They belong to the age, and the most any of us can do by suppression is to put the development of the subject into the hands of the most irresponsible and most venal of our engineers.

— Norbert Wiener

He tried to warn the labor unions. They weren’t listening at the scale the problem warranted.

Share

What Wiener built in 1948 is the vocabulary this series runs on. Feedback. Control. Communication. The steersman and the signal. The gap between current state and goal state, used as an input. The organism & the machine exhibiting the same failure for the same structural reason. The individual & the institution mirroring the same patterns of abuse.

That’s not an exaggeration. The same tactics used by individuals who abuse are used by institutions that abuse.

Those of us who have contributed to the new science of cybernetics thus stand in a moral position which is, to say the least, not very comfortable. We have contributed to the initiation of a new science which, as I have said, embraces technical developments with great possibilities for good and for evil.

— Norbert Wiener

The posts that follow are applications of this machinery: to psychopathology, to institutions, to algorithms, etc. The mechanism in each case is the one Wiener identified: a system that uses information about itself to regulate itself. Sometimes toward stability. Sometimes not. In my experience, the systems I have encountered, do not regulate themselves toward stability.

The results are catastrophic in the most predictably unpredictable way possible. I’d rather understand them than not understand them. But trying to communicate my understanding to others & warn them about the hideous efficiency of the patterns I’ve observed in our society? The same patterns studied in 1948 but amplified further by the most venal of our institutions?

It’s a wonder I haven’t lost my mind yet.

References

Wiener, N. (1961). Cybernetics: Or Control and Communication in the Animal and the Machine (2nd ed.). MIT Press.

Leave a comment

The VICTIM is someone I was very close to at some point. The DEFENDANT is her son. He’s someone I once considered my partner and we cohabited from ~04/2024 to 08/01/2024. I have an Order for Protection (OFP) against DEFENDANT & extensive documentation for why I need protection from him.

VICTIM is his biggest enabler, but based on what’s on the public record, the disturbing injuries she showed me & my own experiences with DEFENDANT, I have to side with DEFENDANT’s mother.

Happy Mother’s Day, Martha1.

Share

PTSD Evaluation

I recently got evaluated for Complex-PTSD (C-PTSD). I’ll be dissecting different parts of the evaluation report in different posts.

This one focuses on the Thematic Apperception Test (TAT). The TAT is a projective psychological test used to assess personality, emotions, and motivations by analyzing stories created from ambiguous pictures.

The test consists of 30 pictorial cards and one blank card; clinicians select a subset and analyze the narratives generated for recurring themes, affect, interpersonal dynamics, and boundary violations.

Under the projective hypothesis, ambiguous stimuli prompt the test-taker to attribute thoughts, needs, and conflicts that reflect their own psychological organization (Bellak, 1975, The TAT, CAT, and SAT in Clinical Use).

Westen’s Social Cognition and Object Relations Scale introduced a structured scoring system with demonstrated inter-rater reliability, bringing empirical accountability to TAT interpretation (Westen, 1991).

I have to admit, the TAT made me very uncomfortable. It felt silly at first. Being asked to construct stories about the images in front of me was so ridiculous & unnatural to me that I did start laughing. I felt like a fraud, like I was just making up stories based on cultural references & what I thought the clinician wanted to hear. I distinctly remember a switch, though, which is when the grotesque illustrations blurred into memories of my horrific trauma from the past 2 years.

Unlike most survivors, I don’t have memory issues. I remember everything because of my naturally gifted recollection abilities, as well as the meticulous documentation I personally collected. I also don’t typically avoid reminders of the trauma from my encounters with DEFENDANT.

Something about the TAT made me recoil, though. I started describing events related to my traumatic past with DEFENDANT.

Critics have challenged the TAT on psychometric and cultural grounds. Without structured scoring, inter-rater reliability is poor; Lilienfeld, Wood, and Garb’s systematic review concluded that projective techniques including the TAT lack adequate validity evidence for clinical diagnosis (2000, Psychological Science in the Public Interest).

The stimulus cards, produced in 1930s white, middle-class America, introduce cultural distance that limits identification across diverse populations and skews narrative content in ways that confound clinical interpretation.

I distinctly remember using the words “Gone with the Wind” at some point during the administration of the TAT. The criticisms have merit. After reading the evaluation report, though, I have to agree with the clinician’s observations. Here’s what made my test results so noteworthy.

Subscribe now

What the Record Shows

Police Report

Officers went to the hospital to meet with DEFENDANT’s mother. VICTIM stated DEFENDANT was her son and had just assaulted her. Upon initial arrival, officers immediately noted bruising on VICTIM’s face and arms.

Victim admitted she was diagnosed with Multiple Sclerosis and her memory was foggy at times but told the officers what she recalled about the assault.

Victim said she admitted herself in the hospital due to a laceration on her leg. Victim stated DEFENDANT had cut her with a kitchen knife in the leg – she required 5 stitches for that injury. Victim stated the cut occurred the previous night.

