AI research publications: the arithmetic, before anyone pays for it.
Our Substack runs the same working we put on a ranked row — what a capability costs at real volume, what it pays back, and where the two do not meet — on cases nobody hired us to look at. It is the cheapest way to check whether our method survives contact with a problem before you buy a sprint measured with it.
Roughly monthly. Free, and no paid tier.
What is actually in it.
Three things, and none of them is a product update. If a post could have been written without a number in front of us, it does not go out.
The arithmetic, in public
What a capability costs to run at real volume and what it pays back in a software business — the same working we show on a ranked row, on a case nobody paid us to look at.
What broke, and what it cost
The deployments that did not clear the bar, written up with the numbers that sank them. These are more useful than the wins and almost nobody publishes them.
What just got cheap enough
Models move every few months, and the thing that failed on cost last quarter sometimes clears this quarter. When a line moves far enough to change a verdict, it gets written up.
Why it is free, and why it stays that way.
There is no paid tier and no posting schedule to hit. It goes out when there is a number worth publishing, which is a lower bar for us to clear honestly than a weekly slot would be.
The publication is the argument for the method, so charging for it would mean charging for the pitch. Everything in it is the work we would have done anyway to keep our own numbers current — writing it down costs us the afternoon and saves you taking our word for it.
How this differs from the blog
The blog is method: how we build a test set, what a “build” verdict has to clear, how we price a system once the demo is over. The Substack is cases — one problem at a time, costed end to end. Read the blog to judge the method, the Substack to watch it run.
Subscribe on Substack — it goes to your inbox, and you can leave any time.