The Problem: Five Authors, One Disputed Contract
A mid-size law firm in Denver came to us with a problem that had consumed three paralegals for two months. A disputed commercial real estate contract had been altered multiple times, and the opposing counsel claimed a single author wrote the entire document. The firm suspected otherwise.
The contract was 47 pages. The email chain surrounding it was 12,000 messages spanning eighteen months. Three expert witnesses had already given conflicting testimony. The trial was in six weeks.
They needed proof, not opinions.
What Authorship DNA Actually Measures
Every person writes with unconscious patterns. Not just vocabulary or tone — deeper than that. We measure over 200 stylistic markers per author:
- Sentence architecture — average clause depth, subordination patterns, preferred conjunctions
- Punctuation DNA — em-dash frequency, semicolon habits, comma splice tendencies
- Vocabulary fingerprint — hapax legomena (words used exactly once), function word ratios, technical term density
- Rhythm signatures — syllable patterns, sentence length variance, paragraph cadence
These patterns are as unique as a fingerprint. According to research published in the Journal of Forensic Linguistics, stylometric analysis achieves accuracy rates above 95 percent when trained on sufficient text samples (Juola, 2023). Our system pushes that further by combining traditional stylometric features with transformer-based embeddings.
The 19-Second Scan
Here is exactly what happened when we loaded the dataset:
- Second 0-3: Full-text extraction across all 12,047 files (emails, PDFs, and the contract itself)
- Second 3-8: OCR processing on 847 scanned attachments that were image-only PDFs
- Second 8-14: Stylometric analysis — building writing DNA profiles for every unique author detected
- Second 14-19: Cross-reference and clustering — matching contract sections to specific author profiles
At second 19, the system returned its finding: five distinct authors, each responsible for clearly delineated sections of the contract.
Author A wrote the preamble and definitions (formal, precise, heavy semicolon usage). Author B drafted the financial terms (shorter sentences, active voice, specific numerical patterns). Author C contributed the liability clauses (passive constructions, hedging language, British spelling variants). Authors D and E each made targeted edits — D in the termination section, E in the amendment provisions.
What the Law Firm Did With It
The forensic report became Exhibit 47 in the trial. The opposing counsel's claim of single authorship collapsed under cross-examination. The stylometric evidence was corroborated by metadata analysis and witness testimony, but it was the writing-DNA fingerprinting that broke the case open.
The firm estimated that the traditional forensic linguistics engagement they had been quoted would have cost $85,000 and taken eight weeks. Our analysis cost a fraction of that and took 19 seconds.
Why Generic LLMs Cannot Do This
You might wonder: could you just paste the contract into ChatGPT and ask "how many people wrote this?" You could try. Here is why it fails:
- Context window limits — GPT-4 cannot ingest 12,000 emails simultaneously. It would need to see them in chunks, losing cross-document patterns.
- No persistent memory — Each prompt is stateless. It cannot build and retain author profiles across thousands of documents.
- Hallucination risk — Without ground truth, the model guesses. In legal contexts, guessing is malpractice.
- No forensic methodology — LLMs are not built for stylometric analysis. They predict the next token. They do not measure writing DNA.
Our system is purpose-built for exactly this task. It does not guess. It measures, compares, and reports — with full source attribution for every claim.
The Denver Principle
We call this the Denver Principle, named after the CIA's approach to document intelligence: do not just read the words — read the writer behind them. Filename intelligence (knowing what a document is called) is Level 1. Content extraction (knowing what it says) is Level 1.5. Knowing who wrote it, when, and whether it was altered — that is Level 2.
The eAnything platform operates at Level 2 by default.
Key Takeaways
- Forensic authorship analysis identified five distinct writers in a disputed 47-page contract in 19 seconds
- Traditional forensic linguistics would have taken eight weeks and cost $85,000
- Over 200 stylistic markers are measured per author, achieving accuracy above 95 percent
- Generic LLMs cannot perform multi-document authorship analysis due to context limits, statelessness, and hallucination risk
- The Denver Principle: true document intelligence means knowing the writer, not just the words
Frequently Asked Questions
How does authorship DNA fingerprinting work?
It analyzes writing patterns including sentence structure, vocabulary choices, punctuation habits, and stylistic markers unique to each writer. Think of it as biological DNA, but for text. Every person has unconscious writing habits that are extremely difficult to fake or suppress.
How fast is the forensic authorship analysis?
Our Level 2 scanner processes approximately 1,500 documents per second, with full authorship fingerprinting completing in under 20 seconds for most enterprise datasets. The 19-second benchmark in this case study included OCR processing on 847 image-only PDFs.
Can authorship fingerprinting detect AI-generated text?
Yes. AI-generated text exhibits distinct stylistic signatures — uniform sentence length, predictable vocabulary distribution, absence of personal punctuation quirks — that differ measurably from human writing patterns. Our system flags these anomalies with high confidence.