The GPUs arrived on a Wednesday afternoon.
Three top-end RTX 8090s, couriered in a plain box to avoid questions from the neighbours. Craig signed for the parcel in silence, nodded once at the driver, and closed the door without saying a word. He carried the box to the spare room, moving with the tension of a man preparing for surgery.
The purchase had hit his credit card like a hammer—just over £3,600—but he told himself it didn’t matter. If things went wrong, the debt wouldn’t matter anyway.
He used three machines.
The first was a burner: a refurbished Dell workstation he’d bought second-hand with cash two weeks prior, wiped clean and reloaded with a minimal Linux distro. It had no connection to any of his personal accounts, devices, or patterns. He’d installed a rotating VPN chain, chained across three different providers, with exit nodes in countries that didn’t honour UK data requests. The MAC address was spoofed. The logs were off. The box was configured to die clean.
The second machine—the one in the next room—was different.
That one was air-gapped. No network card, no wireless chip, no modem. He’d stripped out the mic and camera, too, just to be sure. It would never touch the internet. It sat there humming quietly, waiting for data. Waiting for truth.
The USB key Jeremy gifted him wasn’t just a data stick—it was the whole system. The scripts weren’t copied off it—they were executed directly from the device itself. And the key contained more than code. It housed a firmware-based cryptographic handshake module—something closer to a security fob than a drive. Once slotted in, the host machine bootstrapped the tunnel and performed a low-level handshake against a buried whitelist deep inside the civil service’s infrastructure.
Craig understood instantly: this wasn’t some credential dump. This was a physical identity.
As long as it stayed plugged in, the machine became—just briefly—someone with clearance.
The USB key initiated access by spoofing an internal call via a deprecated SIP gateway—an old internal comms route that should have been decommissioned years ago. It tricked the system into thinking a request was coming from inside the civil service’s telecom core.
Then it rode a buried legacy tunnel—barely documented, practically invisible. A tiny sliver of exposure buried inside a forgotten node in the infrastructure.
The connection blinked green. He was in.
There was no interface. Just command-line prompts and access to a data lake that spanned the whole of the United Kingdom. One simple text file on the USB explained the database schema, so Craig knew what tables and columns to navigate to via SQL commands.
He scoped his query tightly.
Hemel Hempstead postcodes only.
Movement data was first: hourly GPS coordinates for each smartphone. Twenty-four data points per person per day. Roughly 100,000 residents, but most children and the elderly could be excluded. Adults only. He estimated 60,000 phones.
He chunked the requests. Random delays between each segment. Output files disguised with names like Q4_TransitLogs_Archive.csv and Legacy_Mapping_Review.csv.
It took just under seven hours to complete the download.
1.7 terabytes in total.
He moved the data to an encrypted drive, and triggered the kill script, watching the screen flash once as the encryption keys self-deleted and the drive zeroed itself into silence—nothing left now but noise and melted circuits.
It would never boot again.
He walked the drive into the next room and plugged it into the air-gapped box.
The machine spun up silently. No fans screaming. Just the quiet pulse of electric focus. The room felt stiller with that machine running—as if even the dust was holding its breath.
The LLM was his own build. A local fork of LLaMA 14.1, fine-tuned over months for NLP and temporal pattern recognition. He’d originally trained it to spot fraud in time-logged maintenance records. Now it would be doing something darker.
He fed the full dataset into the model—names, addresses, marital links, precise movement logs. No abstractions, no redactions. The MESSEJ logs decrypted beside them, blunt and raw, mapping one life to another with terrifying clarity. Reading them felt invasive, like he was pressing his ear to a stranger’s confession booth. But the AI didn’t flinch. It parsed the patterns without hesitation—who went where, who said what, who shouldn’t have been anywhere near whom.
The prompt was simple, but devastating:
“Identify anomalous movement patterns inconsistent with stated home location. Cross-reference against message content indicating secrecy, romantic intent, or covert coordination. Sort by confidence. Output list of user tokens with relevant notes.”
The model ran for ten hours.
He didn’t sleep. Just sat in the room, drinking lukewarm coffee and watching terminal output flicker. At one point, he caught himself whispering names aloud—U031, U074, U119—like they meant something.
He didn’t want to open the final CSV. But of course, he did.
147 targets.
All cross-referenced. All mapped. Many paired. Several confirmed with extremely high confidence scores—92%, 95%, 99.7%. According to MESSEJ chat logs: married people. Teachers. A nurse. A detective constable. One MP's aide. All in Hemel Hempstead.
The files showed where they’d gone, with whom, and when. The AI had even extracted snippets of MESSEJ chat—pet names, location planning, late-night emojis.
He sat with the data open in front of him.
The mailer script was half-done.
He had ten Bitcoin wallets generated, randomized assignment.
A VPN-routed freemail setup.
He'd built a header-stripping SMTP relay chain.
All deployed on his third machine that would run the mailshot and be nerfed soon after.
It would all work.
He hadn’t sent anything yet.
But he looked down at the draft window for Target #001.
The subject line was already filled in:
“I saw you with Vanessa Peters at Premier Inn on 7th February. Please open this email to save your marriage.”
Craig stared at it for a long time.
Then he closed the lid.
Night had fallen. He didn’t turn on the light.