Kaer Labs publishes what it learns
Frontier agents, RL and applied alignment. The lab is not a mood board. It is the group that writes down what broke.
Kaer Labs Ltd makes Kaer, and its research notes are at kaerlabs.uk. They do not replace the product blog, and they do not exist to make the marketing site look like a university.
When a training run, an evaluation, or an alignment note is fit to publish, it goes to kaerlabs.uk. When a price, an architecture choice or an integration is fit to publish, it stays on this site.
The split is deliberate. Research that is still wet does not belong next to a pricing table.
kaerlabs.uk describes the work: small reasoning models, studies of how models behave in real use, and tests for the services that rely on them. The lab's own models get the same tests as everyone else's.
Everything it publishes comes with its method, its cost and its limits. Its first model in development, Kaer-R1, is a small reasoning model with a learned halting head: it decides, question by question, how long to think, and the same signal flags questions it is likely to get wrong so they can go to a person or a larger model.