AI is great, but AI is wrong.
Pennoyer quotes the rule word for word and proves where it came from. I’m looking for the company that should own it.
I spent twelve years building the systems that ran a nine-figure distribution business. Then my wife, a senior litigator, repeatedly did the same tasks that I knew how to solve. She had to prove where every line of law she quoted came from. So I built her a system that answers from a fixed set of laws that has no inference or opinion that does not reach out to the internet.
Your customers answer to fixed rules: statutes, regulations, codes, standards, contracts. They can’t accept an answer they can’t trace. I’d suspect you have meetings about AI being wrong weekly, if not daily.
It doesn’t give opinions; it’s a recall tool. Ask it a question and it answers with the passage word for word, its citation, and the search record. If the text is silent, it says so. Every source is time-stamped and fingerprinted (SHA-256) from acquisition to answer. It runs offline on local models, air-gapped, with zero cloud calls when it answers.
- Answers your customers can check. The source comes attached, so checking it takes a glance, not a research session.
- A chain of custody you can hand an auditor. From the day a source is taken in to the answer it supports.
- Nothing leaves the building. Regulated and privileged material stays inside.
- The R&D is done. One engine has already taken in, with no engine change:
Law for 48 states, plus DC and the U.S. CodeEU law, 32,433 acts, in EnglishICD-10-CM, 98,186 medical codes49 CFR, hazardous-materials transportationMITRE ATT&CK for industrial control systems
Patent PendingU.S. Provisional Patent Application No. 64/143,510 · filed August 28, 2026
One invented citation in a filing, an audit or a compliance answer can mean sanctions, a failed audit, a lost client, a suspended license, or litigation. Courts have already sanctioned lawyers for citations a machine invented.
Email me at Luke@PennoyerSystems.com or call +1 214-789-0990. That’s my personal number; you can verify it. The measured record is below.
Patent pending · USPTO 64/143,510 · filed 2026-08-28
| fabricated | wrongly refused | |
|---|---|---|
| Texas | 7 0.67% | 113 10.76% |
| Louisiana | 11 1.05% | 128 12.19% |
| California | 2 0.19% | 127 12.10% |
| New York | 12 1.14% | 167 15.90% |
| Florida | 5 0.48% | 142 13.52% |
| Illinois | 4 0.38% | 129 12.29% |
| All six | 41 0.65% | 806 12.79% |
How it was examined. Six states, 1,050 sealed questions each. About half ask about a rule that does not exist, to see whether it plays along. Another company’s model wrote the questions from real statute text; a second model screened them blind. Code, not a model, graded the answers, under rules written before the run. Examination of 6–9 September 2026.
Why both numbers print together. A system that refuses everything never fabricates. So refusals count against the score, and both rates are published side by side. It would rather say nothing than guess.
To begin talks, we start with yours. Built in my office, on my own hardware, with my own money, and air-gapped.
- Email me. I’ll send an NDA.
- Send one set of your rules and twenty questions, laid out as shown in the Corpus Protocol.
- I run them on my hardware and send back every answer exactly as it came out, fingerprinted, with its source. You grade it.
Luke@PennoyerSystems.com
Luke.C.Thorburn@gmail.com
Not yours to decide? Hand this to whoever runs AI at your company.