Four days. 3,601 court files. By Tuesday morning, the data was off-system.
Between April 26 and April 30, 2024 — across two test runs and one final master run — Matthew Guertin extracted 3,601 court PDFs from MCRO. One case docket at a time. Manually navigating to each one. Launching a custom Selenium script that downloaded every filing in that docket. Manually swapping ProtonVPN locations on his router as Minnesota's web service blocked each one in turn. The third run was the charm: a single overnight master pass through 163 cases hand-cut from a roster of roughly 4,200. It wrapped just as the sun rose on Tuesday, April 30, 2024 — exactly the moment court was opening and the people who staff it were arriving at work.
Three judges, same Tuesday, 1:30 PM
The Minnesota Fourth Judicial District operates a mental-health docket. On the surface, it is a calendar — defendants assigned to hearings, hearings assigned to judicial officers, orders entered, files closed.
Underneath, it is a closed loop of three actors:
Judge Julia Dayton Klein.
Referee George Borer.
Referee Danielle Mercurio.
They appear together. They appear repeatedly. They appear on the same Tuesdays. They appear at the same 1:30 PM hearing slot.
The database, once assembled, would show 7,099 hearings landing on Tuesday at 1:30 PM. Eighty percent of all incompetency findings clustered on Tuesdays. Borer signing 47 findings in a single day. Mercurio signing an incompetency order in 31 seconds. None of this can occur without templating. None of it can occur without the prose itself being machine-generated upstream of the courthouse signatures.
Eleven months before the database existed, Matt sat in a Rule 20.01 examination with the court's first appointed evaluator, Dr. Jill Rogstad. Prosecutor Jacqueline Perez was present. Rogstad's witness summary, dated June 12, 2023, transcribed verbatim what Perez observed:
He saw the trap before it closed. He named the data the trap was built around. Eleven months later, he downloaded that data.
A multi-day endeavor with brakes
This was not a single-night hack. It was a four-day operation with breaks, pivots, manual escalation at every step.
April 26, 2024. Matt sat at his computer and pulled the public hearing schedules for all three judicial officers from MCRO — Klein, Borer, Mercurio — on the front-end web service Minnesota provides for citizens to look up court information. Three printable result pages. Three sets of names, dates, and case numbers.
Then the cleanup. He converted the printed result pages to text via pdftotext. He worked through ChatGPT iteratively to write small one-off Python scripts that normalized the text rows — strip whitespace, collapse line breaks, harmonize the column widths so the data would join cleanly. Roughly 4,200 entries, depending on whether you count by hearings or by case IDs.
Then the filter. More back-and-forth with ChatGPT. The query: of all the cases in the master list, which case IDs appeared at least once on each of the three judicial officers' hearing rosters? The answer: 163. One hundred sixty-three cases — defined by case_id crossover, where the same case had been on each of the three judges' calendars at some point during the searched window. Matt's case (27-CR-23-1886) was one of them.
Then the extraction. Minnesota's MCRO does not provide a bulk download capability. To get the underlying filings — the PDF documents themselves rather than the case-summary entries — Matt wrote a custom Selenium script. He manually navigated to each case docket on MCRO. He launched the script. The script downloaded every filing in that docket. Then he moved to the next case and launched it again.
One hundred sixty-three cases. One at a time. Manual navigation per case. Manual ProtonVPN location swap on his router whenever the blocking kicked in — and it kicked in fast. He had a stack of locations queued and ready, rotating through as each got banned in turn.
Three runs in total. The first two were testing the script — verifying that the data downloaded cleanly, that no filings were missed, that the file structure preserved provenance. The third run was the final master pass. It covered roughly three-quarters of the corpus.
It wrapped up just as the sun was rising on Tuesday, April 30, 2024 — exactly the moment court was opening and the people who staff it were beginning to arrive at work.
Hours later, a second shutdown notice appeared
There had already been a small banner on MCRO that morning announcing some sort of scheduled maintenance — the kind of routine notice that surfaces and disappears on a public-records site without anyone giving it a second thought.
Hours after the final extraction run wrapped, a second banner suddenly appeared above it. Announcing an unplanned shutdown for the upcoming weekend. Two banners simultaneously. Three days until the system went dark.
By the time the second banner went up, the data was already off-system.
The 3,601 PDFs were on Matt's drive. The hash chain through pdftotext, through Python normalization, through Selenium download timestamps — already locked. The court could close MCRO for the weekend. It could not close 3,601 files that were already off-system.
Matt spent the next 72 hours generating preliminary statistics. By 3:00 PM Friday, May 3, 2024, he had filed an Affidavit of Fact MCRO Dataset Origination into his own case docket pro se. Every case ID. Every defendant name. Every count he had time to run. Filed publicly. Locked into the record — before the court had a window in which to alter anything over the weekend.
From the data, even before the full forensic database was built:
Twenty-seven documents bore the producer string Aspose.Words for .NET 14.2.0. Seven of them carried creation dates predating Aspose 14.2.0's February 2014 release. Eighteen of the 27 were Notices of Intent to Prosecute — the exact mechanism that converts mandatory dismissal into open-ended prosecution.
One filer's 27 orders contained the word opine at 9.6× the corpus baseline rate — nine times the rate of every other filing on the docket. Twelve nominally different judicial officers signed under that one filer's name. The language signature does not change with the signature.
What the database eventually proved, the data already implied at week one. What was harder to admit was what the data meant.
