Where this is headed.
The AI frontier is advancing on raw capability while the reasoning underneath it stays unsound. Large language models pattern-match: they produce an answer one token at a time, and the chain of thought shown alongside it is a narration, not the computation that generated the answer. Research into chain-of-thought faithfulness has repeatedly found the two come apart, so the explanation a model gives is not a reliable account of how it actually reached its conclusion. Scaling has made this more capable without making it more sound. You can buy more horsepower, a larger model and more compute, but the thought process, the part that decides whether a hard judgment is actually right, is not something the current paradigm produces.
The work that matters sits upstream of the capability race, and we build it from the bottom up. The deepest problem is the thinking itself: getting a model to reason in a way that is structured, grounded, and checkable. That is what Redstone is. It imposes a process from outside the model and enforces it in code, so the model decomposes the problem, states a position before it gathers evidence, checks its load-bearing claims against reality, challenges its understanding and then its plan, and only then executes. Each step exists because the process required it, and the trace is the actual structure of the work rather than the model's account of itself. A sound, stable thinking process is the floor everything else stands on, and almost no one is building it.
On that floor, the next problem comes into reach: software that improves itself. Recursive self-improvement is becoming a real research direction, and the systems that work validate every self-modification against a benchmark, code that compiles or a score that moves. The decisions with the highest stakes are exactly the ones with no benchmark to check against. We route that judgment through the person: in Redstone the human is the verification signal, and because the structure of the work is captured as it runs, a change can be evaluated where no automated test exists. Furthest out, that captured process becomes something to train on directly, a model shaped by how good reasoning actually moves rather than by the finished answers the internet already holds.
This is the order we believe it has to happen in, and the order we build in. Get the thinking right, make its improvement verifiable, then learn from the process itself. Instance B exists to solve those problems, on a foundation solid enough to hold the ones above it.