Research

This is where the lab does that. Not a feed of opinions about where the field is heading, but a record of what we put our hands on: a preview project carried past the demo, an eval run until it said something, a tool shipped and left open to inspection, a sprint run on a real codebase and reported down to its numbers. Research, here, is the work that survived contact with reality.

The frontier moves monthly, and most of what gets written about it is forecast: confident, unfalsifiable, gone by the next release. We would rather be wrong in a way you can check. So every finding we publish carries its workings. When the work is ours to open, that is the repo, the commits, the evals, there for you to re-run. When it is a client's, it is the dated measures and a plain account of what we hold back and why. What we will not publish is a result with nothing behind it but our word: that is only a well-dressed guess.

That standard is why the room fills slowly. An entry earns its place here by being true, not by being timely. We would rather you find four things that hold than forty that flatter.