Price, and the work it replaces

What it costs, what work it replaces, and where to read further.

Start with a pilot. Pay per plant after it.

Not per device. Not per tag. Not per connection. The shape of the price is the part we are committing to, because a licence that grows with every instrument is the thing that makes this category expensive to live with once a plant has a few hundred pieces of equipment.

NOK 300,000a pilot: one plant, three months, a fixed price, with the pass marks written down before it starts
from NOK 450,000per plant, per year, after the pilot

A pilot is how we would test anything else, so it is how we sell. Before it starts we write down together what it has to show: how many of the plant's tags can be wrong at most, from a spot check your own people answer, and how long the same jobs take with and without the console, timed by you. If it misses, you know before you have bought a year. Why these numbers: a pilot and two years come to NOK 1.2 million, under the national threshold of NOK 1.3 million above which a public buyer has to announce a tender, so a municipality can start without one. Prices are excluding VAT. We have not yet run a year of this at a paying plant, and the work it replaces once, mapping the tags, does not alone pay for a year; the pilot's timing is what has to show the rest, and we will publish what it shows.

We write down what would prove us wrong, and then publish it when it does.

28 experiments written down before their data was opened. Many came back falsified or only partly right. Every one of them is on this site or in the repository, with the criteria that were fixed before the run and the number that missed them. That is not modesty. It is the only way to know whether any of this works, and it is the fastest way we know to build something true.

10 of 16We publish the failures.

Every experiment is written down before it runs, with the number it has to beat. Ten of sixteen came back under it. All ten are published, with the reasoning we got wrong. One cost us a tenth of a capability that already worked. Another spent a day teaching a model to do a job that ten lines of if-statements then did better, and we shipped the if-statements and wrote that down too. An idea that survives this is worth building on. One that was never tested is a rumour.

175 msWe measure before we rewrite.

A compile took 2.2 seconds and the obvious answer was to move it to C++. Measuring it first showed 96% of the time inside one standard library call. It now takes 175 ms and there is no case for rewriting anything.

0No number reaches a screen unless a measurement made it.

Every figure written by an agent here has to trace back to something computed from the plant's own data. The finished sentence is read again and thrown away if a stray number appeared in it. A confident wrong number in a control room is worse than silence.

1One founder, one laptop, no data centre.

Everything on this page runs offline on the machine in front of you, including the model. Nothing about this company needs permission from somebody else's cloud, which is also the only honest answer to a plant that will not let its data leave.

Metal and process a potline, a casthouse, and what a reading is worth there Grid infrastructure bays, breakers, and a transformer worth more than the station Power and water turbines and pumps, and the cost of a reading nobody trusts Municipal water the plant with no data engineer, which is most of them What it costs to run measured compute, and the seven jobs of work this actually replaces What the words mean a pot, a busbar, a tag that is railed: every term on this site, explained

Nothing leaves the plant, including to us.

The compiler, the model and the agent all run on hardware inside the fence. Not a private cloud region, not a tenancy with a European flag on it. The export folder is dropped on a machine you own, which is worth saying plainly because "hand it over" would mean the opposite and would trigger a data classification review instead of avoiding one. The site keeps working when the line to the outside does not.

That claim has a size attached to it. A 9,511 tag plant compiles in 449 seconds inside 0.6 GB, the language model that reads the paperwork is 3.0 GB on disk, and pinning the whole thing to a single core makes no difference this machine can measure. The measurements are here, along with the seven jobs of work it replaces and the two it is bad at.

What actually takes the time, and what we do about it.

Not the connection: the category already connects in days, and so do we. Two things bind. The person who knows what the tags mean, whose afternoon is the only configuration there is, and whose answers we make count for more than one tag. And the supplier security assessment, which is a legal duty for your buyer under NIS2 and takes longer than everything else combined. We shrink what it has to cover: nothing leaves the plant, no port is opened, and the sealed run that proves it is published.

What we do not control, and will not put a clock on: your procurement, the NDA, that assessment, and an operator signing off the model.

Back to the front page