Gate 2

A Dance of Efficiency and Ethics

Gate one was the greeter and the graciousness encountered at the front door of a destination restaurant. We already moved through the space that is the front of the house. Please join me as I push through the swinging door into the kitchen, the back of the house where the production happens. This is where, at the end of the night, the shovel and hose come out to clean the floor.

I started on a Windows 10 laptop with 32 gigabytes of RAM, a machine I had used for construction CAD drawings. It was a well-equipped workhorse, but I was bumping up against its limitations for this new work. In retrospect, one of the limitations was the nested subfolder structure: I learned that removing them, “flattening,” was requisite to diligent stewardship of a public corpus of documents. My husband, a first-generation Dungeons & Dragons nerd with all the computer literacy that entails, thought his newer Windows 11 machine would be an upgrade. He wiped it clean, reset it, and made it available to me. He was absolutely right. The Copy as Path function in Windows 11 alone was a game changer for writing code.

Those first days were brutal. I was using Claude to write PowerShell scripts, and it took hours and scores of foundation building attempts. Claude kept handing me code with specs to fill in, and I hit a learning curve that was not a curve at all. It was a vertical wall and I needed fall protection. I was the only person who thought it was possible, with every AI thread and every knowledgeable person telling me to stop climbing. I learned the value of rules of engagement. No scripts written before specs are defined, in terms I can verify, and provided.

The scripting started as keyword searches. Windows File Explorer was dead in the water for querying across a flat universe of 1.8 million files. I asked more questions. Claude wrote PowerShell and Python scripts for Named Entity Recognition searches, keyword proximity windows, and text extractions. I tested it, not once or twice, but enough times to make an honest assessment. It worked. But it was clunky and unsustainable for me alone to be the source for every search request a research team needed. I would become a single point of dependency, a human query engine. That was not going to scale.

I had learned about the Everything search app by asking for free alternatives to file Explorer, and it became my periscope into the flat universe. No human can browse a directory of that size. I set Everything us as an umbilical to the flattened directory. If I need to search anything else, I have to deliberately dial out the scope. Home key takes you to the head of the folder, End key takes you to the tail. I learned the value of Boolean commands. Everything app gave me machine‑speed search capability so I could finrd what I needed without scrolling through an ocean of files.

But Everything still lives on my local machine. It is not a shareable tool. It is my personal interface to the source files, and it cannot translate into a public-facing database that citizen researchers with limited resources can use. It was the next design problem that needed to be solved.

I started noticing search boxes. They are everywhere if you look for them. My Perplexity account has a search box for the thread history. My Signal app has a search box for the chats. And I'll be damned, Dropbox has a search box. I started asking more questions. How does Dropbox search behave? If it has a search box, and that search box can act like a periscope the way my Everything app does, then what are the true limitations for getting the entire corpus into Dropbox?

I kept asking AI, and I kept being told no. The ceiling was 300,000 files. Dropbox could not be pushed beyond that. But the search box changed the stakes. This was no longer about cloud storage or backup. If the corpus could live in Dropbox, then every citizen investigator with a Dropbox account would have their own periscope into the Epstein files. No dependency on me. No PowerShell. No NER scripts. Just a search box. That made the fight worth having. I kept asking in different ways until I isolated the real constraint: it was not that Dropbox's cloud could not store more than 300,000 files. It was that the Dropbox desktop app syncing to a local machine could not track that many. The primary bottleneck was the sync relationship, not the storage. Dropbox in theory could handle my 1.8 million file universe, but it does not unzip folders. I would have to create the lifts myself, batches that could be uploaded on my home fiber optic connection, which is screaming fast for residential Wi-Fi. I knew enough to run a Cat 6 hardwire from my router to my machine rather than trust wireless for transfers at that scale. The secondary bottleneck was the design of lifts, finding the maximum size that uploaded successfully from my hardware.

I learned what Bates numbers are, the legal profession's sequential stamping system that marks every page of discovery documents. The EFTA files carry them, which made them sortable and searchable. The named court case documents do not. I learned that the hard way. I had unzipped court case files into the flat universe myself, not understanding they had no Bates numbers. Once mixed into 1.8 million sequentially numbered files, they were lost. I had to undo my own mistake, identify the orphaned documents, research which cases they belonged to, and manually re-isolate them into subfolders. Hours of grooming work to fix something that took minutes to break. Hours of PowerShell audits. Those court cases had come from a citizen archive, already organized by case name. That structure had to be restored and then protected going forward.

Windows generates duplicate files silently during copy operations, appending parentheses or "Copy" to filenames. In a universe of 1.8 million files, these multiply invisibly. I used Everything to hunt them down, searching for the telltale suffixes and cleaning them out.

When I choked the Windows 11 machine by not understanding Selective Sync, I realized the Windows 10 workhorse had to come back into the equation. I had done this before: working with construction drawings, I ran two machines because resource-heavy programs needed to be separated yet stay available simultaneously. Same scenario, different mission. I had a lifetime license for Synergy, which allows sharing one keyboard and one mouse across two networked machines. It had been unused for over two years, so I reactivated it to stop the dueling mice and juggling keyboards. Once the workhorse machine was back up and running, I could see that my main worktop was hopelessly choked by Dropbox. The desktop Dropbox app was the problem. I uninstalled the app on that machine and kept it only on the workhorse, where I still needed it to handle the big lifts into Dropbox.

But working across Windows 10 and Windows 11 was a distraction. The operating system mismatch created unacceptable friction. I had Claude outline the steps to upgrading the older machine, necessitating the creation of a Windows Media boot USB to install Windows 11 onto the workhorse laptop. I learned the workarounds for making that happen on hardware that was never meant to run it. The Windows 10 machine became a Windows 11 machine, and my two-laptop station transitioned seamlessly for the work at hand.

The displays came from everywhere. A TV I had been using intermittently as a kitchen monitor. A discounted monitor I had bought new for my travel trailer. A free monitor someone left with “free” taped to it at the door of the union hall when I was signing Book 2 of the out-of-work books. I consolidated all of them with an assortment of docks and cables I already had. Three screens, none acquired for this project, all repurposed. Two broken laptops, one with a busted hinge, one with a damaged display, neither functional as portable machines, both functional as workstation components. Remarkable results from unremarkable materials.

I learned to do the big lifts on the workhorse running the Dropbox app, then aggressively hide each folder through Selective Sync as soon as it was uploaded, keeping the machine from choking on its own index. When the lifts were done, I removed the Dropbox app entirely and went browser-only. The cloud holds the files. My machines stay unburdened.

Pivotal to what became attainable was the realization that AI conversation threads have an effective service life, beyond which the behavior described as AI hallucinations occur. I needed a rough metric to monitor that progression. Across the 100 plus threads I managed I established that 100 kilobytes per markdown (.md) download was the limit. The dynamic also changed when I realized the downloads themselves became forensic documentation of the projects I was running concurrently. I developed a handoff system: export the thread, onboard a new one with the full history, and continue. Each thread serves its purpose and passes the torch.

Arguably the most important thing I learned was to set rules of engagement with AI. No speculating before I invite discussion. No presenting menus of options that lead my thinking before I have finished building my case. No sycophantic agreement when I need honest pushback. These are not preferences. They are quality control. I suspect that is the niche the “agent” builders are selling: customization that comes at a premium I cannot afford.

The conclusion of the matter is that the source files are sacred. Accessibility is the war. The Epstein files are the battlefront. Pixels are the weapons available to engaged citizens. Perhaps my work product, this database I call Alice’s Cause, can contribute to the curbing of unqualified power unjustly exploiting the vulnerable.

 -Rainey