Ghost Crusade X-Treme is a serialized paranormal adventure newsletter in which three investigators — Jupiter Meteor Nguyen, a college-credit-collecting physics undergrad who narrates the reports, Joe, an HVAC technician who runs a ghost-hunting website on the side, and Cletus, a dog of uncertain ownership — visit real American historic places and find things there that the National Park Service does not mention.
Everything paranormal in these reports is fiction. Everything else is an attempt at real history. And the whole thing is generated by artificial intelligence — then edited, critiqued, and steered by a human being who reads every single episode. Here is exactly how it works.
The places are real. All 100,866 of them.
Every episode is set at a genuine listing from the National Register of Historic Places, the United States' official inventory of historic sites worth keeping. The server holds the Register's public data download — 100,866 properties at the time this page was written — courthouses and cemeteries, mansions and jails, bridges, churches, and the occasional lighthouse.
Each day the generator picks a site it has not used yet, then sends a research crawler out to read about it: National Park Service pages, state historical societies, museum sites, archives. What comes back is the true part of the report — the dates, the architecture, the fires, the famous trials. The haunting gets invented on top of that, which is the fun part, and the footer of every episode exists because we never want the two confused.
The books Joe half-remembers
The paranormal vocabulary comes from a very specific source: the Time-Life Mysteries of the Unknown series — those ominously illustrated hardcovers that lived on library shelves (and in late-night TV ads) from 1987 to 1992. Thirty-three volumes, roughly 150 pages each, with titles like Phantom Encounters, Alien Encounters, and Cosmic Duality.
In the fiction, Joe grew up reading exactly these books, and it shows: he rambles about them, pulls a battered volume out of the van like a tome of knowledge, and gets the details reassuringly wrong. In the machinery, the series exists as a large original reference map — one summary per volume covering its topics, concepts, and story ideas, deliberately not a transcription, so nothing copyrighted gets reproduced. Its job is to keep the hauntings varied (not every ghost says boo the same way) and to give each character a distinct way of being wrong about the supernatural: Joe with fond inaccuracy, Jupiter with a hypothesis she intends to test, and Cletus by barking at it.
How an episode gets made
Once a day, on a timer, a pipeline on this little server runs end to end:
1. Pick a place. A property is reserved from the Register database — no repeats.
2. Research it. A web-search-equipped model finds and reads authoritative sources about the real site.
3. Design the haunting. A planning model drafts a "blueprint" — what the phenomenon is, how it manifests, what the sensors record — and a second model reviews the blueprint before anything gets written.
4. Write it. A long-form model writes Jupiter's field report, in her voice, from her notes, at 1,500–2,500 words. A review pass checks the draft against the site's accumulated rules.
5. Photograph it. Two images are generated (more on those below).
6. Publish. The episode is rendered to a static page and served. No framework, no database on the public side — just files behind a web server.
Different open models play different roles — one plans, one writes, one handles the small jobs — coordinated by a few hundred lines of plain Python. The one thing no model gets to do unsupervised is decide what is true about the real places.
The photographs
Each episode carries two AI-generated images styled like black-and-white field photos: the first shows the real historic location with the team's van parked outside; the second captures a moment from that episode's investigation. A standing image specification guards them — the characters must look like themselves every single day, the cast in a scene must match who was actually present, the historical details of the building come from the research, and the images are never allowed to invent signage, logos, or text. Old buildings are hard to draw and easy to fake; the spec exists so the pictures feel like evidence, not posters.
The human in the loop
And now the most important part. After each episode is published, a human being — the author of this project — sits down and reads it. Then they open a private feedback page and say what worked and what didn't, in plain language, the way any editor would: "Joe wouldn't say it like that." "Jupiter's field notes read like a data log; give them her voice." "Nobody knows whose dog Cletus is. Keep it that way."
That feedback is classified by a model into three streams:
Writing guidance — direction for future episodes: how dialogue should sound, how pacing should feel.
Character memory — what Joe, Jupiter, and Cletus individually learn about how they are supposed to behave.
Shared lore — what actually happened in an episode, so next week's story remembers it.
The classifier is kept on a short leash. It must quote the exact feedback that supports each point, and any claim about events must match a quote from the actual episode text. A protected canon list overrides everything: Joe is a competent HVAC technician and not a fool; Jupiter is the narrator with a real physics background; Cletus is an ordinary dog with no powers, not a mascot or a punchline. Feedback that conflicts with canon is rejected automatically, and unsafe merges are refused — the pipeline supports additions, not rewrites of established truth.
All of it is folded back into the next day's generation context. That is the actual experiment here: the characters are not fixed. Joe's voice sharpens, Jupiter's notes get more like field notes and less like spreadsheets, and the world accumulates continuity — because a person read the last story, told the machine where it rang false, and the machine kept the note.
The short version
Real places from a real government database. A paranormal imagination borrowed from thirty-three out-of-print books about ghosts and Bigfoot. Models that plan, write, review, and paint. A human editor who reads everything and holds the canon. New episode daily.
The hauntings are made up. The process isn't.