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

Verificate launches Compile-Time Inference at Santa Clara's AI Infra Summit

SANTA CLARA, Calif. / SYDNEY, Sept. 15, 2026

Companies have poured money into chatbots and agentic workflows that search their own files. The answers are still produced by someone else's model. The files are used as context. Buyers are told a better model means a better response and less hallucination. The data and the know-how keep being absorbed by the model.

Verificate Pty Ltd, a Sydney software firm, is selling the opposite stack. On Tuesday, at AI Infra Summit in Santa Clara, Calif., it is launching Compile-Time Inference.

The company first compiles a base layer from public sources that match the task. That may be statutes and judgments, a live news feed, or open biomedical records. That layer is shared. The customer's own files then write on top of it: matters, policies, holdings, internal notes. The combined store is what the institution owns. A language model is used to read a question and to put an answer into a response or a tool call. It does not choose the facts, and it is not the asset.

The usual method, retrieval-augmented generation, or RAG, works the other way. Software finds passages that look similar to the question. A model then guesses the next word from its own vocabulary. The files are prompt filler. The intelligence stays with whoever trains the next model.

Verificate compiles first. If a stored fact covers the question, the system returns that fact. No chatbot required. If the question is a decision, software picks the highest-scoring action the compiled records support. Identical files, identical action. If the store has nothing, the system is silent.

"The public layer gets you into the domain. The customer's data is what no one else can rent," said Craig Atkinson, Verificate's founder and chief executive. "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."

The company is showing three versions of that stack.

Kevin, at verificate.ai/kevin/workbench, starts from the public Australian legal record. A firm can then add its own matters and advice. Verificate says the tool will not cite a case that is not in the held corpus, and will flag when a precedent is no longer good law. It is a research preview, the company said, and not legal advice.

A signed-video system, at hybrid-vlm.verificate.ai, starts from the News24 live broadcast. Each frame is stamped and indexed as it airs. A newsroom or rights holder can layer its own archive and questions. Answers come in plain English with a link to the clip. Viewers check the claim by watching.

A genomics demonstration, shown privately, starts from public biomedical sources. A lab or company can add its own assays and notes under nondisclosure. The rule is the same: cite what the store holds, decline when it does not.

Atkinson said the long game is a public fact commons, closer to Wikipedia than to a foundation model. People and institutions would contribute and maintain cited facts: history, law, culture, politics, science. Compile-Time Inference would serve those facts. Open-weight models would only read the question and write the sentence. They would not be the memory.

That, he said, is how a small onshore model could beat a frontier chatbot on recall. The large model is guessing from weights trained last year. The commons is the current, cited record. Australia could start with its own layer. So could any country. Contributors would add facts, not pay for the next training run.

"Wikipedia already proved strangers can maintain a record of the world," Atkinson said. "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."

Atkinson is due in Providence, R.I., the following week. On Sept. 24-26 he will give an oral paper at Brown University's ML Symposium on Justice in Healthcare, Education, and Public Policy. The paper, "Intervention Bias as Disparate Impact," measures what happens when a chatbot, not 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 threshold. The gap held when the team scored the flags against who actually passed the course. Telling the model to be more careful flipped the error: it then missed students who needed help.

Policies compiled from labeled outcomes, the paper says, cut the gap and gave the same student the same advice every time. Atkinson said that result and the Summit product are one argument. The files decide. The model does not.

Verificate software can run on a customer's own machines. Production answers need not leave the building.

"We want our customers' learned knowledge, often compiled over decades, to matter more than the model that makes that intelligence accessible," Atkinson said.

The company was founded in 2024. It also sells a code-review gate for AI-written software and an inference engine, HELIX, that attaches a confidence score to model output. Compile-Time Inference is the layer that turns public records plus customer files into a store the customer keeps.

At the Summit, Verificate is looking for customers and partners to move the conversation off datacenters and tokens-per-watt, and onto whether the institution owns the answer.

"The future of this work is how you mobilize what you already know," Atkinson said. "Compile-Time Inference is how we serve a large cited record on hardware customers already own, without sending the answer back to someone else's model."

Media
Verificate Pty Ltd, Sydney
info@verificate.ai
verificate.ai

About Verificate
Verificate Pty Ltd (ACN 681 762 818) compiles a public fact base in a domain, then lets a customer write their own records on top. Language models extract and narrate. The store holds the facts. Compiled policies pick the action. The firm is based in Sydney.