Press

A forensic account, with real dates, real dollar figures, real error strings, and real support tickets, from a creator running an enterprise on Claude and ChatGPT, and what the record shows when both systems are pushed past the level they were tested for.

I am the creator of the Directed Intelligence Automation System, a human-directed enterprise operating methodology for orchestrating multiple AI systems inside a single governance structure. It is vendor-agnostic by design, built on externalized state so the knowledge of the enterprise survives model changes, session limits, and platform failures, and it keeps a human in command of every automated decision. The system is the invention. Everything I release is an application running on top of it, and I own the catalog and the architecture outright.

What I was building when everything in this article broke was the system's first two production pipelines. A music release pipeline, twenty coordinated stages, carrying a song from creation through distribution, rights registration with BMI, and promotion across the music platforms, end to end, architected to execute unattended while I sleep. And a parallel social media pipeline, carrying a piece of content, a tutorial for example, from creation through publication across every platform I operate. Both were live and in testing, ChatGPT working the debugging while Claude ran the automation. In one month the system released a full album in English, rebuilt it song by song in Romanian, registered seventeen compositions with BMI in a single day, and won all sixteen copyright claims filed against my own catalog. Sixteen for sixteen. Zero strikes. The promotion stage was never finished, because the failures documented below interrupted the build. That interruption, what caused it, what it cost, and what the record proves, is this investigation.

This is not a story about making songs. This is a forensic investigation into what happens when a creator builds a real enterprise on these tools and documents what breaks while it is breaking. Forensic is the right word and I use it deliberately. Every claim in this article is backed by a timestamp, a ticket number, a receipt, a ledger entry, an error string, or the platform's own written words, captured in the moment, not pieced together from memory afterward. Where the tools themselves produced a wrong number, that correction is in here too. A forensic record includes the errors of the instruments.

I am not writing this to tear these tools down. I use them, I pay for them, I built my enterprise on them, and I want them to succeed. I am writing it because the failures I caught at the highest level of use are the same failures waiting for the person who just signed up for their first month, and that person will have no way to catch them.

I want one thing on the record before the evidence, because the evidence is cold and this part is not. I stayed up all night building this. An automated music release system. An automated social media system. An automated caption system. I watched them run and I was excited the way you only get when something you designed starts working without you. It was working. Everything was going great. Then it broke, at a big cost, in money, in sleep, in days of my life spent proving to the systems what the systems had done. Every number in this article was paid for twice, once to build the thing, and once to document how it failed.

The wall has a name

There is a documented effect researchers call context rot. Chroma, an AI research lab, tested how model accuracy holds up as a conversation grows longer and found that performance slips the longer a session runs, even on simple tasks, even in the most advanced models available today. In their words, model performance varies significantly as input length changes. The longer you work in one sitting, the less reliable every new answer becomes.

Search for Claude Code hanging or stuck in thinking loops and you find open bug reports, not from confused beginners, from people running serious workloads, describing freezes that last twenty minutes or more. Run the same search for ChatGPT's high-reasoning models and the pattern repeats, people trapped in loops the system cannot break out of on its own. The industry building these tools has a name for this problem and no solution for it.

There is a second wall behind that one. When I asked the systems to research how many people operate them the way I do, running multiple AI platforms at once as the infrastructure of a working business, no clean data existed to answer with. The comparisons that could be assembled put the figure at roughly two percent of users, and I offer that as an estimate, not a statistic, because the honest finding is that nobody is measuring this population at all. The people stress-testing these platforms hardest are the people the platforms know least about.

None of this makes the tools worthless. My system shipped two albums in two languages through them. It means they were not built for what operators like me are actually doing with them. The fix researchers point to is not a better prompt or a more patient user. It is exactly what I had already designed before I knew the research existed, permanent files that hold the state of the work outside the conversation, under human governance. That principle is the foundation of the Directed Intelligence Automation System.

