Retrieval for authoritative legal text that answers only in verbatim, cited spans.

Air-gapped and quote-locked: every answer is a passage that exists in the corpus, with its provenance attached. There is no free-generation path to hallucinate through.

Texas · signed 2026-08-26 · build 7a710b64bbf3  —  Louisiana · signed 2026-08-27 · build 75201b816067

The cheapest route to a zero fabrication rate is silence. Both rates are fixed before the run and published together — the second is the price of the first.

Other states measured better than Texas on fabrication and none of them shipped, because each missed a coverage floor set in advance. The gates are real.

How it is proven

• The rules are locked first: failure definitions, sample size and coverage floor, signed before any result exists.
• The answer key is the law, not a model. Every question is anchored to a real row of enacted text.
• The question-writer is not the test-taker. The examination is drafted by a different company’s model.
• The grader is code. Same input, same grade, every time. Never an AI judging another AI.
• The exam is sealed and run once. Infrastructure failures count against the score.
• The failures are published beside the passes.

The one honest gap is that I commissioned the questions, so replace them. The reproduction kit stands the system up on hardware you control and runs the same examination against any question set, including one never shown to me.

Beyond law

It is not built for Texas. It is built for governed published text. On 2026-09-01 the machine took in the federal health-privacy regulations — 45 CFR, 5,617 provisions — with the two state statutes that sit on top of them and can be stricter: Texas Health and Safety Code chapter 181, 31 provisions, and California’s medical-confidentiality act, 39 provisions, each acquired from the official publisher and fingerprinted first. The same week it loaded the medical-coding standard, ICD-10-CM, 98,186 codes, every one accounted for. The engine was not edited for any of it.

For context: an independent 2024 Stanford study measured leading commercial legal-AI research tools at 17–33% hallucination on comparable questioning. Those figures are two years old and describe products that have since changed.

One reply, direct to me: luke@pennoyersystems.com

Graded by machine against rules fixed in advance; the result is signed however it came out. No laboratory or standards body has graded anything here, and none is claimed. No attorney endorses any signed result.