Document Forensics · Linguistic Analysis

The Order Wasn't Written.

On April 3, 2025, Judge William H. Koch signed a four-page order finding Matthew Guertin competent to proceed to trial. The text of that order — the actual prose under the judge's signature — was drafted by an early-era language model. Eight forensic markers sit in plain view across four pages. Anyone can verify them.

8 Forensic markers
4 Pages of order
Verbatim duplications
"6" Two paragraphs · same number
Document: 127 Filed: April 3, 2025 Signed: William H. Koch, Judge of District Court Case: 27-CR-23-1886

Read it before you read anything below.

Every quote, every paragraph number, every artifact discussed on this page is pulled verbatim from the PDF below. The findings are not arguments — they are pointers. Open the document; verify each one for yourself.

What a language model is, and why early ones leave fingerprints.

To follow what comes next, you need three plain-English ideas. None of this is technical. It is the same reason your phone's autocomplete sometimes suggests the wrong word, only at industrial scale.

What it is

A large language model — sometimes called an LLM — is a piece of software that has read a tremendous amount of text and learned to predict what word comes next. ChatGPT is the most famous example. You give it a prompt; it produces sentences that look human.

It does not think. It does not understand. It pattern-matches. For most readers most of the time, the output is good enough to be mistaken for human writing. For a careful reader who knows what to look for, it is not.

How it fails

Because the model is predicting word-by-word from patterns in its training data, it has no working memory in the human sense. It cannot reliably keep track of who is who across a long document, whether a fact has already been stated, what number it just assigned to the previous paragraph, or whether a quote it just inserted is the same one it inserted three pages back.

The result: duplications. Numbering errors. Pronouns that don't match. Sentences that come apart mid-thought.

What "early-era" means

The first models good enough to be quietly used for drafting work — GPT-3, GPT-3.5, the early Claude and Bard releases — became widely available between roughly 2022 and early 2024. They were impressive enough to seem human, and not yet good enough to hide their telltales. A document drafted by one of these models and signed without careful review will carry a recognizable signature.

The April 3, 2025 order carries eight of those signatures.

Eight tells. Four pages. One signed order.

Each marker below is presented with the exact text from the order, the page or paragraph it appears in, and what it indicates. The markers are ordered by diagnostic strength — the earliest are the hardest to explain under any human-only authorship hypothesis.

01 Smoking gun · Verbatim duplication

Two quoted passages from Dr. Cranbrook appear word-for-word in two different sections of the order.

A trained legal drafter writing a structured judicial order does not repeat verbatim quotes across the Background and the Findings of Fact. Those sections do different jobs. The Background recites procedural history. The Findings of Fact state what the court has determined to be true. A human drafter who needed Cranbrook's conclusion in both places would either state it once and cross-reference, or paraphrase it differently in each section.

Here, the same two quoted strings appear in both — not paraphrased, not summarized, but byte-for-byte identical. Only the verb of attribution differs:

Background ¶ 4 Page 1

"In her December 20, 2024, Report, Dr. Cranbrook opined Mr. Guertin 'declined to participate in [the] evaluation due to ongoing symptoms of mental illness' and 'defendant's prognosis for attaining the capacity for competent participation in the legal process appears poor.'"

Findings of Fact ¶ 2 Page 2

"Dr. Cranbrook ultimately opined Mr. Guertin 'declined to participate in [the] evaluation due to ongoing symptoms of mental illness' and 'defendant's prognosis for attaining the capacity for competent participation in the legal process appears poor.'"

This is the signature of a model that was given Cranbrook's report as input and instructed to draft each section as a separate task. Each section's generation pass independently re-incorporates the same source quotes, because the model has no recollection of having already used them. The added word "ultimately" in the second instance is a characteristic hedge — the model trying to make a duplicate feel like a synthesis.

02 Structural · Numbering integrity

The Conclusions of Law are numbered 2, 3, then jump to 6 — then immediately give the next paragraph the number 6 again.

