Constrain
Bounded iteration, checked arithmetic, explicit absence, region-managed memory and a narrow API catalog replace broad, ambient access to C and libc.
01
The idea came first. Practice raised the bar.
A language refined by doing.
We began with a realization: AI models work better within enforced constraints—and every constraint a programming language can enforce is ultimately formal in nature.
A constraint must be expressible as a rule the toolchain can evaluate, rather than advice the model must remember. That led us to experiment with a small set of purpose-built C primitives whose individual contracts we could verify with Frama-C.
With a tiny core, we built a Minsky machine, demonstrating that a tightly constrained vocabulary could still express general computation. From there, we guided state-of-the-art AI models through increasingly demanding projects, producing hundreds of thousands of lines and adding primitives only when real workloads justified them.
Kibrahacha is the resulting deep module: the accumulated constraints, evidence, interfaces and lessons needed to let AI focus on the program’s logic.