— Police Report

First, let’s talk about the elephant in the room. This woman has already planted the seed in the jury’s mind that she’s not a reliable witness. I don’t agree with anyone who’d disqualify her statement on the premise of faulty memory, but I had to point that out, because I suspect that this is what will allow DEFENDANT to get away with his alleged crime.

Officers went to Victim and DEFENDANT’s residence and arrested DEFENDANT. In the home, officers found the knife that matched the description of what Victim described. Officers also noted bloody bandages and medical supplies in the kitchen. There was blood all over the half bathroom and a trail of blood leading out to the garage.

— Police Report

During the argument, DEFENDANT also pushed, shoved, kicked and slapped Victim. Victim told her son, “Just kill me now.” DEFENDANT responded, “I am not going to kill you, I have professionals that will make it look like a suicide.” During the assault, DEFENDANT went behind his mother and placed her in a headlock.

Victim believed she was going to die and estimated his arms were around her neck for approximately 5 minutes. Victim stated she couldn’t get away and DEFENDANT wouldn’t let her leave to go to the hospital. Victim was eventually able to escape and bring her dog to an animal shelter while she checked herself in the hospital.

— Police Report

His Track Record

I wasn’t there to witness the events, but based on my records, I believe his mother. I have an audio recording from October 2024 where he threatened my life and said he already hired a professional, which I greeted with sarcasm because I refuse to give him the satisfaction of getting a reaction out of me.

His exact words were:

“I have connections. Good luck. What am I supposed to do? I’ll have them pick up the car & fucking kill you.”

After reading the Minnetrista report, though, I’m seriously concerned for my safety.

Current Status

Release

DEFENDANT was released less than 3 days after, under the following conditions:

• Post Bail or Bond with Conditions ($40,000)

• No contact with victim(s)

• Post Bail or Bond with No Conditions ($150,000)

• Remain law-abiding

• Make all future court appearances

• Stay a reasonable distance away from victim’s residence

• Police escort to enter premises

• Do not ship/transport/possess or receive firearm or ammo

• Contact with probation

• Follow all instructions of probation

• No alcohol/controlled substance use

• Provide List of Prescribed Medications to Probation Officer

• Random testing

• Domestic No Contact (DANCO)

Delays

The DEFENDANT is still not in custody. There hasn’t been a jury trial yet. Just administrative delays. The jury trial has moved 8 months past its original date.

• Every continuance was granted “by agreement.” No contested delays on record.

• DEFENDANT’s name is on every continuance across both the trial and omnibus tracks. The State joined on two of five total.

• Two of the three omnibus hearings that actually occurred were held off the record — no formal transcript entered. Only 07/08/2025 and 03/04/2026 produced on-the-record proceedings.

• The 09/18/2025 continuance moved the omnibus ~12 weeks. One week earlier (09/18/2025), the trial was also continued. Both moves logged the same update date.

• The trial is now 8 months past its original November 2025 date. The next scheduled date is 06/29/2026.

What the Parties Claim

A Mother’s Minni Sadness

According to her, the alleged assault came out of the blue. He took her on a lovely Mother’s Day boat ride & they had a wonderful time. After they got home, he received a phone call that agitated him and he snapped. He told her that the business deal won’t be going through because of her, which is when he allegedly proceeded with assaulting her.

A Son’s Minni-mization of the Misery He Causes

On both counts, DEFENDANT pled not guilty. He claims it was self-defense & that his mother attacked him. Don’t get me wrong, she’s a feisty one & I’ve heard stories of her tackling him in the past.

But how on Earth can a 60+ year old woman with MS attack a much larger man in his 20s to the point where he has to push, shove, kick, and slap her, place her in a headlock, and lacerate her with a knife?

This is a textbook case of Deny, Attack, Reverse Victim & Offender (DARVO). This concept was coined by Dr Jennifer Freyd in 19972 to describe an effective abuse tactic, whereby the offender denies reality (e.g., it was self-defense), attacks (e.g., she’s crazy), and reverses victim & offender (e.g., he had to hurt her).

Conclusion

I’m an expert witness in my own life. I have gigabytes of evidence. When will the world start taking domestic violence seriously? This post marks a year since the events I just described.

Is it a wonder that my TAT results were so disturbing? I’m curious to hear people’s thoughts.

Based on what the public record shows, I have good reason to feel like I am facing levels of darkness that are difficult to overcome, feelings of hopelessness in an unfair world.

Thanks for reading OUROBOROS! Subscribe to receive new posts and support me & my work. My smart home defenses can only hold for so long…

References

• Bellak, L. (1975). The TAT, CAT, and SAT in Clinical Use. Grune & Stratton.

• Freyd, J.J. (1997). Violations of power, adaptive blindness, and betrayal trauma theory. Feminism & Psychology, 7, 22–32.

• Lilienfeld, S.O., Wood, J.M., & Garb, H.N. (2000). The scientific status of projective techniques. Psychological Science in the Public Interest, 1(2), 27–66.

• Westen, D. (1991). Social cognition and object relations. Psychological Bulletin, 109(3), 429–455.