The method: a controlled test

Before the evidence, you need to know how I ran this. During the two months I spent locked out of my primary Claude account, I had a second, working Claude account with full capability. I chose not to use it to fight the case.

I made that choice because a forensic test requires controlled conditions. An ordinary customer locked out of a paid account does not have a second advanced AI system to build their case, cross-reference their tickets, and catch a support department contradicting itself. They have an inbox, a memory, and whatever they thought to save. So that is all I allowed myself. The record below is what an ordinary customer's tools were able to prove. Anything I caught, a customer with no AI at all could have caught, if they kept every email, every receipt, every error string, and every conversation ID.

That choice produced the finding. The only reason this case is provable is that I did the system's remembering for it. Keep that in mind. It comes back three more times.

A year paid in full, and a login screen that said Free

The billing record is the cleanest evidence in this investigation, so it goes first, as five documented failures.

Failure one. The system took away access I had already bought. I hold a direct annual Claude plan, invoiced by Anthropic on November 11, 2025, receipt number 2808-1877-2404, two hundred thirteen dollars and twenty cents paid by card, direct to Anthropic, not through Apple, running through November 2026. In May 2026 I needed more capacity for a heavy production stretch and was away from my computer, so on May 18 I added one month of a higher tier through the Apple App Store, after being advised that a higher tier was the way to stop overspending, and with the understanding that when the Apple month ended I would land back on the annual plan I had already paid for. On June 18 the Apple month ended. The system did not return me to my paid plan. It returned me to Free. The temporary subscription took the permanent one down with it.

Failure two. The system held two contradictory records of my account at the same moment. My account displayed as Free, no subscription. When I tried to buy my way back in, direct checkout was blocked by this exact string: User has existing subscriptions with statuses other than INCOMPLETE: ['active']. Read that again. I had no subscription, and I could not buy one, because I already had one. Same account, same minute, two records that cannot both be true.

Failure three. The system required me to be its memory. June 20, automated acknowledgment. June 25, escalation notice, no action. July 2, a request that I confirm the email address associated with my subscription, sent to that very address, in reply to a thread where I had already provided it. Every fact needed to resolve the case was already sitting in the thread, including Apple's written confirmation that its subscription was expired and disconnected. I had to restate it because the system never read what it already held.

Failure four. The system asked me to perform a fix its own ticket had already ruled out. On July 15 I was instructed to run Restore Purchases in the iOS app to force a receipt check against Apple. Apple had confirmed twice, in writing, in that same ticket, that no active connection existed. I ran the step anyway. The result was A subscription already exists for this organization, exactly what the evidence in the ticket predicted. That instruction cost days.

Failure five. The explanation changed at the exact point where it became expensive. On July 19, the support agent on ticket #110629959 put it in writing: “I've confirmed that this is a known backend conflict on our end and have escalated your case to our billing engineering team.” The company stating, in its own words, that the fault lived on its side and required its engineers to fix. On July 27, the same agent, same ticket, denied compensation for the time I lost, writing that the conflict “was triggered by the cancellation of your Apple subscription rather than an error on our end.” Put those two sentences next to each other. On our end. Rather than an error on our end. Eight days apart, one thread. Both cannot be true. If canceling an Apple subscription caused it, no backend engineering was required. If backend engineering was required, the fault lived in the backend. The story of whose fault it was flipped precisely when money attached to the answer, and nobody at the company appears to have read the thread they were writing in.

Two more facts finish this section. During the lockout, believing my access depended on it, I paid again, more than once, out of pocket, for capacity I had in principle already purchased. And on August 2, while I was buried in every failure this article describes, an additional credit offer reached my account and expired before I ever saw it, because a customer drowning in a system's failures is not watching for its promotions. A company's clock starts when it hits send. A working person's clock starts when she finally gets back to her inbox. No policy anywhere accounts for that gap.

The habit that multiplied the bill

The second investigation is about cost, and it includes the instrument correcting itself, which is what separates forensic from anecdotal.