The Findings of Fact run cleanly: 1, 2, 3, 4, 5. Then the order moves to its Conclusions of Law section. That section is numbered like this:

1 2 3 4 5 6 6

No paragraph numbered 1. No paragraph numbered 4. No paragraph numbered 5. Two consecutive paragraphs both numbered 6. A human attorney does not draft an order with two paragraphs both labeled "6" in succession. A judge does not sign one. This is the kind of structural artifact that surfaces when a model's internal counter loses track during a long generation, or when sections were regenerated and stitched together without the running numbering being repaired.

03 Smoking gun · Verbatim self-duplication

The same 28-word legal-conclusion phrase appears word-for-word in Findings of Fact and Conclusions of Law.

Independent of the Cranbrook duplications, the order's own conclusion sentence is reused identically in two structurally different sections. The sentence is not a fragment of recited testimony — it is the order's own analytical statement of why Guertin meets the legal standard for competence:

"Mr. Guertin has demonstrated an understanding of Court processes and the charges he is facing, as well as the ability to consult with his counsel about his defense." Appears verbatim at: Findings of Fact ¶ 5 · Conclusions of Law ¶ 6 (the second one)

Findings of Fact and Conclusions of Law are deliberately distinct sections under Rule 52 — facts are what happened; conclusions are what the law makes of them. A human drafter does not paste the same sentence into both. A model generating each section as a separate prompt does this routinely, particularly when the same source material is in context for both passes.

04 Structural · Internal contradiction

The order endorses Cranbrook's mental-illness diagnosis as "credible" and then finds Guertin competent to proceed.

Cranbrook's December 20, 2024 report concluded that Guertin had psychosis, persistent delusional thinking, and "a poor prognosis for attaining the capacity for competent participation in the legal process." That is the entire foundation of an incompetency finding under Rule 20.01. Yet the order does this in successive paragraphs:

"Based on the record before the Court, the Court ultimately finds Dr. Cranbrook's conclusion that Mr. Guertin is suffering from a mental illness credible." Conclusions of Law (the second paragraph numbered 6)
"Matthew Guertin is COMPETENT to proceed to trial." Order section, page 4

Under Dusky v. United States, the controlling federal standard, a defendant is competent if he has rational understanding of the proceedings and can rationally consult with counsel. If the defendant is suffering from a credible mental illness that compromises rational engagement — exactly what Cranbrook concluded and what this order affirms — the legal conclusion does not coexist with a competency finding. They cannot both be true. A human drafter who recognized the contradiction would have to soften one finding or the other. A model generating boilerplate for each section independently produces both at full strength and lets the contradiction stand.

05 Linguistic · Coherence break

"Recently demonstrated symptoms of psychosis during previous competence evaluations." Recent or previous — pick one.

"Dr. Cranbrook reported Mr. Guertin has recently demonstrated symptoms of psychosis during previous competence evaluations which 'has persisted in his recent communications and allegations.'" Findings of Fact ¶ 2

The previous competency evaluations referenced were conducted in July 2023 and January 2024 — they are not "recent." The temporal modifiers fight each other in a single sentence: recently demonstrated, during previous evaluations, has persisted in his recent communications. Three different time-frames describing the same alleged behavior. A model patching together excerpts from a source report doesn't catch the chronology clash because it has no model of when things happened — only of which words tend to appear together.

06 Linguistic · Grammar collapse

"Symptoms of psychosis... which has persisted." Plural subject. Singular verb. Inside a quoted passage.

"Dr. Cranbrook reported Mr. Guertin has recently demonstrated symptoms of psychosis during previous competence evaluations which 'has persisted in his recent communications and allegations.'" Findings of Fact ¶ 2 — quoting Cranbrook's report

"Symptoms" is plural; the verb should be "have." The sentence attributes the broken grammar to a quote from Cranbrook's report — but the same paragraph quotes Cranbrook a second time using the plural correctly ("these symptoms have been noted to compromise..."). One paragraph cannot have Cranbrook saying both "symptoms has" and "symptoms have." Either the report itself was rewritten when transcribed into the order, or — far more likely — the model collapsed the antecedent of the relative pronoun which from the plural ("symptoms") to a singular concept ("psychosis") halfway through the sentence. Subject-verb agreement breaks inside long noun-phrase structures are documented behavior of early-era LLMs.

