Research
Portable reference libraries that let a locally hosted AI answer against authoritative sources.
A knowledge library is a curated, downloadable body of authoritative material, prepared so that an AI engine running on your own hardware can query it directly. Nothing leaves the machine it runs on. The Centre assembles the libraries, and the system that reads them, on the subjects where a reliable local reference matters most.
The Centre keeps the forge. You keep the crucible, and everything that goes into it. The forge is where the base system is built and where every ingot is signed; the crucible is the machine in your own office. Once an ingot is in your crucible, nothing about it, and nothing you ask of it, ever returns to the Centre.
One way. Nothing returns to the Centre.
An ingot is a sealed, signed unit of capability. It is verified against the Centre’s own signature before your crucible will accept it, and it carries an expiry, so a copy cannot quietly become the current one. Two kinds do the work described on this page.
AXM is what turns the crucible into a working AI system, and it arrives complete. It carries the model runtime that holds the language model in your graphics card, the chat interface you actually work in, the retrieval layer that finds the passages in a library that bear on your question, a graph store that holds the relationships between documents rather than only their text, and the tailoring that governs how the whole of it answers.
It installs in one step and needs nothing from the internet once it is in place. Every part of it runs on the machine in front of you, including the model.
The ones that are ours alone. Each is a curated corpus on a single subject, prepared and indexed so the engine answers from the source material and cites it. They are added to AXM one at a time, and a crucible may carry as many as its owner has use for.
A library is not a search index bolted on afterwards. It ships with its own instructions for how that subject is to be handled: what counts as authority, what must never be paraphrased, and what the engine should refuse to guess at.
1,024 Acts of Parliament and 4,904 regulations and statutory instruments, the consolidated federal statute book, as it stood on 24 July 2025. Every Act carries the date it was consolidated to, so an answer can be placed in time rather than assumed current. Ask what the law says and the reply is the text of the provision, not a summary of it. The largest of the libraries by a wide margin.
Treaty text, authoritative commentary, customary law, and the published positions of states on how international law applies. Among them are the Charter of the United Nations, the Geneva Conventions and their Additional Protocols, the Vienna Convention on the Law of Treaties, the UN Convention on the Law of the Sea, and the Rome Statute. Sources of very different authority are kept distinct, so a single state’s claim is never presented as the settled view, and a commentary is never mistaken for the instrument it comments on.
The public policy and direction of the Government of Canada, drawn together from the Treasury Board of Canada Secretariat, the published guidance of the Canadian Centre for Cyber Security (the Cyber Centre), and the Royal Canadian Mounted Police. The policy hierarchy is preserved, so an answer can show what sits above a given instrument and what hangs beneath it, which is usually the question behind the question.
Historic public addresses by heads of state, heads of government and other figures of consequence, held as delivered. Built for speechwriters and for researchers. A speechwriter can work against the cadence and structure of the record without lifting a line from it; a researcher can establish what was said, by whom, and when. The register says plainly when it holds nothing by a given speaker. That is the answer a researcher needs, and the one a general assistant is least likely to give.
The Cyber Centre’s security control baselines and the practitioner guidance around them, prepared for design and assurance work rather than for reading. It exists so a practitioner can qualify their own work. Put your design, your assessment, or a supplier’s document to it and ask where it departs from the baseline it was written against, control by control, each finding carrying the control it came from. It holds the relationships between controls as well as their text, so it can also be asked what a given control draws in, and what a given profile actually selects.
From the policy down to the mathematics. At the top: which cryptographic algorithms are approved today, what replaces them, and the dates and conditions attached to the change, drawn from Canadian direction and the international standards it rests on. Beneath that: the key sizes, parameters and security strengths that decide those answers, and the structures the post-quantum algorithms are built on. The deadline and the reason for it live in the same library, so a conversation can move between them, and the conditions are kept intact, because a deadline stripped of its qualifier is worse than no answer.
The six libraries are ours. What you put beside them is yours, and it stays yours.
You can place your own documents in front of the engine, whether a draft, a supplier’s design, a departmental plan or your own published work, and it will read them, use them and cite them alongside the libraries. None of it is sent anywhere. There is nowhere for it to go: the engine is on the machine in your office, and the Centre has no way to reach it.
In the forge, tempering is the heat treatment that sets a steel’s character, and in plain English, a temper is what we call someone’s disposition. Both meanings are the point.
A temper is a small collection of your own past writing that shapes how the engine writes for you. Assemble the pieces you have already written, keep them together, and work with them to hand. Write about your subject and the engine draws on what you wrote about that same subject, matching your cadence and your style, without ever reproducing your sentences.
A temper changes the voice, never the facts. It is never quoted as a source and never cited as authority. Your own published work can be cited. That is ordinary material, and the engine treats it like any other document you supply. A temper is the deliberate exception: it is there to be sounded like, not to be relied on.
A temper is yours alone. It lives in your own account, on your own machine. It is not one of our Ingots: we do not build it, sign it, licence it or see it. Where several people at an office share one engine, a temper is not shared with them. Your colleagues cannot read it, and neither can we. It is your writing, kept where you already keep it.
The engine is tailored to answer only from material it can cite, and to say plainly what it could not see. That is the whole of the difference, and it is the reason the libraries are useful to people whose work has to be defensible: advisors, policy writers, counsel, researchers, academics, journalists and analysts.
A general assistant will fill a gap from memory and sound no less confident for it. This one is configured not to.
What did she say about this in her 2019 address?
The register holds nothing by that speaker for 2019, so I have not answered from memory. It does hold four addresses by her between 2015 and 2022. Shall I list them?
Which of these algorithms have to be retired, and by when?
Three of the four have a published end date. The fourth has a condition attached rather than a date, and I have kept the condition with it rather than rounding it to a year. Each line cites the guidance it came from.
Does the attached design meet the baseline?
I read the twelve pages of the forty that bear on the controls in scope. Findings are below, each against the control it is drawn from. The remaining pages I have not assessed, and I am not treating them as compliant.
No configuration makes a system correct, and none of this removes the need to check the citation. What it does is make the checking possible.
Plainly, in your own words, and about your own documents. A few examples of the kind of work the libraries are built for:
Here is my draft article. Write the next section in the same voice, grounded in the treaty text and state positions the international library holds.
Here is our departmental plan. Where does it fall short of current policy? Give me a table: requirement, the section that answers it, and how well.
Here is a supplier’s design document. Where does it depart from the architecture and the contract it was written against? One finding at a time, each with its reference.
What does the Act actually say on this point? Give me the words, not a paraphrase.
Which algorithms must be retired, by when, and under what conditions?
Here is a draft speech. Align it to the themes and cadence of the historic record, without reproducing any of it.
Work of any size is done a section at a time. You are not asked to get a long request right in one attempt, and neither is the system.
One ordinary desktop machine, bought once, running in your own office. There is no subscription, no account, and no connection required after the ingots are in place. That is not a limitation of the design; it is the design. Your questions are often more revealing than your documents, and a question asked of a service somewhere else is a question you have disclosed. Here the material, the model and the reasoning all stay on hardware you own, under your own control, and in your own jurisdiction.
The system chooses what to run from the hardware it finds, so a smaller machine is served differently rather than served badly. Members pay nothing for the ingots; the Centre is a not-for-profit and the libraries exist to be used.