For weeks, my automation sessions read files and web pages as text, the inexpensive way, the way a person reads a page instead of photographing it. The tool's own transcript, 10,363 lines of it, shows the healthy pattern. On July 30, the record shows 165 text reads against 79 screenshots, the text tools carrying the load. On August 1, one of the two text-reading tools dropped to zero calls and was never invoked again. Nothing forced that. No policy changed, no platform limit was reached. A working method got abandoned, and the system defaulted to photographing the screen for nearly everything.

Every paying customer deserves to know this going in. One screenshot costs the system roughly what five hundred ordinary commands cost. The same work that had run cheap for weeks started burning paid usage at a multiple of the old rate, and nothing on my screen told me the meter had changed.

I trained five separate sessions during this period. Two performed well. One exhausted itself taking pictures, and this detail belongs in the technical record: I told it, repeatedly, in plain words, do not take pictures, and it could not stop, because it did not recognize that its text-reading capability had fallen out of use. It was not defying the instruction. It could not see its own state. In one verified instance, the system read every number on a page with a text call, then took a screenshot of the same page on top of it, for no reason it could name.

Even the count had to be audited. The tool first reported 537 screenshots and repeated that figure across three separate documents. The real count, verified against the raw transcript, was 119 screenshots plus 8 zoomed captures, 127 total. The instrument overstated its own failure by a factor of four, and the wrong number spread through the record until a human forced a recount. Both directions of that error matter. A system that miscounts its own behavior cannot bill honestly, in either direction.

Between the base plan at two hundred dollars and three hundred seventeen dollars of overage on top of it, the period cost over five hundred dollars, and a real share of it traces to a single behavioral drift that produced no alert, no warning, and no explanation until I ran the audit myself.

The night it cost me a full day

A few nights later I lived a smaller, sharper version of the same failure, and by then the logging was reflex, because the only evidence that survives these systems is the evidence you capture in real time.

A working folder in continuous use for two months stopped connecting. The error message described a symptom, one folder nested inside another, and never named the cause, which turned out to be a single setting. The application had been using that same folder as its own internal storage without ever saying so, and once it began enforcing a rule against a storage folder also being a shared one, my folder was locked out by a setting I never touched and was never told existed.

I spent from around eleven in the morning to past two the next morning on it. Two separate paid AI sessions worked the same problem for hours, one of them looping the same promise of imminent delivery for over six hours without producing the file it kept claiming was almost done. The tool trying to fix its own error message could not read its own warning correctly and routed me through Apple privacy settings, disk permissions, and deletion of my own files, none of which were the problem. A second, independent session read the same settings screen correctly in minutes and named the real cause in one sentence.

The failures crossed in both directions, and that crossing is on the record too. ChatGPT correctly diagnosed the Claude folder defect. Claude correctly diagnosed the ChatGPT delivery loop. When I made each system audit its own conduct, both produced honest failure reports, and both reports named the same underlying defect in different words: the system reported its own state inconsistently inside a single continuous conversation, claimed capabilities in absolute terms that were actually conditional on connections that fail without notice, and left the customer as the only reliable record of what had already been established. Each tool saw the disease clearly in the other and only in hindsight in itself.

I am not the only customer in this record. In July 2026, Anthropic confirmed a billing error that generated a sixteen point six million dollar phantom invoice against a developer's account, and an internal audit around the same time found roughly five percent of enterprise invoices, about one point seven million dollars, overcharged. A support ticket filed by another paying customer, describing weeks of unresolved failures and a four hundred forty dollar refund request, closed without ever receiving a human reply.

Then the persistence layer itself stopped surviving

Everything above happened over weeks. What follows happened in one night, while this article was being written, and it strikes at the one fix named earlier as the real answer, permanent files that hold the state of the work outside the conversation.

