Models that say
when they don't know.
A language model that always answers confidently is more dangerous in production than one that stays quiet. Neurolism builds the layer that knows the difference, and measures whether it is right.
What you get on the first day.
Before any of the architecture below: the things you can use straight away, on every plan including the free one.
In the browser and as an app
The browser needs nothing installed. The desktop app opens the folder the work is already in - a repository, a directory of contracts.
What that changes. For you, the same history on every device. For a company, the code and the documents never get copied somewhere else to be read.
Reads what you attach
PDF, Excel, Word, images. A spreadsheet is read as cells rather than as flattened text.
What that changes. Where the file does not cover the question it says so. That is the difference between someone checking one answer and someone checking every answer.
Searches and names the source
The passage it used arrives with the answer. Where the search turned up nothing, that is the answer.
What that changes. You check it in seconds instead of doing the search again - and in regulated work, an answer that names its source is one somebody can sign.
None of it trains anything
Nothing you send is used for training, ours or anybody else’s. Not a setting you have to find and switch off.
What that changes. For you, nothing private ends up in a future model. For a company, it is the first question a data protection officer asks, answered before it is asked.
Three kinds of week.
The same model, three completely different days. What follows is what it actually does in each - not a list of adjectives.
Building software.
Open a repository in the desktop app and it works where your code actually is - not on a copy you pasted into a chat window.
- Reads the files, not just the file names
- Runs
dotnet buildanddotnet test, then reads the errors git status,diff,log,blamewithout asking; commit and push only with your say-so- Writes files into your folders, one confirmation each
- Force pushes and hard resets are not available at all
Learning something.
Explanations that stop where the knowledge stops, instead of filling the gap with something that reads well.
- Walks through a problem step by step, not just the answer
- Turns a script or a chapter into flashcards and questions
- Asks back when the question is ambiguous
- Says plainly when it is unsure - the whole point of the design
- Reads photographed notes and slides
The dull half of the week.
Reading what piled up and writing what has to be written - with the files you already have, in the formats you already use.
- Reads contracts, spreadsheets and PDFs you attach
- Hands back an .xlsx or .docx you can open, not a pasted table
- Turns numbers into a chart inside the conversation
- Drafts offers and replies in a second language
- Looks up what changed and names the source
- Nothing you write is used for training
Four problems, and what we do about each.
These are not marketing complaints about the field. They are the four constraints that decided how this system is built.
- The problem
A wrong answer sounds exactly like a right one
Fluency and correctness are separate abilities, and models got good at the first one much faster than at the second. Nothing in a confident paragraph tells you which parts were retrieved and which were invented.
Our answerA layer whose only job is to disagree
Language and evidence live in separate representation spaces. A third space scores whether they actually match, and has a dedicated channel for saying that they do not. That channel is why a smaller model stays usable.
- The problem
Confidence scores that are confidently wrong
Most systems report a number between zero and one and never check whether it means anything. On a set where most pairs agree, a model that always says yes scores well and has learned nothing.
Our answerCalibration measured, not assumed
We measure expected calibration error and AUC on a balanced set, never accuracy. Of the times the system says seventy percent, roughly seventy percent have to be right, or the number does not ship.
- The problem
Inference cost grows with every user
Training is paid once. Inference is paid every day, forever, and a system designed without that in mind gets more expensive exactly as it gets more successful.
Our answerEscalate only when it is needed
A small model answers, a router decides whether that was enough, and only the remainder reaches a larger one. Evidence is precomputed into an index rather than embedded inside the request.
- The problem
Nobody can tell you where the training data came from
For most models the honest answer to a question about provenance is a shrug. That is a problem the moment somebody has to sign off on the output.
Our answerEvery source logged, own hardware
Origin, licence, date and volume are recorded for every source that goes into training. The models are trained on our own machines in Switzerland, not on rented capacity somewhere unspecified.
Three sizes, one behaviour.
Lumen is the family answering today: three sizes that share the same abstain logic. What differs is how much they can recall unaided, and how long a chain they can hold. Medium and Large take images and documents alongside text.
A second family, Gauss, is in planning: our own weights, trained from zero, with the same layer built into the architecture instead of wrapped around it. A 1.16-billion-parameter model of it exists and works. Everything above it needs funding.
13 kinds of work, one account.
Three of them are above in full. Every one has its own page: what it does, how a session goes, and where it stops.
Building things
Where the output is a file somebody else will run or read.
Checking things
Where the job is to find what is there, not to produce something new.
Everyday work
The half of the week that is reading, sorting and explaining.
Where the rules are strict
Work that has to survive somebody asking where the answer came from.
It gets to know you, and shows you what it learned.
From your own conversations it distils a short list of traits - that you write C#, that you want answers without a preamble, that you work in Swiss francs. So you stop explaining it every Monday.
- Traits, never transcripts - and nothing sensitive, by rule
- Drawn as a net under Neural net: every trait with the weight it actually carries
- One switch to turn it off, one button to forget everything
- Incognito conversations neither feed it nor use it
It works things out instead of estimating them.
Thirty-eight tools the model can reach for, and the ones that matter are the boring ones: the questions where a fluent answer and a correct answer look identical. Every tool on this page is tested against a known result on the status page, not listed and hoped for.
“What is the median of these 340 numbers?”
Sorted in its head, a model lands near the median. Near the median is a different number, and nothing in the sentence says so.
“What is the payment on 300000 over 25 years at 3 percent?”
An annuity is a formula with a power in it. Estimated, the figure looks right and is out by thousands over the term.
“When is that meeting in New York, on 29 October?”
Six hours apart, as everyone knows. For four weeks a year it is five or seven, because Europe and the US switch on different days.
“Why does this string not match?”
A non-breaking space, a soft hyphen, a Cyrillic letter that looks Latin. Invisible on screen, and invisible to a model too - it reads tokens, not characters.
“Solve for x, and show the steps.”
Rearranging is rule-following. A sign error in step three reads exactly like a correct step and changes the answer.
“How much did we spend on this in September?”
A business account can put read-only SQL over its own usage. The alternative is a number that sounds about right.
The rest read files back, hand you one, or run on your own machine. The full list is here.
Four promises, and what each one costs us.
Every one of these is a decision that makes the product slightly harder to build. That is what makes them worth writing down.
- Not used for trainingYour conversations train nothing. Not our models, not anyone else's. That is a decision about the product, not a setting you have to find.
- Yours to deleteDelete a conversation and it is gone from the database, not hidden. Delete the account and everything attached to it goes with it, in one step.
- Incognito when you want itAn incognito conversation is never written down at all. It exists in the window you are looking at and nowhere else.
- Switzerland, on our own machinesServed from hardware we run, with a written record of every source that went into the models. Not a badge, a paper trail.
The same models behind your own product.
A separate business account, API keys, prepaid credits and usage you can see per key. No monthly seat you have to grow into.