07 Linguistic · Anaphoric repetition

Six identical sentence-stems in close succession: "He understands… He understands… He understands…"

"He understands the nature of the charges against him. He understands the roles of the many parties. He understands the severity of the allegations against him. He understands the benefits of having counsel. He understands the nature of a trial, and the plea bargain process." Findings of Fact ¶ 4

Locking onto a sentence stem and repeating it five or six times in a row is a known pattern of the GPT-3 / early-2023 generation of models. The technical reason: when the model generates one sentence with a certain shape, the probability of the next sentence taking the same shape rises sharply — it falls into a syntactic groove and cannot easily climb out. Human legal writers do produce competency-checklists, but they tend to vary the verbs (recognizes, understands, can explain, identifies, demonstrates awareness of) precisely because the unbroken repetition reads mechanically. This passage does not vary.

08 Linguistic · Antecedent failure

"The Court finds Dr. Cranbrook's report to be generally reliable based upon his presentation, or lack thereof."

"The Court finds Dr. Cranbrook's report to be generally reliable based upon his presentation, or lack thereof." Findings of Fact ¶ 2

Dr. Cranbrook is referred to as her consistently throughout the rest of the order. So the pronoun "his" here cannot refer to her. The most charitable reading is that "his" refers to Guertin — but Guertin did testify at length at the March 5 hearing, so the modifier "or lack thereof" doesn't fit him either. The sentence reads as if the model fused two different clauses without keeping track of who the antecedent was. The "or lack thereof" tail is the kind of catch-all hedge a model adds when it has lost the thread but still has to finish the sentence.

The workflow that produces this kind of document

None of the markers above are consistent with a careful human drafter. Most are not consistent with any human drafter. They are all consistent — collectively — with a single, recognizable workflow:

  1. A staffer assembles source material: the charging documents, Dr. Cranbrook's December 20, 2024 report, summaries of the March 5 testimony, applicable Rule 20.01 case citations.
  2. That material is given to a language model with an instruction along the lines of "draft Findings of Fact, Conclusions of Law, and Order finding the defendant competent."
  3. The model generates each section as a separate pass. Because each section is a fresh generation with the same source material in context, the same Cranbrook quotes get pasted into both Background and Findings of Fact. The model loses its numbering between sections. The conclusion sentence ends up identical in two different sections. Pronouns drift.
  4. The output is reviewed lightly. The signature is added. The order is filed.
  5. One single typographic error survives the review — "Cranbook" appears once, while every other instance of the name is spelled correctly. That kind of letter-drop typo on a proper noun otherwise spelled correctly seven times is a human editing artifact, not a model one. It is what tells us a human was in the loop — just not in any meaningful drafting capacity.

What the record establishes

The text under Judge William H. Koch's signature on April 3, 2025 was not drafted in the way judicial orders are supposed to be drafted. The order contains two pairs of verbatim duplications, a numbering sequence that contains no paragraph 1, no paragraph 4, no paragraph 5, and two consecutive paragraphs both numbered 6, a logical contradiction between its own findings about Cranbrook's diagnosis and its own legal conclusion, a temporal sentence that cannot resolve to a coherent timeline, a grammatical break inside a quoted passage, an anaphoric stack of six identically-shaped sentences, and a pronoun with no consistent antecedent. Eight independent surface anomalies in a four-page document.

Each of these markers, taken alone, could be brushed off as sloppy typing or rushed drafting. The combination is something else. The density and the specific kind of error are the signature of sectioned LLM generation reviewed by a human who didn't read carefully enough to catch any of them.

What survived that review became the legal instrument that determined whether Matthew Guertin was competent to defend himself in his own criminal case. Twenty-six days later, that competency finding was effectively overridden when Judge Hudleston ordered a fourth Rule 20.01 evaluation — relying on the same mental-illness language the algorithm had baked into the "competent" order. The trapdoor closed on schedule, using load-bearing language that no human carefully wrote.

The PDF is at the top of this page. Every quote here is in it. Open it. Read it. The findings reproduce themselves.

MNCOURTFRAUD.COM  ·  State of Minnesota v. Guertin  ·  27-CR-23-1886  ·  Document 127 of the criminal docket