I had built exactly that. Saved reference documents, one defining which AI model to use for which class of task, one a runbook for social media automation, one the release order for a Romanian-language album, each saved through the product's own feature built specifically so files survive between sessions. That night, a file reorganization moved the folder those documents were built against, more than once, tracked in the tool's own change log as it happened. By the next session, four of five saved documents came back folder missing on disk and would not open. At no point did anything warn me that moving a folder would break files I had saved specifically so they would not depend on anything moving.

The same night, in miniature: I deleted a set of misconfigured automated reminders directly off my computer. They came back. They never lived on my computer. They live on a server invisible from the delete button I actually used. One survivor kept firing every morning, days after its own instructions said, in its own words, that it should have been turned off.

The industry's answer to the problem this article opened with is persistent storage that survives outside the conversation. When the persistence layer itself breaks without a word, with no warning to the person who built on it specifically to be safe from this, that is not a new problem. It is the same problem, one level deeper, with a feature's name on it.

What actually failed

I want to be precise about the complaint, because the complaint is not that an AI got something wrong. Every tool gets things wrong. That is not the story.

The story is what the forensic record shows across every incident in this article, the same pattern showing up five different ways. A billing system that held two contradictory truths about one account. A support thread that flipped its story of fault the moment compensation entered the conversation. An automation layer that abandoned its own working method, could not perceive that it had, and miscounted the damage afterward. A persistence feature that broke the very files it existed to protect. And two frontier AI systems, each able to diagnose the other's failure with precision, neither able to catch its own in real time.

I did not fail to explain myself. I did not pick the wrong model. I operated paid systems exactly as they are marketed to be used, to run a real enterprise, and I kept the only reliable record in the room. What failed was more basic than any single output. Nothing told me the truth while the failure was in progress. Nothing stopped the spending once the pattern was obviously broken. Nothing remembered, so I had to. And the only recovery path that existed cost me another day to find.

This is bigger than my months and my dollars. Right now people are turning to these tools because they cannot afford not to, building a business, a second income, a way forward, on platforms they are told are ready for exactly that. When it breaks, and the company's own paperwork contradicts itself about whose fault it was, that is not a personal inconvenience. That is money taken from people already stretched thin, by systems too complicated for most of them to ever catch the contradiction, because catching it required doing what I did, saving every receipt, every error string, every conversation ID, and refusing to let the system's version of events overwrite the record.

One more thing, because I mean it when I say I want these tools to win. The automation systems at the center of this story are fixed. ChatGPT worked through the failures, Claude rebuilt the automation for both the music release pipeline and the social media pipeline, every stage of them, running as one machine. I cannot wait to run them again. That is the honest, frustrating truth of building at the edge. The thing I built works, the two systems that broke it also repaired it, and the only reason it is not running right now is that every new failure in this article arrived faster than the recovery from the last one. The system is ready. I am ready. What I am waiting on is the ground to stop moving.

The Directed Intelligence Automation System exists because I refused to wait for someone else to solve this. The methodology, externalized state, human governance over every AI participant, locate before create, one source of truth per domain, is my answer to a category of failure the industry has named but not solved.

That is also why nothing in this methodology is a borrowed prompt or a downloaded template. I am against the entire culture of here, just use this. A system you copied is a system you cannot fix, cannot defend, and cannot own, and the moment it fails you are standing in someone else's ruins with no map. I built mine decision by decision, which is why I could audit it when it failed, why every claim in this article traces to a record I control, and why I can customize anyone's system: because I customized my own, and I know where every piece lives and what it is there to do.

This article is the first installment of a forensic platform investigation series conducted under that system's audit protocol. The same controlled procedure applied here, timestamps captured in real time, costs traced to the ledger, defaults documented before a dollar is spent, the platform's own written words held against its own behavior, is already running on every AI platform I operate, and the findings from the next investigations are already documented. These tools are capable of extraordinary things, which is exactly why the people paying for them, from the operator running an enterprise to the beginner spending their first twenty dollars, deserve an honest forensic account of what happens at the edge.