Articles·News & press·15 Sept 2026·6 min read

Verificate launches Compile-Time Inference: facts first, then the decision

On Tuesday 15 September 2026, at AI Infra Summit in Santa Clara, California, Verificate launched Compile-Time Inference. The official announcement was published by the Summit as a press release. This article is the plain-English companion: what launched, why it is built the other way round from RAG, and the evidence that was presented alongside it.

What launched

Most enterprise AI of the last three years has the same shape. A chatbot or an agent searches the company's own files, stuffs the passages it finds into a prompt, and someone else's model produces the answer. The files are used as context. The know-how is absorbed by the model, and buyers are told that a better model means a better answer.

Compile-Time Inference is the opposite stack, and it runs in a fixed order:

Compile a public base layer for the domain. Statutes and judgments, a live news feed, open biomedical records — whatever matches the task. That layer is shared.
The customer's own files write on top. Matters, policies, holdings, internal notes. The combined store is what the institution owns.
The model reads the question and phrases the answer. It does not choose the facts, and it is not the asset.

At query time the behaviour is equally fixed. If a stored fact covers the question, the system returns that fact — no chatbot required. If the question is a decision, software selects the highest-scoring action the compiled records support: identical files, identical action. If the store holds nothing, the system is silent.

“The public layer gets you into the domain. The customer's data is what no one else can rent. Change the model and the cited answer should sit still. Add their files and the system should get more specific to their business. Ask something that is not in the store. You should get silence, not a story.”— Craig Atkinson, founder and CEO, Verificate

Why it is not RAG

Retrieval-augmented generation works the other way. Software finds passages that look similar to the question, and a model then guesses the next word from its own vocabulary. The customer's files become prompt filler; the intelligence stays with whoever trains the next model. Compile-Time Inference moves the decision off the model's vocabulary and onto the customer's data. The full argument, layer by layer, is on the Compile-Time Inference page.

Three deployments on one rule

Legal — Kevin. Starts from the public Australian legal record; a firm adds its own matters and advice. It will not cite a case that is not in the held corpus and flags when a precedent is no longer good law. A research preview, not legal advice. Open the workbench.
Video — the signed-broadcast system. Starts from the News24 live broadcast. Each frame is stamped and indexed as it airs; a newsroom or rights holder layers its own archive and questions. Answers arrive in plain English with a link to the clip, so viewers check the claim by watching. How it works · live system.
Genomics — shown privately. Starts from public biomedical sources; a lab or company adds its own assays and notes under nondisclosure. The rule is the same: cite what the store holds, decline when it does not. Genomics.

The evidence presented with it

The week after the Summit, Craig Atkinson presents an oral paper, “Intervention Bias as Disparate Impact”, at Brown University's ML Symposium on Justice in Healthcare, Education, and Public Policy. It measures what happens when a chatbot, rather than a compiled store, decides which students to flag for help.

On a public Open University data set, a commercial RAG-style system falsely flagged disabled students 63.3% of the time and non-disabled students 32.8% of the time — an adverse-impact ratio of 0.55 against the common 0.80 four-fifths threshold. The gap held when the flags were scored against who actually passed the course, and telling the model to be more careful simply flipped the error: it then missed students who needed help. Policies compiled from labelled outcomes cut the gap and gave the same student the same advice every time.

The paper and the product are one argument: the files decide; the model does not. The underlying research is the EAV-DT paper on arXiv; the decision layer is described on the architecture section.

The long game: a public fact commons

Atkinson framed the destination as a public fact commons — closer to Wikipedia than to a foundation model. People and institutions contribute and maintain cited facts across history, law, culture, politics and science; Compile-Time Inference serves them; open-weight models only read the question and write the sentence. That is how a small onshore model can beat a frontier chatbot on recall: the large model is guessing from weights trained last year, while the commons is the current, cited record. Any country could start with its own layer.

“Wikipedia already proved strangers can maintain a record of the world. We want that record to be compiled, cited and refuse when it is empty, then spoken by whatever open model you have on the desk. You do not out-train OpenAI. You out-remember it.”— Craig Atkinson

The launch press release — AI Infra Summit, 15 September 2026.
“The most expensive place to do your thinking is at query time” — the infrastructure argument, by Craig Atkinson; our companion article.
The Compile-Time Inference brochure (PDF) — the four-page product briefing.
Verificate at AI Infra Summit — what we are showing in Santa Clara.

FAQ

What is Compile-Time Inference?

Compile-Time Inference is Verificate's alternative to RAG. Instead of retrieving passages into a prompt and letting a language model guess the answer, the facts are compiled first — a public base layer for the domain, with the customer's own records written on top — into a cited store the institution owns. At query time the system returns the stored fact, selects the highest-scoring action the records support, or stays silent when the store holds nothing. The model only reads the question and phrases the answer.

How is Compile-Time Inference different from RAG?

RAG finds passages that look similar to the question and a model then guesses the next word from its own vocabulary; the files are prompt filler and the intelligence stays with whoever trains the next model. Compile-Time Inference compiles the facts before any question is asked, so identical files produce identical answers, every answer carries its source, and swapping the model does not move the cited answer.

What was shown at the launch?

Three deployments of the same stack: Kevin, a legal research preview built on the public Australian legal record; a signed-video system built on the News24 live broadcast with every frame stamped and indexed as it airs; and a genomics demonstration, shown privately, built on public biomedical sources with a lab's own assays layered on under nondisclosure.

Can Compile-Time Inference run on our own infrastructure?

Yes. The software runs on the customer's own machines, so production answers need not leave the building. The compiled store — public layer plus the customer's records — stays inside the customer's estate.

Where is the official launch press release?

The launch press release, "Verificate launches Compile-Time Inference: facts first, then the decision", was published by AI Infra Summit on 15 September 2026 and is linked from this article.

Own the answer. Compile the record.

Verificate Compile-Time Inference is the alternative to RAG: a public fact base compiled for your domain, your records written on top, and every answer resolving to its source — on infrastructure you control.

See Compile-Time Inference Download the brochure (PDF)Talk to us