A model that shows its work: not after the fact, but live, as it thinks. That's the whole point of Logos: you shouldn't have to trust the reasoning. You should be able to watch it.
Most models hand you an answer and hide the work. We do the opposite. Interpretability isn't a safety feature we bolt on at the end. It's the native language of a model you've actually looked inside.
We read the residual stream directly: which features fire, where the manifold integrates, the moment the model stops formatting tokens and starts thinking. Sparse autoencoders break a representation into the features the model actually uses; the topology tells us where those features unify; the circuits tell us how they connect. Three instruments, one legible account of the computation.
On one screen, you talk to the model. Alongside the conversation, a live feed of what's happening inside: the features lighting up, the workspace integrating, decoded as it thinks. A raw stream of firing features is a firehose, so a second agent watches it for you and surfaces the moments that matter: this is interesting, look here.
And when something does matter, you click it, and the whole thing opens on a second screen: the model's thinking as a structure you can browse, zoom into, and explore. Not a chart of the answer. A map of the reasoning that produced it.
Every layer, on real prompts, in the running model — not a post-hoc reconstruction.
The instruments were built in the same model they measure: the kernel taps its own residual stream as it runs, so what you watch is the actual computation, at native precision, in real time. The process observes itself.
Nomos makes the computation lawful. Logos makes the reasoning legible. Verify the one, read the other.— The two pillars, together