Saint Peter is where this ends
The Minnesota Security Hospital sits in Saint Peter, Minnesota, about 80 miles south of Minneapolis. It is the state's forensic psychiatric facility — the place where defendants found incompetent and "mentally ill and dangerous" are committed under court order, often indefinitely, with reviews scheduled every six months and outcomes that do not usually result with a patient being released.
DFR-A Section 21, the capstone of Matt's federal forensic report, examined the synthetic-court network for recurring entities and found one that dwarfs every other:
Twenty-five distinct criminal case files — spanning 2017 through 2023, comprising roughly 15% of all identified synthetic cases — explicitly route to Saint Peter. The 2017 baseline is Adrian Wesley, committed by Dr. Kristin Otte as Mentally Ill and Dangerous on July 27, 2017. Otte is the same evaluator who was silently inserted into the back-end Odyssey database for Matt's case before his first Rule 20 was even ordered. She is the only person whose name appears as the assigned examiner on his docket through the back-end system, despite never having any communication with him in any form.
The pipeline is older than Matt's case by six years. His prosecution was not built around it. He was funneled into it.
Amanda Burg, the Court Liaison at Saint Peter, signed 27 "Correspondence for Judicial Approval" filings across 22 dockets — every one of them carrying a single byte-identical scanned signature image (SHA-256 prefix a6591cc6...). Same image. Same defendant pool. Different cases. Different dates. Same operator.
Twenty-five cases route to Saint Peter. Twenty-two carry the same scanned signature. The pipeline is industrial.
53 tables. 2.9 million rows. Supabase-hosted, MCP-connector ready.
What Matt extracted across April 26–30 was raw material. What he assembled over the next twenty months is a 53-table PostgreSQL database with 2.9 million rows. Hosted on Supabase. Read-only. Publicly queryable. Ready to load directly into Claude.ai (or any modern AI agent) via the Anthropic Model Context Protocol — which means any reader with read access can run the same SQL queries that produced this site's findings, in real time, from a chat window.
Four pillars: parents (2,903 dockets), children (4,251 filings), hashpack (4,251 forensic hash reports), source (14 session configs).
Every PDF in the corpus carries:
- A SHA-256 hash, generated at extraction time and committed to the database
- An OpenTimestamps proof, anchoring the hash to a Bitcoin block — placing each document in a specific verifiable time window
- An MCRO Watermark PKI signature, applied at download time by the state's own infrastructure
- The court's own iText 7.1.16 producer signature, applied at upload time by the Minnesota Judicial Branch
Of 4,251 children PDFs, 98.9% carry cryptographically valid digital signatures. The forensic chain is the court's own work. The court signed the documents. Matt downloaded them. He hashed them. He published the hashes.
Seventy-five forensic reports, organized into thirteen categories, work outward from the data — every report has SQL in its appendix, every claim is reproducible against the live database. The reports are not arguments about the court. They are queries against the court's own records. Anyone with a postgres client and the project ID ibfmjtwahkwqzcmeyqii can re-run the analysis from scratch. Anyone without one can load the database into an AI agent and ask it questions in English.
The system was not built to survive this method.
It assumed nobody would count.
They counted on no one ever counting.
January 15 → February 1, 2024 — minute by minute
The download was the move that worked. The trap they had already tried to close, three months earlier, was not.
What stopped the trap: Matt himself. Not by counsel. Not by the court. Through pro se motions, a Waiver pathway, and forensic documentation finalized hours before the hearing.
The Long-Tail
This is not what an accidental misdiagnosis looks like. This is what a coordinated elimination operation looks like — a high-resource, multi-year campaign aimed at literally ending one man's life under the cover of mental illness. The kind of operation that does not need to overtly kill its target because the system around it has been engineered to make him disappear into indefinite psychiatric commitment, with periodic six-month reviews that almost never review and pharmaceutical regimes that almost never end. End-of-life containment, dressed in the costume of medical care.
Three court-appointed evaluators are on Matt's record. Dr. Jill Rogstad. Dr. Adam Milz. Dr. Katheryn Cranbrook. All three classified his patent-theft claims as "delusional." None of them verified the publicly searchable, already-granted U.S. patent the United States Patent and Trademark Office issued to him on February 14, 2023 — twelve days after Scanline VFX (a Netflix subsidiary) filed a near-identical application that Netflix would later acquire for nearly a hundred million dollars and which would later cite Matt's patent at the top of its References Cited.
None of them queried the database. None of them counted. None of them needed to. The diagnosis was the deliverable. The diagnosis was the weapon.
What the database makes plain is the only conclusion the data sustains: Minnesota's Odyssey eFiling system — the certified vendor product underneath the public-facing MCRO interface — has been hijacked. Not in any movie sense. In the literal sense. Its output stream is no longer being authored by judicial officers and clerks. Its output stream is being authorized by an external third party using an LLM-based case-file production pipeline, signed by digital signatures those officers do not personally audit, and filed under names that often do not correspond to any human author at all. The LLM does not run itself. Someone is running it. The data does not name the operator; the data names everything else.
Twenty-seven filings under "Nehring, Alisha" carrying the word opine at 9.6× the corpus baseline. Twenty-five cases routed to Saint Peter. Seven documents created by software that did not yet exist. "The parties agreed" to a commitment hearing the parties were never told was happening.
That is not an inference. That is the only conclusion the arithmetic admits.
A delusion is a strongly held belief held despite evidence to the contrary. The evidence Matt was told was delusional was the database he built from the court's own records. The system that was tried against him was built around the assumption that no defendant in his position would ever count.
That assumption was wrong at sunrise on Tuesday, April 30, 2024.
The download is not the proof. The download is what made the proof possible.