Latent Intuition Labs
The mind is the latent variable. Everything we build starts there.
We study what a person actually knows, believes, and can do — the things nobody can accurately report about themselves — and build tools that use that understanding to make them more capable, not less necessary.
Nobody can accurately report their own mind. Eleven observations can pin down a piece of it. Getting that estimate honestly is the part everyone skips, and the part we refuse to.
Mission
Capability, not dependency.
Understanding a person is power over them. The same model of your mind that could teach you something can also keep you scrolling, sell you what you did not want, or quietly decide what you never get shown. Nothing in the technology settles which of those you get; the product does. That choice — made quietly, a hundred times over, in the design of a tool — is the entire difference between software that grows a person and software that hollows one out.
A tool that hands over the answer leaves you exactly where it found you, only faster. It produces output, and output is worth something. But understanding produces judgement, and judgement is what you bring to the problem the tool has never seen. One of those compounds across a career. The other has to be bought again every single time you need it.
We are not interested in whether people can be replaced. We are interested in how far a person can be taken by a machine that genuinely understands them and is genuinely on their side. That is a design problem before it is a modelling problem, and almost nobody is treating it as one.
Every product we ship has to pass one test: does the person close the tab more capable than they opened it? Not better served — more capable.
The problem
Every tool already has a model of you in it.
Every piece of software that helps a person contains a theory of that person. A difficulty setting is a theory. An onboarding flow is a theory. A recommendation is a theory, and an engagement score is a very specific and very unflattering one. These theories are mostly implicit, mostly crude, and mostly never checked against the human sitting there.
The gap between that cartoon and the actual person is where tools fail — not because the software is badly built, but because it is aimed at an average that describes nobody. What someone knows, what they believe, where their skill actually stops, what they are quietly avoiding: none of it is on the settings page, and none of it can be got by asking. People are not being evasive. Most of what governs how a mind works is simply not available to its owner to report.
It can, however, be inferred. That is old science — psychometrics has estimated unobservable traits from observable behaviour for a century — and it has been stuck at the scale of a paper test. What changed is that a model can now read what a person produces, in their own words and their own work, and keep a picture of them current. A specific, honest, living model of one mind is finally something you can build.
Elicit
Evidence from what a person does under conditions we chose, rather than from what they report about themselves.
Infer
A model of the state you cannot see — what they know, what they can do, where the edge is — carrying its uncertainty in the open instead of hiding it.
Act
A product built on that model and moving on the person’s behalf: the next thing they need, not the next thing that holds their attention.
Verify
Every model of a mind predicts something. Check it against what actually happened, later, or it was a story.
That model is not the product. It is the substrate. Once you know where a person actually is, you can build a tool that starts there — and the same understanding, aimed differently, is precisely what makes software addictive. Both are available to anyone who does this work. We are only interested in the first.
What’s next
One product is a proof, not a company.
Cognometrist exists to show the method holds somewhere it can be checked. Learning is a good first domain precisely because it is unforgiving: either the person can answer the question in June or they cannot, and no amount of pleasant interface covers the difference. A model of a mind that survives that is worth pointing at something else.
So the roadmap is not a list of features for one product. It is the same machine — elicit, infer, act, verify — rebuilt wherever an honest model of the person would change what the software ought to do. We are deliberately not naming those here. Announcing a product before it works is the sort of claim this page has spent several sections arguing against, and we would rather be judged on what has shipped.
If you are working on something in that shape, or you want to know what we are building next before it is public, the fastest route is to write to us directly.
How we work
Principles we’d like to be held to.
These are not aspirations posted on a wall. Each one is a rule that has already cost us something — a feature we did not ship, a claim we could not make, a shortcut we did not take.
-
Custody
The model of you serves you.
Everything we learn about a person exists to act on that person’s behalf. We do not sell it, we do not rent it out for targeting, and we do not tune it against time-on-app. A company whose core asset is understanding people owes a plain answer about where that understanding is pointed.
-
Disclosure
Name the machinery.
We tell people what the software actually does — what it measured, what it inferred, and how sure it is — instead of selling “AI” as a mood. Users make better decisions when the mechanism is legible, and we write better software when we have to describe it plainly.
-
Evidence
Earn the claim.
If we say an output is grounded in your own material, there is a citation you can open and check. Every sentence of marketing should map to a mechanism a sceptical engineer could go and verify. Where we cannot show the mechanism, we do not make the claim.
-
Authorship
The person stays the author.
Our tools do not do the work for you. This is not a compliance posture; it is the product thesis. Software that does the work for you removes, one convenience at a time, the reason to be able to do the work at all.
-
Failure
Design around the failure mode.
Language models are confidently wrong sometimes — a property of the technology, not a bug awaiting a patch. So we build for it: retrieval instead of recall, citations instead of assurances, and measurement that checks the person rather than taking anyone’s word for it.
-
Scope
Ship the whole loop.
A prompt wrapped in a text box is not a product. The valuable work is everything in between: eliciting the evidence, holding the structure, evaluating, scheduling, and the state that has to survive someone closing the laptop for a week and coming back.
-
Horizon
Prefer the durable result.
We optimise for what someone retains in a month, not for how productive they felt in a session. Those two numbers disagree far more often than the industry likes to admit, and when they do, we take the first one.
Contact
Get in touch.
Questions about our products, press, partnerships, or a bug you would like fixed — one address reaches us, and a person reads it.